weltliteratur.net

DraCor Platform Update (July 2025)

2025-07-14T00:00:00+00:00

tl;dr: Enhanced API, Frontend and Schema, Comprehensive Documentation, and Upcoming Summit in Berlin.

After bigger updates in August 2021 and December 2023, we are excited to share that DraCor just got another upgrade!

This latest release introduces a wide range of improvements – from enhancements to the API and frontend to refinements in the underlying TEI/XML data model. It is the result of months of collaborative work by our team, all in the spirit of keeping DraCor open, transparent, and easy to hack, build on, and explore. The new features come without any breaking changes.

DraCor was built as an open digital infrastructure for the computational study of European drama – spanning from ancient Greek tragedy to 20th-century theatre. Since our last major technical update, the platform has grown not only in functionality but also in content, thanks in large part to our global community of researchers and enthusiasts who have contributed new corpora. Today, DraCor hosts more than 4,000 plays across nearly 30 corpora, representing over 20 languages – and the platform continues to grow.

Before diving into the roundup of changes, we would like to give credit where it is due. The lion’s share of this update was made possible thanks to the tireless work of our technical lead, Carsten Milling, and co-lead, Ingo Börner. Development was supported through funding from CLS INFRA and DraCorOS.

💡 A Smarter, Cleaner API

The DraCor API has seen its most comprehensive revision since the project began. Among the many changes are the following:

DTS Support: We have implemented a /dts endpoint to align the API in line with the Distributed Text Services (DTS) specification. A DraCor Notebook documenting the new endpoint is available here.
Plaintext Endpoint: Users can now fetch a plain text version of any play via a dedicated endpoint – making it easier to use DraCor data in NLP pipelines and other tooling.
Corpus Metadata Enhancements: We now track Git commit hashes from the corpus repositories, adding another layer of provenance and traceability to our data.
Improved Filtering: The spoken-text endpoint now features refined relation filter parameters, offering more precise ways to query dramatic dialogue.
Inclusive Metadata: The API now supports both sex and gender attributes, enhancing its ability to reflect diverse encodings of character identity.
Better Developer Experience: We have extended and refined our OpenAPI specification, added more comprehensive tests, and introduced a robust CI workflow for building Docker images.
Upgrades: Under the hood, we have moved to eXist-db 6.4.0 and dracor-metrics 1.5.1, improving both performance and maintainability.

Full API release notes here: dracor-api changelog.

🖥️ A Refined Frontend for Exploration

The DraCor frontend has also received a series of upgrades that improve usability and extend functionality:

New “Tools” Tab: Data from each play can be sent directly to third-party analysis tools, such as Voyant Tools, Gephi Lite or the CLARIN Switchboard.
Corpus Status Overview: We have added a status page that provides an overview of the current state, maintenance responsibilities, and licencing of all DraCor corpora – useful for developers, curators, and researchers alike.
Rendering Fixes & UI Improvements: A variety of bugs were squashed, including issues with rendering segment lists. The toolchain has been updated, and we have migrated to pnpm for improved dependency management.

Full frontend release notes here: dracor-frontend changelog.

🧾 TEI Schema 1.0: A Solid Foundation

One of the most substantial and long-awaited developments is the release of the first stable version of the DraCor TEI schema. It all began during a coffee break at the CCLS 2024 in Vienna – and ended in a full overhaul of how we define and validate DraCor corpora. Highlights include:

Based on tei_drama (TEI 4.9.0): The schema builds on the official TEI customisation for drama, ensuring compatibility with broader TEI ecosystems.
More Flexibility & Clarity: We have revised the structure to make it easier for contributors to encode new plays and for the schema to accommodate varied metadata.
New Elements & Better Validation: We now support elements such as <region>, <orgName>, and <addName>, and introduced Schematron rules tailored to DraCor’s specific needs.

Full schema release notes here: dracor-schema changelog.

📜 Comprehensive Documentation

Our contribution to the EU project CLS INFRA is documented in several detailed reports that provide comprehensive insights into DraCor:

D7.1: The report “On Programmable Corpora” presents DraCor as a prototype for an infrastructural ecosystem of Programmable Corpora, defined as corpora that expose an open, transparently documented, and a research-driven API, enabling machine-actionable access to texts. This concept builds on on Aaron Swartz’s vision of “A Programmable Web” (2013), in which applications form an interoperable, dynamic ecosystem through the “natural growth” of APIs from the resources they make accessible. (doi:10.5281/zenodo.7664964)
D7.2: The “Report on API Libraries for R and Python for the Programmable Corpora Prototype DraCor” introduces the DraCor libraries, (rdracor for R and pydracor for Python), which provide researchers with direct programmatic access to DraCor’s corpora in their preferred programming language. (doi:10.5281/zenodo.15302236)
D7.3: The report “On Versioning Living and Programmable Corpora” addresses the challenge of ensuring research reproducibility with living corpora by recommending Git-based versioning for dynamic corpora and containerisation for complex programmable corpora infrastructures. (doi:10.5281/zenodo.11081934)
D7.4: The final report of our CLS INFRA work package, “Report on the Implementation of Programmable Corpora”, documents the evolution of DraCor into a standards-compliant, extensively documented service. It now features a stable API (version 1.x), Distributed Text Services (DTS) endpoints, and new tools such as EzDrama and Who-is-@who, all of which contribute to transforming the experimental prototype into a sustainable piece of European infrastructure for computational literary studies. (doi:10.5281/zenodo.15301341)

📅 Save the Date: DraCor Summit 2025

From 1 to 5 September 2025, the DraCor Summit will take place in Berlin – hosted by the Freie Universität Berlin and the University of Potsdam. It will be five days of workshops, talks, a corpora conference, a computational drama analysis workshop, a barcamp, and even co-working time. The Summit is open to anyone working in or curious about computational literary studies, corpus development, or cultural analytics. There is no registration fee. General registration opened on 6 June – we would love to see you there!

🧭 Follow the Drama

To stay up to date with DraCor news, project updates, and community events, subscribe to our mailing list and follow the official DraCor account on Bluesky.

(Portions of this blog post were composed with assistance from Claude Opus 4.)

How to Identify Key Passages in Literary Works? Using Algorithms and Machine Learning!

2023-12-19T00:00:00+00:00

Context

In the DFG-funded project What matters? Key Passages in Literary Works (which is part of the special priority programm Computational Literary Studies) we set out to identify and characterize key passages in literary works.

We understand key passages as passages that are particularly important to expert readers when interpreting texts. In a mixed-methods approach, we investigate empirically which textual characteristics of literary genres can be revealed through patterns of citation and quotation.

Corpus

Our main corpus consists of two literary works Die Judenbuche by Annette von Droste-Hülshoff and Michael Kohlhaas by Heinrich von Kleist with 44 and 49 scholarly articles, respectively. Fortunately, we could build on the previous work of the ArguLIT project and their annotation of all direct quotations.

Automatic Identification of Quotations

Scholarly texts contain different types of quotations. For example, verbatim quotes of single words to longer quotations spanning multiple sentences, and indirect quotations in the form of summarizations or re-narrations. In the first phase of the project, we focused on the automatic identification and linking of direct quotations starting with quotations of a length of five or more words. In Lotte and Annette: A Framework for Finding and Exploring Key Passages in Literary Works¹, we outline the current landscape for text reuse detection and the development of our tool Quid. Although there are a number of existing tools, we found that all had limitations for our specific use case. We evaluated Quid and compared it to the existing tools.

Approach	Die Judenbuche			Michael Kohlhaas
	Precision	Recall	F₁	Precision	Recall	F₁
	BLAST	0.59	0.61	0.60	0.37	0.59	0.45
Copyfind	0.85	0.75	0.79	0.76	0.79	0.78
SimilarityTexter	0.91	0.64	0.76	0.83	0.74	0.79
Textmatcher	0.69	0.37	0.48	0.68	0.42	0.52
Quid	0.82	0.90	0.86	0.70	0.90	0.78

Table 1. Comparison of different approaches for text reuse detection with an evaluation on our corpus.

Considerably more difficult to identify are quotations which are shorter than 5 words. In A Novel Approach for Identification and Linking of Short Quotations in Scholarly Texts and Literary Works², we develop and compare two approaches to tackle this challenge, ProQuo and ProQuoLM.

For ProQuo, we use the (page) references for long quotations as examples to tell apart (page) references for short quotations from other text in parenthesis. This includes references like those to the Bible or other literary works. We then relate short quotes to their source in the literary work by figuring out the relationships between the quotes and references. We also use the positions of long quotes as guides to link short quotations to the correct passage of the literary work.

For our second approach, ProQuoLM, we fine-tune a German BERT for classification. First, we identify potential short quotes, and then use the fine-tuned model to filter them.

Approach	Die Judenbuche			Michael Kohlhaas
Approach	Precision	Recall	F₁	Precision	Recall	F₁
Baseline	0.65	0.78	0.71	0.59	0.75	0.66
ProQuo	0.87	0.72	0.79	0.87	0.66	0.75
ProQuoML	0.88	0.75	0.81	0.87	0.69	0.77

Table 2. Evaluation results of our two approaches compared to a baseline which always links a quotation from the scholarly work to the first matching instance in the literary work.

QuidEx – Visualization and Exploration

We created QuidEx, a website for visualization and exploration of the results, which is shown in this screenshot:

On the left, there’s a heatmap that displays the distribution of quoted passages in the entire literary work. The darker the area, the more frequently it has been quoted, suggesting its significance. Right beside the heatmap is the literary work itself. The grayscale indicates how many scholarly works quote any part of a crucial passage. This means the level of gray remains constant for the entire key passage. The font size is adjusted based on how often a minimal segment is quoted. At the bottom, alongside the literary text, there’s a list of all scholarly works.

The source code of the website is available as a white-label version which facilitates adoption by others, for example, for the comparison of intertextual relations of literary texts.

Summary and Outlook

In the first phase of the project, we developed tools to identify, link, visualize, and explore direct quotations of all lengths. In August 2023, the project went into its second phase, titled Is Expert Knowledge Key? Scholarly Interpretations as Resource for the Analysis of Literary Texts in Computational Literary Studies. One important task we are currently working on, is the identification and linking of indirect quotations, that is, summarizations and re-narrations.

For daily key passages, follow us on Bluesky or try Quid online with our web interface.

Lotte and Annette have since been renamed to Quid and QuidEx, respectively. ↩
Accepted at JCLS 2023 and soon to be published. ↩

Streamlining the DraCor API

2023-12-01T00:00:00+00:00

Again, What Is DraCor?

DraCor, the Drama Corpora platform, is an infrastructure built to facilitate the digital research on European drama. We have currently 25 (!) TEI-encoded DraCor corpora in 19 (!) languages/dialects:

Alsatian, Bashkir, Czech, Dutch, English, French, German, (Ancient) Greek, Hebrew, Hungarian, Italian, Latin, Polish, Russian, Spanish, Swedish, Tatar, Ukrainian, Yiddish

16 of these corpora are already running on our production server (dracor.org), the others are still in the development phase and will be released when ready. Some corpora (like the Shakespeare corpus) have a stable number of plays. Depending on the corpus strategies of their maintainers, other corpora can generally grow and be updated with new plays. We call them ‘living corpora’.

DraCor API 1.0

The TEI-encoded plays are stored in an eXist database. Various encoding layers are extracted and made available via our API: customised data for individual research purposes. You can easily access the full text or text slices (a list of characters, spoken texts, stage directions) or co-occurrence networks. All API endpoints are documented following the OpenAPI specification.

A little bit of history. The DraCor API has grown organically ever since we introduced it sometime in August 2017. Yet the first publicly documented release on GitHub was 0.72.0 in September 2020.

New endpoints were added on a frequent basis and made DraCor what it is today, a research-prone platform serving data for computational literary studies. We always tried to use meaningful names and adhere to naming patterns, but over the years some inconsistencies have slipped in.

Henny Sluyter-Gäthje revisited the consistency of names when working on her pydracor library, a wrapper for DraCor API endpoints (see GitHub ticket). Building on her work, we revised the DraCor API, making it more consistent and sustainable.

This work took a lot of time and after the streamlined API has worked well on our staging server for a while now, we release today the first version of the DraCor API that we consider stable, our 1.0.0.

Fig. 1. This mockup is part of the new DraCor Media Kit by Mark Schwindt, released along with the new API version.

Legacy API

From here on out, we would like that everyone who works with the DraCor API to use version 1.x. The URL paths on dracor.org now have a /v1 prefix (e.g., https://dracor.org/api/v1/info).

However, the last pre-release version of the API will remain to be available under URLs featuring a /v0 prefix (e.g., https://dracor.org/api/v0/info). Old URLs without a version prefix will be redirected to the versioned ones. For instance, https://dracor.org/api/info now redirects to https://dracor.org/api/v0/info. This should allow old scripts to function as long as they follow the redirect.

We plan to keep the API version 0.x available for at least a year. After this period, we may phase it out so as not to have to maintain multiple versions of the API indefinitely. Also see dracor.org/doc/faq.

pydracor and rdracor

Our API wrappers will be updated to work with our API version 1.0.

pydracor 2.0.0 is already available via PyPI, the updated rdracor will be available via CRAN.

Other Updates

We introduced a new /wikidata/mixnmatch endpoint to automatise the integration of DraCor permalinks into Wikidata. A corresponding DraCor property was also proposed and we wait for approval.

We retired the /corpora/{corpus}/plays/{play}/segmentation endpoint so as to reduce redundancies (the same information can be obtained via the /corpora/{corpus}/plays/{play} endpoint).

For now, the RDF endpoint has been removed from the OpenAPI specification as it needs more work. This is something we prioritise and plan to release rather sooner than later, probably with a 1.1 release of our API.

With this release, we also updated our eXist-db from 6.0.1 to 6.2.0.

The Team Behind the Update …

… includes Carsten Milling, Henny Sluyter-Gäthje, Ingo Börner, Peer Trilcke, Daniil Skorinkin (all University of Potsdam), Frank Fischer, Viktor J. Illmer, Mark Schwindt, Julia Beine, Heinz-Alexander Fütterer (all Freie Universität Berlin) and Ivan Pozdniakov. The implementation was largely carried out as part of the “CLS INFRA” project (funded through EU’s Horizon 2020 programme) and the Cluster of Excellence “EXC2020 Temporal Communities” (funded through DFG).

DraCor was founded as a community project and, as always, numerous friends and colleagues have contributed feedback and ideas while working on their corpora or conducting research based on data from DraCor.

The platform, once praised for its original guerilla strategy – although this is debatable 😊 – has become a well-frequented part of the international research infrastructure. We trust that this API 1.0 release will contribute to making digital research on European drama (and beyond) more reliable and sustainable.

Goodbye for now, may your requests be valid, your responses swift, and your status codes always successful!

Reverse Engineering "Frembdsch", A Fantasy-Language Radio Play by Dagmara Kraus and Marc Matter

2023-02-09T00:00:00+00:00

First aired on the German radio station SWR2 on December 17, 2022, “Frembdsch” (which could be translated into English as Fornlish or Foreysh) is an audio play composed by author/translator Dagmara Kraus and media artist/sound poet Marc Matter. In the performance the composers are joined by the voices of Susanne Reuter, François Cavaillès, and Uwe-Peter Spinner. For the most part, the dialogue of the play is in an invented language, while the stage directions, which were derived using the German Drama Corpus at dracor.org, are spoken in relatively understandable, mostly standard High German. Hörspiel und Feature describes the piece as being based on the concept of “mots fantômes,” incorrectly read and written phantom words that have found their way into dictionaries and then into the mouths of speakers. What is produced is a broad polysemy with very indefinite semantics.

What sort of language is Frembdsch? A multilingual speaker of a Romance, Germanic, or Slavic language might be reminded of Hungarian for its opaque semantics and familar phonology. Common word endings include -os and -ć which are reminiscent of the Greek noun and the Polish verb. Some words are repeated, among “gudem,” “finestre,” and “mazgrab,” which, as a German speaker, sound like “good,” “gloomy” or “window,” and “mass grave” respectively. The R is in most cases an alveolar trill, as in Italian or in Slavic languages. Phonologically it is clearly European, although which language it most closely resembles differs based on the speaker. Its least Indo-European feature is rather prevalent reduplication (as in some creole languages), whose function seems to be primarily narrative. The only period in which I detected a consonant unused in Indo-European languages was between ‘20:00 and ‘21:00, where an actor produces what sounds like a dental click (k͡ǀ), a sound only found in Africa and in the Damin ritual jargon of Australia. To this, the person saying the stage directions says “Befremdet” (tr. disconcerted or estranged).

As stated above, the stage directions were produced with the help of the German Drama Corpus. Many of these stage directions appear word for word in the corpus (some many times), other have been distorted and denatured, so as to make their identification impossible. Some were simply invented by the authors. I did a search in order to see how many times each stage direction used in Frembdsch appears in the German Drama corpus. The top ten are:

Stage direction	Count
Für sich. (tr. to oneself)	3015
Heftig. (tr. severely)	1537
Verwirrt (tr. confused)	234
Zuckend. (tr. twitching/flinching)	216
Geheimnisvoll. (tr. mysteriously)	199
Spöttisch. (tr. mockingly)	189
Von innen. (tr. from within)	187
Taumelt. (tr. staggering)	162
Neugierig. (tr. curiously)	146
Befremdet. (tr. disconcerted/estranged)	99

To oneself, severely, confused, flinching: these are the bread and butter of German stage directions.

I also compiled a list of the works in which the stage directions from Frembdsch most often appear. The top ten:

Play	Count
Ignorabimus by Arno Holz	173
Das Konzert by Hermann Bahr	103
Das Haus der Temperamente by Johann Nestroy	82
Sonnenfinsternis by Arno Holz	80
Die Industrie-Ausstellung by Friedrich Kaiser	75
Der Vampyr by Wilhelm August Wohlbrück	74
Einen Jux will er sich machen by Johann Nestroy	74
Der Talisman by Johann Nestroy	65
Zar und Zimmermann by Albert Lortzing	65
Faust by August Klingemann	61

These are generally the works in which the above most common stage directions most often appear. In Arno Holz’s Ignorabimus, for example, the word “heftig” appears no less than 93 times. “Für sich” appears 101 times in Nestroy’s Das Haus der Temperamante.

I wanted to see which of the stage directions could be traced back to single sources, that is to say, which of them are distinctively from a single work. The 31 stage directions with a distinct, exact match in the German Drama Corpus were:

Time	Stage direction	Play
00:43	Kein Licht. (tr. No light)	Ödipus und die Sphinx by Hugo von Hoffmansthal
02:56	Noch halb im Fass. (tr. Still half in the barrel.)	Eulenspiegel by Johann Nestroy
04:00	Der Wind stöhnt. (tr. The wind moans.)	Draußen vor der Tür by Wolfgang Borchert
05:00	Springt mit allen Zeichen des Entsetzens zurück. (tr. Springs back with every indication of distress.)	Der Sohn by Walter Hasenclever
05:37	Von großer Erregung übermannt. (tr. Overwhelmed by great agitation.)	Der Sohn by Walter Hasenclever
07:32	Tod geht ab. (tr. Exit death.)	Draußen vor der Tür by Wolfgang Borchert
08:32	Übertrieben elegant. (tr. With exaggerated elegance.)	XYZ by Klabund
08:41	Immer mehr schwatzhaft und schnodderig. (tr. Ever more loquacious and flippant.)	Draußen vor der Tür by Wolfgang Borchert
08:47	Absichtlich gemein. (tr. Purposefully base.)	Draußen vor der Tür by Wolfgang Borchert
10:27	Abwechselnd lauschend und sprechend. (tr. Alternately listening and speaking.)	Meister Oelze by Johannes Schlaf
11:38	Nimmt sich das Rätsel vor. (tr. Gets to work on the riddle.)	XYZ by Klabund
12:41	Bricht disharmonisch ab. (tr. Discontinues dissonantly.)	XYZ by Klabund
13:01	Kugelfüsse. (tr. Spherical feet.)	Der Wetterfürst by Paul Scheerbart
13:35	Wird immer leiser. (tr. Grows ever quieter.)	Draußen vor der Tür by Wolfgang Borchert
13:45	Mit strengem Ton. (tr. With a harsh tone.)	Almansor by Heinrich Heine
14:10	In größter Ungeduld. (tr. Extremely impatient.)	Eulenspiegel by Johann Nestroy
14:14	Tut pikiert. (tr. Acts piqued.)	Meister Oelze by Johannes Schlaf
14:26	Robust, um nicht zu heulen. (tr. Stolidly so as not to sob.)	Draußen vor der Tür by Wolfgang Borchert
15:45	Perplex zurück. (tr. Back, perplexed)	XYZ by Klabund
16:05	Mit komischer Gravität. (tr. With comic gravity.)	Zar und Zimmermann by Albert Lortzing
16:36	Mit ungeschickter Ängstlichkeit. (tr. With clumsy anxiety.)	Eulenspiegel by Johann Nestroy
18:14	Mit den Bewegungen einer Marionette. (tr. With the movements of a marionette.)	Eulenspiegel by Johann Nestroy
18:23	Alle fahren erschrocken herum. (tr. All wheel about, shocked.)	Der Sohn by Walter Hasenclever
19:36	Auf beide los. (tr. Coming at both.)	Zar und Zimmermann by Albert Lortzing
21:29	Und ziehen sich zurück. (tr. And draw back.)	Hidalla oder Sein und Haben by Frank Wedekind
22:13	Mit affektiertem Schmerz. (tr. With affected anguish.)	Almansor by Heinrich Heine
22:43	Sich etwas erhebend. (tr. Rising somewhat.)	Aslauga by Friedrich de La Motte Fouqué
23:58	Alle in größter Verwirrung ab. (tr. All exit in great confusion.)	Eulenspiegel by Johann Nestroy
28:09	Fällt rauschend ein. (tr. Collapses noisily.)	Zar und Zimmermann by Albert Lortzing
29:04	Buntes Gewühl. (tr. Pied turmoil.)	Almansor by Heinrich Heine
29:52	Burleskes Ballett. (tr. Burlesque ballet.)	Almansor by Heinrich Heine

I then wanted to see which plays have more than one distinctive line in Frembdsch. They were:

Play	Count
Draußen vor der Tür by Wolfgang Borchert	6
Eulenspiegel by Johann Nestroy	5
XYZ by Klabund	4
Almansor by Heinrich Heine	4
Zar und Zimmermann by Albert Lortzing	3
Der Sohn by Walter Hasenclever	3
Meister Oelze by Johannes Schlaf	2

It seems Borchert, Nestroy, Klabund, Heine, Lortzing, Hasenclever, and Schlaf, entered Frembdsch relatively unharmed. Note that only one of these plays (Lortzing’s Zar und Zimmermann) is present in the previous list of works in which stage directions from Frembdsch most often appear. Only two authors are represented in both lists: Lortzing and Nestroy. Nestroy also appeared three separate times in the previous list, suggesting a general stylistic affinity between his stage directions and those of Frembdsch.

A closer look into Nestroy’s Eulenspiegel, the play with the second-highest number of distinctive lines in Frembdsch, reveals narrative borrowings that would otherwise not be captured by this strategy of exact matching. Reminiscent of Beckett’s Endgame, the narrative, as it is, of Frembdsch begins and ends with stage directions indicating a character exiting and then again entering a barrel. For example:

Time	Stage direction
01:05	In der Ecke steht ein Fass. (tr. A barrel stands in the corner.)
01:10	Lauscht ein Kopf heraus. (tr. A head listens out. [sic!])
02:56	Noch halb im Fass. (tr. Still half in the barrel.)
03:28	Aus dem Fass steigend. (tr. Climbing out of the barrel.)

Then near the end:

Time	Stage direction
31:52	Lauscht in das Fass. (tr. Listens into the barrel.)
31:56	Steigt ins Fass. (tr. Climbs into the barrel.)
31:59	Bückt sich, schlägt Deckel zu. (tr. Bends, closes the lid.)

The only exact match with the German Drama Corpus is “Noch halb im Fass.” This comes from Nestroy’s Eulenspiegel, a play with a great deal of business with people in barrels. A closer look into Eulenspiegel reveals two near misses: instead of “Steigt ins Fass,” Nestroy’s text has “Steigt ins Faß,” and instead of “aus dem Fass steigend,” Nestroy’s text has “aus dem Fasse steigend.” It seems that Nestroy’s work, and Eulenspiegel in particular, is disproportionately represented in Frembdsch.

Vossian Antonomasia Extraction Using Pre-Trained Language Models

2022-11-23T00:00:00+00:00

Hi there, let’s get back to one of our favorite topics: Vossian Antonomasia (VA). In our 2019 EMNLP-IJCNLP paper, we tried to detect VA automatically using rule-based methods and a simple neural network approach. The latter showed very promising results, so we continued along this line and brought in some heavier machinery – the Michael Jordans in the field of natural language processing (NLP): pre-trained language models (PLMs).

Neural networks and especially PLMs like BERT have shown that they can improve a wide range of NLP tasks, especially those for which large labeled datasets are not available.

The advantage of language models is the pre-training, which is conducted with large amounts of unlabeled text data. For example, BERT was trained on the English Wikipedia and BooksCorpus. In the pre-training phase, the model learns a basic understanding of language that can be used further for downstream tasks, for example, named entity recognition (NER), sentiment analysis, or part-of-speech tagging.

We decided to try this on Vossian Antonomasia for our latest paper published in Frontiers in Artificial Intelligence. Instead of classifying complete sentences, that is, deciding whether a sentence contains a VA expression or not, we reformulated the task. Now, the machine learning model is trained to identify all parts of a VA within a sentence, that is, the source, target and modifier, and distinguish them from one another. This is called sequence tagging.

Words:	A	Spice	Girls	of	hip-hop	,	the	Wu-Tang	Clan	offers	something	for	every	kind	of	rap	fan
Tags:	-	B-SRC	I-SRC	-	B-MOD	-	B-TRG	I-TRG	I-TRG	-	-	-	-	-	-	-	-

For the training of neural networks, we annotated our VA dataset on the word-level, that is we marked for each word in 3,066 sentences whether it is part of a source, target, or modifier or does not belong to a VA at all.

We then used the BERT base model and fine-tuned it with the annotated data. In particular, we added an additional layer on top of the BERT model that computes a tag for each word of the input sentence. Then the parameters were re-computed based on the input data.

In addition, we also trained a neural network model from scratch using a concatenation of ELMo and GloVe embeddings and a bidirectional long short-term memory (BLSTM) neural network with a conditional random field (CRF) on top that also tags each word.

Task	Approach	Precision	Recall	F1
Classification	Baseline	0.876	0.880	0.878
	BLSTM-ATT	0.921	0.074	0.947
	BERT-CLF	0.971	0.977	0.974
Sequence-Tagging	BASELINE	0.765	0.616	0.682
	BLSTM-CRF	0.908	0.907	0.907
	BERT-SEQ	0.908	0.944	0.926

First, we could improve the sentence classification task by almost 0.1 points in F1 score. Also, we could achieve strong results (0.93 in F1 score) on the new sequence tagging task, where BERT outperforms the BLSTM-CRF model.

In addition to the evaluation on our annotated dataset, we conducted a robustness study on real-world newspaper data. We also studied the ability of the model to predict new types of VA focussing on new types of source entities (e.g., organizations, locations, fictional characters) and on new syntactic variations around the source (e.g., “a SOURCE on”, “of SOURCE of”). In total, the model identified around 10,000 VA candidates in the NYT corpus that our previous models were not able to find. The following table shows the most frequently predicted source candidates. Due to limited capacity, we only evaluated samples of these candidates.

Source Candidates	Count
Holy Grail	116
Cadillac	88
Pied Piper	85
RollsRoyce	71
Paris	60
Harvard	58
Microsoft	43
Venice	42
Demon Barber	39
King	37
Switzerland	37
McDonalds	35
Darth Vader	34
Wild West	33
Cinderella	32
Goliath	29
Woodstock	29

In summary, in our newest paper, we developed new models for extracting VAs on the word-level, that is, the models tag all words that belong to a VA expression in a sentence. In addition to the high evaluation scores on our annotated dataset, we showed in multiple robustness studies that the best model is able to predict new versions of VAs regarding syntactic variations and also types of named entities.

The full annotation and deeper analytics of the predicted candidates is one of our next projects. If you are interested in participation, feel free to contact us. Stay tuned!

Automatic Extraction of Network Data From Amazon Prime Videos (Using "1917" As an Example)

2022-02-16T00:00:00+00:00

Social networks extracted from movies or series are interesting research material. You can either download them (via pages like Moviegalaxies) or create them yourself (e.g. using our tool ezlinavis). They are also well suited as a basis for tool-based introductory courses in network analysis.

Often, however, we don’t know exactly how the data came about (for example, there is still the problem that Frodo appears twice in the network extracted from the first of the three “Lord of the Rings” movies on Moviegalaxies – how did that happen?). Another problem is that sometimes it just takes too long to extract your own network data in a scene-by-scene fashion using ezlinavis.

So it might come in handy if we introduce another way to retrieve network data.

If you happen to have an Amazon Prime Video account (or know someone who has 😊), you may have noticed the X-Ray feature. When you press the ⏸️ button, you see background information about the current production, including information about the characters/actors currently on screen.

Turns out that this information is stored in a JSON file that you can download for each movie/series in a roundabout way. Once you have this file, you can do some analysis with it, like visualising the screentime of characters or extracting co-presence networks, which is really neat. (One advantage is that you can very easily compile datasets on your own favourite movies or series, so long as these productions are available on Prime Video.)

To start with a visual argument, here is the result for Sam Mendes’ movie “1917”, followed by a description of how we did it:

We are far from being the first to discover this possibility. In fact, there’s a blog post from 2016 on this topic – and although things usually change all the time on the internet, the method described still works almost 6 years later.

The first step is to download the JSON file using your browser’s built-in developer tools.

The resulting JSON file is a bit convoluted, but it is relatively easy to find your way around it. Amazon’s X-Ray data divides the entire movie into 14 scenes ("id":"/xray/scene/1" etc.):

Scene	Starts at
1	00:00:00
2	00:00:48
3	00:01:01
4	00:11:21
5	00:16:29
6	00:25:45
7	00:30:50
8	00:44:37
9	00:54:57
10	01:06:33
11	01:19:56
12	01:22:59
13	01:38:20
14	01:49:32

These scene divisions are, of course, contingent and can also lead to rather poor results for some movies. But for “1917”, a station drama, it works quite well.

Let’s look at Scene 6, which features only two characters, Schofield and Blake:

{
  "changesCollection":[
    {
      "changeType":"AddItem",
      "itemId":"/name/nm1126657/Lance Corporal Schofield",
      "timePosition":1548000
    },
    {
      "changeType":"AddItem",
      "itemId":"/name/nm2835616/Lance Corporal Blake",
      "timePosition":1560000
    }
  ],
  "initialItemIds":[],
  "timeRange":{
    "endTime":1850000,
    "startTime":1545000
  }
}

As you can see, in addition to the mere co-presence in a scene, there is also a temporal marker for the appearance of a character. But if you think this would make it easy to measure the overlap of the presence of two characters, alas, no. Because, unfortunately, there are only markers for their entrance, not their exit, so a character is always “present” until the end of a scene.

To extract the co-presence network data, we wrote a little Python script (based on the example here). We also included shared scenes as edge labels and fed all this into Gephi, the result of which you have seen above.

Conclusion

We don’t know how Amazon Prime’s X-Ray data is encoded in practice, so we can’t assume that all JSON files were created in a similar process. But the fact that we can read this JSON data relatively easily, despite the lack of documentation, allows us to do some meaningful things with it.

And there’s a lot more we could do with the X-Ray data. For example, the JSON also contains IMDb IDs of actors (e.g. "id":"/name/nm1126657/Lance Corporal Schofield" → imdb.com/name/nm1126657/). But for now, as Ed Wood used to say so enthusiastically, that’s a wrap!

Quantification of Scholarly Articles on German Drama

2021-10-12T00:00:00+00:00

Context

In our project “What matters? Key passages in literary works” we analyse how scholarly articles cite literary works and whether these citations can be used to identify key passages. We started with a corpus of 100 scholarly articles dealing with the interpretion of one of two novellas:

Michael Kohlhaas by Heinrich von Kleist (1808)
Die Judenbuche by Annette von Droste-Hülshoff (1842)

In the next phase of our project, we want to focus on German-language drama and are therefore building a corpus of scholarly articles dealing with German plays. For our distant-reading approach, this corpus should be substantial, that is, for each play we would like to have a number of scholarly articles dealing with it. So one of our questions for operationalisation was: “Which plays should we pick so that we find a reasonable number of scholarly works per play?” Or more generally: “Which German plays have been interpreted most frequently by researchers?”

Data Sources

To answer this question, we utilise two data sources:

GerDraCor, the German Drama Corpus, which is part of the larger DraCor project and currently comprises 545 German plays from the period 1657¹ to 1947². DraCor provides the full texts of all plays (which is crucial for our project) as well as detailed metadata such as title, author and year of publication.
The BDSL online catalogue, short for Bibliographie der deutschen Sprach- und Literaturwissenschaft (Bibliography of German Linguistics and Literature), a comprehensive bibliography of more than 300,000 scholarly works on German language and literature published between 1985 and 2010. It offers broad search options, in particular it is possible to search for articles that refer to a specific work (by title) or an author (by name).

Implementation

1. We downloaded a GerDraCor snapshot on August 31, 2021, by cloning its git repository (git clone git@github.com:dracor-org/gerdracor.git).

2. Using a small Python script, we extracted titles (XPath: tei:teiHeader/tei:fileDesc/tei:titleStmt/tei:title) authors (XPath: tei:teiHeader/tei:fileDesc/tei:titleStmt/tei:author) of each play. The resulting TSV file contains 545 plays. The name of authors is stored in the format “Surname, Forename”, titles of plays separated from it by a tab character.

Here the first entries of the resulting list:

Alberti, Konrad   Brot!
Alberti, Konrad   Im Suff
André, Johann   Der Comödienfeind
Anzengruber, Ludwig   Das vierte Gebot
Anzengruber, Ludwig   Der Gwissenswurm
[...]

3. Looping over the list of plays, we queried BDSL on the same day, August 31, 2021. It is possible to search for “Behandeltes Werk” (work treated), which allows us to restrict the search to scholarly articles whose subject is a specific work (using the title of plays). As some titles are ambiguous, initial tests have shown that we need to further restrict the search to “Behandelte Person” (person treated), which names the author of the work.

⚠️ There are a few limitations to our approach: We do limit our search to certain fields, but ultimately rely on the string matching of BDSL’s search engine, so we miss some hits. If, for example, we search for Wagner’s “Tannhäuser” using its short rather than its full title (“Tannhäuser und Der Sängerkrieg auf Wartburg”), we would arrive at 16 hits instead of just one. BDSL also employs keywords for works, using the keyword “Wagner, Richard / Tannhäuser und der Sängerkrieg auf Wartburg” we even get 44 hits. All in all, working with the BDSL interface is a bit opaque, interoperable IDs for both authors and works (aligned with the Gemeinsame Normdatei GND) would be a crucial improvement here.

Results per Play

For the following 172 plays we were able to find an article (that is, for 373 we could not):

Play	Author	Articles
Faust	Johann Wolfgang Goethe	2037
Dantons Tod	Georg Büchner	309
Nathan der Weise	Gotthold Ephraim Lessing	307
Penthesilea	Heinrich von Kleist	304
Iphigenie auf Tauris	Johann Wolfgang Goethe	243
Emilia Galotti	Gotthold Ephraim Lessing	193
Prinz Friedrich von Homburg	Heinrich von Kleist	172
Die Räuber	Friedrich Schiller	171
Wilhelm Tell	Friedrich Schiller	147
Der zerbrochne Krug	Heinrich von Kleist	143
Maria Stuart	Friedrich Schiller	123
Leonce und Lena	Georg Büchner	116
Die Jungfrau von Orleans	Friedrich Schiller	116
Die Hermannsschlacht	Heinrich von Kleist	114
Elektra	Hugo von Hofmannsthal	111
Torquato Tasso	Johann Wolfgang Goethe	109
Kabale und Liebe	Friedrich Schiller	102
Die letzten Tage der Menschheit	Karl Kraus	100
Amphitryon	Heinrich von Kleist	97
Reigen	Arthur Schnitzler	85
Der Schwierige	Hugo von Hofmannsthal	77
Die Zauberflöte	Emanuel Schikaneder	71
Miß Sara Sampson	Gotthold Ephraim Lessing	70
Egmont	Johann Wolfgang Goethe	66
Die Familie Schroffenstein	Heinrich von Kleist	57
Die Soldaten	Jakob Michael Reinhold Lenz	50
Der Rosenkavalier	Hugo von Hofmannsthal	47
Frühlings Erwachen	Frank Wedekind	47
Draußen vor der Tür	Wolfgang Borchert	45
Die natürliche Tochter	Johann Wolfgang Goethe	45
Ariadne auf Naxos	Hugo von Hofmannsthal	44
Judith	Friedrich Hebbel	43
Maria Magdalene	Friedrich Hebbel	43
König Ottokars Glück und Ende	Franz Grillparzer	42
Die Verschwörung des Fiesco zu Genua	Friedrich Schiller	41
Die Jüdin von Toledo	Franz Grillparzer	40
Der Turm	Hugo von Hofmannsthal	40
Philotas	Gotthold Ephraim Lessing	39
Libussa	Franz Grillparzer	37
Jedermann	Hugo von Hofmannsthal	36
Stella	Johann Wolfgang Goethe	35
Die Frau ohne Schatten	Hugo von Hofmannsthal	34
Geschichten aus dem Wiener Wald	Ödön von Horváth	34
Vor Sonnenaufgang	Gerhart Hauptmann	31
Die Büchse der Pandora	Frank Wedekind	31
Die Juden	Gotthold Ephraim Lessing	30
Ein Bruderzwist in Habsburg	Franz Grillparzer	29
Die Hermannsschlacht	Christian Dietrich Grabbe	28
Herzog Theodor von Gothland	Christian Dietrich Grabbe	28
Die Schwärmer	Robert Musil	27
Anatol	Arthur Schnitzler	26
Der grüne Kakadu	Arthur Schnitzler	26
Die Ratten	Gerhart Hauptmann	25
Agnes Bernauer	Friedrich Hebbel	25
Liebelei	Arthur Schnitzler	23
Der gestiefelte Kater	Ludwig Tieck	23
Don Juan und Faust	Christian Dietrich Grabbe	21
Arabella	Hugo von Hofmannsthal	21
Der Unbestechliche	Hugo von Hofmannsthal	21
Kasimir und Karoline	Ödön von Horváth	21
Professor Bernhardi	Arthur Schnitzler	21
Clavigo	Johann Wolfgang Goethe	20
Proserpina	Johann Wolfgang Goethe	20
Gyges und sein Ring	Friedrich Hebbel	20
Der Marquis von Keith	Frank Wedekind	20
Scherz, Satire, Ironie und tiefere Bedeutung	Christian Dietrich Grabbe	19
Der Traum ein Leben	Franz Grillparzer	19
Weh dem, der lügt!	Franz Grillparzer	19
Cardenio und Celinde oder Unglücklich Verliebete	Andreas Gryphius	19
Hermanns Schlacht	Friedrich Gottlieb Klopstock	18
Die Kindermörderin	Heinrich Leopold Wagner	18
Der Tod des Tizian	Hugo von Hofmannsthal	17
Der Talisman	Johann Nestroy	17
Das weite Land	Arthur Schnitzler	17
Des Meeres und der Liebe Wellen	Franz Grillparzer	15
Herodes und Mariamne	Friedrich Hebbel	15
Die Wupper	Else Lasker-Schüler	15
Die Gründung Prags	Clemens Brentano	14
Ödipus und die Sphinx	Hugo von Hofmannsthal	14
Das Liebeskonzil	Oskar Panizza	14
Die Ahnfrau	Franz Grillparzer	13
Das Salzburger große Welttheater	Hugo von Hofmannsthal	13
Die Zwillinge	Friedrich Maximilian Klinger	13
Der Zerrissene	Johann Nestroy	13
Der Triumph der Empfindsamkeit	Johann Wolfgang Goethe	12
Hannibal	Christian Dietrich Grabbe	12
Medea	Franz Grillparzer	12
Judith und Holofernes	Johann Nestroy	12
Erdgeist	Frank Wedekind	12
Ponce de Leon	Clemens Brentano	11
Ein treuer Diener seines Herrn	Franz Grillparzer	11
Der Sohn	Walter Hasenclever	11
Genoveva	Friedrich Hebbel	11
Julius von Tarent	Johann Anton Leisewitz	11
Der einsame Weg	Arthur Schnitzler	11
Die Journalisten	Gustav Freytag	10
Sappho	Franz Grillparzer	10
Blunt oder der Gast	Karl Philipp Moritz	10
Alceste	Christoph Martin Wieland	10
Der letzte Held von Marienburg	Joseph von Eichendorff	9
Almansor	Heinrich Heine	9
Die Familie Selicke	Arno Holz	9
Zriny	Theodor Körner	9
Komödie der Verführung	Arthur Schnitzler	9
Leben und Tod der heiligen Genoveva	Ludwig Tieck	9
Florian Geyer	Gerhart Hauptmann	8
Sturm und Drang	Friedrich Maximilian Klinger	8
Der junge Gelehrte	Gotthold Ephraim Lessing	8
Einen Jux will er sich machen	Johann Nestroy	8
Datterich	Ernst Elias Niebergall	8
Der Alpenkönig und der Menschenfeind	Ferdinand Raimund	8
Wallensteins Lager	Friedrich Schiller	8
Alarcos	Friedrich Schlegel	8
Die verkehrte Welt	Ludwig Tieck	8
Franziska	Frank Wedekind	8
Ugolino	Heinrich Wilhelm von Gerstenberg	7
Der Engländer	Jakob Michael Reinhold Lenz	7
Wallensteins Tod	Friedrich Schiller	7
Canut	Johann Elias Schlegel	7
Halle	Achim von Arnim	6
Jerusalem	Achim von Arnim	6
Die Freier	Joseph von Eichendorff	6
Des Epimenides Erwachen	Johann Wolfgang Goethe	6
Demetrius	Friedrich Hebbel	6
Freiheit in Krähwinkel	Johann Nestroy	6
Moral	Ludwig Thoma	6
Faust	Friedrich Theodor Vischer	6
Perdu! oder Dichter, Verleger und Blaustrümpfe	Annette von Droste-Hülshoff	5
Erwin und Elmire	Johann Wolfgang Goethe	5
Babel und Bibel	Karl May	5
Magdalena	Ludwig Thoma	5
Die Laune des Verliebten	Johann Wolfgang Goethe	4
Die Mitschuldigen	Johann Wolfgang Goethe	4
Satyros oder der vergötterte Waldteufel	Johann Wolfgang Goethe	4
Der Tod Adams	Friedrich Gottlieb Klopstock	4
Der Diamant des Geisterkönigs	Ferdinand Raimund	4
Der Verschwender	Ferdinand Raimund	4
Der Kammersänger	Frank Wedekind	4
Hidalla oder Sein und Haben	Frank Wedekind	4
Codrus	Johann Friedrich von Cronegk	3
Merlin	Karl Immermann	3
Tristan und Isolde	Richard Wagner	3
Die Zensur	Frank Wedekind	3
Kaiser Friedrich Barbarossa	Christian Dietrich Grabbe	2
Uriel Acosta	Karl Gutzkow	2
Gabriel Schillings Flucht	Gerhart Hauptmann	2
Colberg	Paul Heyse	2
Alkestis	Hugo von Hofmannsthal	2
Ein Trauerspiel in Berlin	Karl von Holtei	2
Die Koralle	Georg Kaiser	2
Der Kreidekreis	Klabund	2
Der Erbförster	Otto Ludwig	2
Das Haus der Temperamente	Johann Nestroy	2
Der Barometermacher auf der Zauberinsel	Ferdinand Raimund	2
Die Piccolomini	Friedrich Schiller	2
Der Bettler	Reinhard Sorge	2
Heimat	Hermann Sudermann	2
Musik	Frank Wedekind	2
Johann Faust	Paul Weidmann	2
Die Nase des Michelangelo	Hugo Ball	1
Der blaue Boll	Ernst Barlach	1
Narziß	Albert Emil Brachvogel	1
Brutus	Joachim Wilhelm von Brawe	1
Die Frau im Fenster	Hugo von Hofmannsthal	1
Simsone Grisaldo	Friedrich Maximilian Klinger	1
Franz von Sickingen	Ferdinand Lassalle	1
Die Makkabäer	Otto Ludwig	1
Golo und Genovefa	Maler Müller	1
Der Weibsteufel	Karl Schönherr	1
Ernst Herzog von Schwaben	Ludwig Uhland	1
Tannhäuser und Der Sängerkrieg auf Wartburg	Richard Wagner	1
Faust	Hermann Ludwig Wolfram	1

Although it is no surprise to find Goethe’s “Faust” with all its variants in first place, it is surprising that the second-placed play, Büchner’s “Danton’s Death”, lags behind by an order of magnitude in terms of the articles in which the play is discussed.

Results per Author

To see which author was written about most often, we can group and sort the result by author:

Author	Plays	Articles	Mean Articles per Play
Johann Wolfgang Goethe	22	2610	118
Heinrich von Kleist	7	887	126
Friedrich Schiller	11	717	65
Gotthold Ephraim Lessing	12	647	53
Hugo von Hofmannsthal	17	478	28
Georg Büchner	2	425	212
Franz Grillparzer	13	247	19
Arthur Schnitzler	13	218	16
Friedrich Hebbel	13	163	12
Frank Wedekind	12	131	10

We can also sort this table by the average number of articles per play per author, although the result should be treated with caution for several reasons:

Author	Plays	Articles	Mean Articles per Play
Georg Büchner	2	425	212
Heinrich von Kleist	7	887	126
Johann Wolfgang Goethe	22	2610	118
Emanuel Schikaneder	1	71	71
Friedrich Schiller	11	717	65
Gotthold Ephraim Lessing	12	647	53
Karl Kraus	2	100	50
Wolfgang Borchert	1	45	45
Hugo von Hofmannsthal	17	478	28
Robert Musil	1	27	27

On the one hand, authors who have become known primarily for just one or a few plays appear at the top (Karl Kraus, Wolfgang Borchert, Georg Büchner). However, we only looked for plays that are already represented in GerDraCor. In addition to Büchner’s two plays “Leonce and Lena” and “Danton’s Death”, his famous fragment “Woyzeck” is missing, which limits the usefulness of the above table. All the results presented only make statements about the plays and authors currently contained in GerDraCor.

DraCor Summer Update (August 2021)

2021-08-05T00:00:00+00:00

DraCor, the Drama Corpora platform, can be accessed at dracor.org. Our aim is to provide an infrastructure component for computational literary studies in order to facilitate access to relevant research data and to ensure compliance with the FAIR Principles when dealing with literary corpora. In our case, the focus is of course on drama and theatre, but we plan to extend our concept of Programmable Corpora to novels as part of the H2020-funded CLS INFRA project that started earlier this year.

Today, we are pleased to introduce the next iteration of our DraCor website, along with new API functions that make it much easier to access our data for research. Our front-end, i.e. web design, is now at version 1.1.0, the DraCor-API at version 0.82.0.

We have worked on all levels of the website, the homepage, the corpus overviews and pages of individual plays:

As before, the website should work well on all types of devices, large screens, laptops, tablets and mobile phones. There are many small details to discover in the new design (our favourites are the author images that we automatically retrieve from Wikimedia Commons via Wikidata QIDs, just one small example of the benefits of our highly interoperable approach).

The overall design is even better aligned with the TEI encoding on which all our corpora are based to increase the visibility of our rich annotations.

Here’s a non-exhaustive summary of the new functions added since December 2020:

infoboxes are now context-sensitive (their content differs depending on which tab is activated: “Network”, “Relations”, “Speech distribution”, “Full text”, or “Downloads”)
more explanations on the spot, e.g. the “Network” tab is accompanied by a short description of what the network graph actually shows
we also introduced an ℹ️ symbol that leads to longer explanations in our newly introduced (and still somewhat empty) FAQ section
author information and images sourced live from Wikimedia Commons via Wikidata identifiers
improved readability of network graphs (annotated information on gender is now integrated, colour contrast has been increased)
plays with historical or mythological characters now display a Wikidata QID if available, a new feature provided by our API (not all corpora are yet encoded with this information, but once they are, it will be much easier to compare all plays with e.g. a Faust character or to do a contrastive analysis of the vocabulary of mythological and historical characters in different authors and even between corpora in different languages)
the “Full text” tab now shows the print source in the context box, along with an overview of speaking characters per segment
corpus metadata is now much more accessible, you can download it directly from the corpus overviews (try the CSV version to create some quick charts in MS Excel or LibreOffice Calc)
added a button to copy DraCor IDs to the clipboard
new corpus added: BashDraCor
new plays were added to our in-house corpora, GerDraCor and RusDraCor
Michael Sonkin added numerous character annotations to ItaDraCor, significantly improving the corpus which we originally inherited from Biblioteca Italiana
introduction of TypeScript (to gradually improve the quality of our codebase)

Ok, we’ll stop here. As always the complete history of changes can be traced on GitHub.

How Did We Get Here?

Our React-powered website and the DraCor API have been up and running since 2017. We started with a German and a Russian drama corpus, both encoded in TEI, but later added many more corpora, a handful of which were provided by the community. At the moment we host 12 active drama corpora, and number 13, the rather large French Drama Corpus (more than 1,500 plays so far), is just around the corner. Many colleagues have approached us asking if they can add their corpus in the future, and we very much welcome their contribution. (If you have or want to create or maintain a TEI-encoded drama corpus and add it to DraCor, please drop us a line.)

Looking back, we find it hard to believe that we didn’t initially intend to run a fancy website at all, our focus was on functionality. It wasn’t until we had developed the concept of Programmable Corpora and the website was functional and running stably that we eventually decided we wanted a fancy website too. So in December 2020, we released version 1.0.0 of the DraCor front-end, and since then the project has received much more attention. In the past it wasn’t a big problem when DraCor went offline on a Friday afternoon because of an unexpected server error, we just fixed it the next week. Well, we can’t wait that long anymore. 😊 Now we try to fix a problem as soon as it arises. If this isn’t possible, we collect more complex issues (and also feature suggestions) in the issue tracker.

Another way to measure our impact is through the increasing number of third-party research conducted with our data. The most recent example is Inna Wendell’s PhD thesis at UCLA on Russian five-act comedy in verse. A growing list of DraCor-enabled research can be found on our website. We also just set up a Zotero collection to keep track.

API

DraCor is of course more than a website, as we never tire of pointing out. The crucial element is its API. All currently active functions are documented via Swagger, but this is of course not enough, especially if you are not yet that used to dealing with APIs. We are planning some workshops in the near future to train how to use the API so that the TEI-encoded DraCor corpora can unleash their full power to answer research questions. Stay tuned.

Ways of Interaction

I’m sure we’ve forgotten this or that detail in this blogpost, but hope you enjoy exploring the new DraCor. If you run into an issue, have a question, or want to contribute, you can always check our Credits page to find the right contact person. We also use the GitHub Discussions feature for the dracor-frontend repo (all things website) and dracor-api (all things API functionality).

That’s it for today. It’s been great to do some hacking together over the last months, even if the infamous in-person Potsdam/Moscow hackathon spirit is still unmatched. See you soon.

🌞 🌴 🐬 ⛱️

Ted Underwood's "Distant Horizons": Reviews and Replications

2021-05-18T00:00:00+00:00

Ted Underwood’s most recent book – 206 pages, published in February 2019 – was widely acclaimed and much discussed. Based on the assumption that digital methods and corpora show their strengths especially in diachronic observations over long periods of time, he has provided aspects of a digital literary historiography.

But this “computational monograph” – to quote Maciej Maryl – is not just something to read. Data and code have been published and can be replicated and questioned. This is not always easy and is not yet a living part of our practice in the Digital Humanities.

That’s why this blogpost attempts to bring together the REPLICATIONS made so far to put Underwood’s technical setup to the test. Such efforts to verify scientific claims should be encouraged more and can be exemplary for similar undertakings in the future.

I’m also gathering all REVIEWS of the book I found (as neither the publisher’s website nor the author’s personal page offers such an overview). If you have any additions, please lemmino (you can find me on Twitter).

Fig. 1. Ted Underwood: Distant Horizons. Digital Evidence and Literary Change. Chicago and
London: The University of Chicago Press 2019. (doi:10.7208/chicago/9780226612973.001.0001).

Underwood’s book has a somewhat mediating character. The machine-learning methods used for the analyses are not bleeding edge, he mostly relies on regularised logistic regression on several thousand features. Underwood refutes the claim that machine-learning models are always inscrutable black boxes. He spends a lot of time looking into these black boxes and interpreting the word frequencies and their distribution in order to craft narratives relevant to literary studies.

But we are still at the beginning, as Underwood makes clear: “Statistical models can help us understand the social forces shaping a long arc of literary change. But turning those models into fully satisfying stories could take several more decades.” (p. 109)

I originally wanted to write a review myself, but never found the time to finish it (and finally got discouraged late last year when I encountered Maciej Maryl’s excellent and comprehensive review 😊). The following list of reviews I have compiled may serve to better sort through the rich discussions surrounding “Distant Horizons” and find good entry points.

REVIEWS

Dan Sinykin, Post45 (6 May 2019)
- Tess McNulty responds to his review, Post45 (6 May 2019)
Mareike Schumacher, lebelieberliterarisch.de (25 Jul 2019) 🇩🇪 – followed by a comment from Fotis Jannidis
Laura Jiménez Ríos, Revista de Humanidades Digitales, vol. 4 (2019), pp. 212–215 🇪🇸
Daniel Rosenberg, Modern Philology, vol. 117, no. 3 (Feb 2020), publ. online 22 Nov 2019
Katherine Bode, Modern Language Quarterly, vol. 81, issue 1 (1 Mar 2020), pp. 95–124 (preprint) – a more critical take
Alison Booth, Victorian Studies, vol. 62, no. 3 (Spring 2020), pp. 547–550 – also somewhat critical
Marie-Christine Boucher, KULT_online, no. 61 (April 2020)
Maciej Maryl, JLTonline (17 Oct 2020) – assigns a new genre to the book, calling it a “computational monograph”

REPLICATIONS OF CODE AND/OR DATA

When Underwood submitted his book to the publisher, he published data and code in a ZIP package on Zenodo on 25 March 2018 (doi:10.5281/zenodo.1207277). There’s also a (slightly updated) Git Hub repo.

Portions of chapters 2, 3 and 4 are reprints of prior collaboratory work. This is also the reason why most of the listed replications could take place before the book even went to press.

Andrew Goldstone, “Of Literary Standards and Logistic Regression: A Reproduction”, andrewgoldstone.com, 4 Jan 2016
- Sarah Allison wrote a short piece reflecting on Goldstone’s replication: “Other people’s data: humanities edition”, Journal of Cultural Analytics, vol. 1, issue 1 (18 Dec 2016)
Jonathan Goodwin, “Darko Suvin’s Genres of Victorian SF Revisited”, jgoodwin.net, 17 Oct 2016
Nan Z. Da, “The Computational Case against Computational Literary Studies”, vol. 45, no. 3 (Spring 2019) – among a total of 14 CLS papers, she also discusses Underwood’s “The Life Cycle of Genres”, which expanded into chapter 2 of “Distant Horizons”
- followed by Underwood’s reaction in CI’s response forum, critinq.wordpress.com, 1 Apr 2019
- and Da’s follow-up, “Errors”, critinq.wordpress.com, 2 Apr 2019
Ted Underwood’s reply to Da, “The Theoretical Divide Driving Debates about Computation”, Critical Inquiry, vol. 46, no. 4 (Summer 2020), which re-analyses material in chapter 2 using different methods

The First of May in German Literature

2021-05-04T00:00:00+00:00

[ tl;dr: This blogpost introduces a small dataset comprising 249 mentions
of the "1st of May" extracted from a corpus of German-language fiction. ]

Last Friday, German radio station Bayern 2 asked me for an interview about our research on the role of the month of May in literature (and about the practice of Digital Humanities in general, for that matter). The announcement is here – unfortunately, they didn’t put the interview online, but if anyone wants to listen to it, I can send them an mp3 copy.

I took this opportunity to revisit my earlier research on date extractions from literature, which I started with Jannik Strötgen right after we met at his poster at the 2013 Herrenhausen Conference “(Digital) Humanities Revisited”. Although our work was mostly finished by 2015, we occasionally returned to this topic due to public interest, and it was always great to come back to it since we could never exhaust all the possibilities our research data offered us because we quickly got busy with other things.

The journalists at Bayern 2 probably learned about our research through an article I wrote for Die Literarische Welt two years ago:

Fig. 1. “Als einst am 1. Mai die Welt begann” (publ. 4 May 2019).

This article draws on the main results of our study originally presented at the DHd2015 (in Graz) and DH2015 (in Sydney) conferences.

An interesting by-product of our work was the “TIWOLI” app (“Today in World Literature”) with iOS and Android versions (which we introduced in this blog in 2016). And, of course, there’s the TIWOLI Twitter bot, run by our colleagues at University of Cologne.

In 2017, we finally managed to release the corpus on which our research is based, the “Corpus of German-Language Fiction (txt)” (doi:10.6084/m9.figshare.4524680). The corpus contains 2,735 German-language prose works (mainly novels and short stories) by 549 authors, ranging from about 1510 to the 1940s (the bulk of the texts dates from 1840–1930). [On a sidenote, our corpus was also used by Severin Simmler to fine-tune a German BERT model to adapt it to the literary domain.]

To extract the mentioned month names and concrete days, we used Jannik’s cutting-edge HeidelTime engine. As you can see in the flower-ringed histogram featured on the newspaper page, the month of May is by far the most frequently mentioned month in the corpus. In addition, the 1st of May is, also by far, the most frequently mentioned day of the year, with 249 mentions in the corpus.

Here is an overview with the frequency of all days as we extracted them from the corpus (‘1’ means 0–9 occurrences, ‘2’ means 10–19 occurrences, etc., ‘+’ means 90 or more occurrences):

Fig. 2. Green fields: days mentioned 50+ times in the corpus
(slide taken from our DHd2015 slideset).

HeidelTime can extract various forms of dates, so here it has managed to extract “1. Mai”, “1. May” as well as “erste” / “ersten Mai”. Some example sentences:

“Ich telephonierte Montag nachmittag (also am 1. Mai) dem Witschi nach Gerzenstein, er möge mich in meinem Bureau aufsuchen.” (Glauser: Wachtmeister Studer)
“[…]; den 1. May 1772 reiseten wir aus Gondar ab.” (Knigge: Benjamin Noldmanns Geschichte der Aufklärung in Abyssinien)
“Du sollst doch auch wissen, daß heute der erste Mai ist, du Einsiedler!” (Dahn: Ernst und Frank)
“Am ersten Mai anno domini 1835 war zu Haimbach ein großes Frühstück.” (Stifter: Feldblumen)

Now we’ve always wanted to dig a little deeper into our dataset to go beyond mere quantification, and we’ve demonstrated this idea by qualifying all mentions of the 10th of August:

Fig. 3. Semantics of the 10th of August in our corpus
(slide with German labelling taken from our DHd2015 slideset).

About three quarters of the mentions were plot-related (i.e., something happened on a 10th of August, including date references in epistolary novels and fictitious diary entries). But more than 20% were non-plot-related date mentions, instead a reference to the 10th of August 1792, a decisive event of the French Revolution, when armed revolutionaries stormed the Tuileries Palace in Paris. This date is a constant reference point in our corpus of German fiction and makes for a high number of non-plot-related date mentions.

For our focal date, the 1st of May, this problem is even more complex. Many feasts and customs are associated with this date, such as Walpurgis Night and the Maypole Dance, and since 1889 also International Workers’ Day. All these meanings can overlap when a 1st of May is mentioned and it is very difficult to decide whether a mention of the day refers to a plot-related event or not.

To start somewhere, however, let us first measure the precision of the 249 mentions counted and say something about the reliability of the data we generated and evaluated at the time. There are only 2 false positives, the rest is a problem of corpus composition. It is meant to be a corpus with works of fiction (hence the name), but some non-fiction works have slipped in (as already detailed in the corpus description on Figshare). Some mentions are duplicates, and one work (by Strindberg) made it into the corpus of German originals, although it is a translation). Here’s an overview:

Fig. 4. Precision of ‘1st of May’ mentions extracted from our corpus:
correctly extracted from fiction: 226 – from non-fiction: 11 –
duplicates: 8 – false positive: 2 – from translations: 1.

Although we cannot say anything about recall in our exploratory study, the precision of HeidelTime is really good. We knew that before, of course, but never really detailed this. So this blogpost confirms the special role that May plays in German prose literature, even if the means we used at the time are far from being able to provide an exhaustive answer to the questions of “when literature takes place”. But it was a start.

You can explore our findings yourself in an enhanced subset of our data, which we never formally published before. So here is a list of 1st of Mays mentioned in German fiction (in the ‘timex value’ column, HeidelTime automatically tried to assign years from the surrounding context of a mentioned date, which was not always successful, but we left this column in, even though we never used this data). The file format is CSV:

🌿 sentences-containing-first-of-may.csv 🌱

Vossian Antonomasia, Next Level

2020-11-09T00:00:00+00:00

Our last blogpost on Vossian Antonomasia (VA, or, vossanto) appeared almost two years ago. Since then, we were busy improving our methods, writing papers, and updating our web page with some more data to explore. In this article we summarise the latest developments.

Interactive Timeline for Data Exploration

In the past year, our research group at IBI conducted a small code sprint to create an interface for the visual exploration of Vossian Antonomasia extracted from our New York Times working corpus (covering the years 1987–2007). Thanks to Sjaak Priester and his excellent JavaScript Dateline widget we were able to quickly set up an interactive timeline:

It assembles all Vossian Antonomasia extracted with the method described in our 2019 EMNLP-IJCNLP paper. Each VA is represented by a colour dot and the named entity providing the source (whose properties or qualities are transferred to another named entity by the magic of Vossian Antonomasia). The colour of the dot indicates the New York Times desk responsible for the corresponding article. Clicking on an entry will display more information, as shown in the above screenshot:

a photo automatically retrieved from Wikimedia Commons (if available),
the sentence containing the VA expression with source and modifier highlighted,
the article ID with a direct link to its full text,
the name of the author,
the name of the desk,
source and licence for the photo,
a permanent link to the corresponding VA expression.

Feel free to scroll through the timeline or just follow this example.

In the upper left corner you can find a box for full-text search. After typing in a few letters it shows all matching VA expressions. Clicking on a match scrolls to the corresponding place in the timeline and shows the details, a very convenient way to find interesting Vossian Antonomasia. For example, type in “literature” to acquaint yourself with “the Tupac Shakur of American literature” or “the Madonna of Cuban literature”.

Automatic Detection of Vossian Antonomasia

Our first approach to VA extraction (published in Digital Scholarship in the Humanities 35.1, April 2020) was semi-automated, sentence-based, and strictly centered around humans. We used regular expressions to extract all sentences featuring our VA source pattern “the ENTITY of”. Then our core idea was to use Wikidata as knowledge base for distant supervision. We only kept candidates that included the exact name or alias of a Wikidata entity which had to be an instance of the class human. We used a manually created blacklist to exclude candidates like “the House of” (for ‘House’ being, among other things, a documented alias of botanist Homer Doliver House).

In our 2019 EMNLP-IJCNLP paper we automated the process of extracting VA from large newspaper corpora and compared three new approaches to our initial approach:

The second approach is an extension of our previous method. We replaced the manually curated blacklist with a popularity measure to identify candidates that could be removed after linking an entity to a Wikidata entry. We compared the ‘popularity’ (for which we took the number of Wikidata sitelinks per entity) of a human entity to another entity featuring the same label. For example, the human entity ‘House’ (the botanist) only has 9 sitelinks while the entity ‘House’ (the building) has 178 sitelinks. We removed such candidates since it was unlikely that the label was linked to the correct entity. We also removed all candidates where the source (e.g., ‘Prince’) together with multiple subsequent words (e.g., ‘of Wales’) matched the name or alias of another Wikidata entity. This allowed us to get rid of frequent false positives like the Prince of Wales (who, grammatically, could also be an aspiring Welsh singer reminding of the Artist Formerly Known as Prince).
Since we focus on people as VA sources, our third approach is based on named-entity recognition (NER). Instead of using Wikidata to detect sentence candidates, we now tried the Stanford NER tool to detect entities. We also applied the last step from the first approach to detect false positives.
In our fourth approach we leveraged the annotations from our initial approach to train a neural network. First, we transformed each word of a sentence into a word vector using pre-trained word embeddings (GloVe). Then we fed the vectors into a neural network – a bi-directional long short-term memory layer (BLSTM) with a feed-forward layer – to let it learn to distinguish sentences that do contain a VA expression from those that don’t.

In our case, evaluating the different approaches is especially tricky due to the rarity of the phenomenon. It is unfeasible to determine the recall for the 1.85 million articles in our NYT corpus. Therefore we calculate precision, recall and f1 based on the labelled corpus. The BLSTM performs best, boosting the precision to 87%:

approach	precision	recall	f1
1st: semi-automated	49.8%	–	–
2nd: Wikidata	67.3%	93.0%	78.1%
3rd: NER	71.8%	81.3%	76.2%
4th: BLSTM	86.9%	85.3%	86.1%

We are currently working on various approaches to detect all three parts (source, target and modifier) of a VA expression in a sentence. As a by-product, we will have an enriched corpus where all parts of a VA expression are tagged in each positively labelled sentence.

Modifiers

Since we need labelled data to train and evaluate our machine-learning approaches, we examined more than 3,000 VA expressions to annotate their targets and modifiers. As a result, we can now present more reliable statistics on the most common modifiers in our corpus. The ten most frequent modifiers are:

count	modifier
56	his day
34	his time
29	Japan
17	China
16	tennis
16	his generation
16	baseball
14	her time
13	our time
13	her day

A longer list with many more results can be found on our most recent statistics page. As described in our first paper, this list illustrates an important function of antonomasia, which is the “inculturation” of lesser known phenomena (cf. also Holmqvist/Płuciennik 2010, p. 379).

By filtering the modifiers we can create rankings for specific categories; for example, countries …

count	country
29	Japan
17	China
10	Brazil
8	Iran
7	Mexico
7	Israel
7	India
4	South Africa
4	Poland
3	Spain

… or sports:

count	sports
16	tennis
16	baseball
11	hockey
10	basketball
8	golf
8	football
6	soccer
6	racing
3	women’s basketball
3	sailing
3	auto racing
2	pro football
2	New York baseball
1	Yale football fame
1	women’s hockey
1	women’s college soccer
1	this year’s national collegiate basketball tournament
1	the tennis tour
1	the tennis field
1	the soccer set
1	the racing world
1	the Olympic hockey tournament
1	stock-car racing
1	Rotisserie baseball
1	pro football owners
1	professional basketball coaches
1	professional basketball
1	motocross racing in the 1980’s
1	micro golfers
1	major league baseball
1	Laser sailing
1	Japanese baseball
1	Iraqi soccer
1	horse racing
1	hockey in the former Soviet Union
1	hockey commentary
1	high school baseball in New York
1	harness racing
1	golf criticism
1	football teams
1	football owners
1	football announcers
1	European hockey
1	country-club golf
1	college football underclassmen
1	college football these days
1	college football
1	college basketball
1	Chinese baseball
1	Brazilian basketball for the past 20 years
1	BMX racing
1	biddy basketball
1	basketball announcers
1	basketball analysts
1	basketball analysis
1	baseball’s new era
1	baseball managers
1	baseball executives
1	baseball collections
1	baseball cards

Most modifiers are short and consist of only one to three words, as this plot of the distribution of the length (number of words) of modifiers shows:

The longest modifier we have found so far contains 25 words:

“And while he modestly demurs, Mr. Barker is widely regarded as the Bob Fosse of the carefully choreographed event that consumes Midtown Manhattan with tin whistles, step dancers and some two million spectators on that invariably brisk March 17 morning.” (source: 2001/03/07/1276052)

JSON Data Dump

As additional outcome of the timeline, our dataset is now also available in JSON format. The enriched JSON data dump contains links to Wikidata and to images of VA sources on Wikimedia Commons (including licence information), so that they can easily be reused for further analysis or demos. The following excerpt shows a JSON entry for an exemplary VA expression:

{
   "id": "0039183_0",
   "date": "1987-05-10",
   "sourceId": "Q9696",
   "sourceLabel": "John F. Kennedy",
   "sourceImId": "John_F._Kennedy,_White_House_color_photo_portrait.jpg",
   "sourceImThumb": "thumb/c/c3/John_F._Kennedy%2C_White_House_color_photo_portrait.jpg/180px-John_F._Kennedy%2C_White_House_color_photo_portrait.jpg",
   "sourceImLicense": "pd",
   "fId": "1987/05/10/0039183",
   "aUrlId": "9B0DE3DD1E3EF933A25756C0A961948260",
   "text":
   "In the 60's Mr. Bernstein looked like *the John F. Kennedy of* /culture/.",
   "author": "Botstein, Leon", "desk": "Book Review Desk"
}

The fields contain the following information:

key	content
id	A unique (within the dataset) identifier for a VA expression.
date	Publication date of the corresponding NYT article.
sourceId	The Wikidata ID of the VA source.
sourceLabel	The (English) Wikidata label of the source.
sourceImId	The name of the source’s image on Wikimedia Commons.
sourceImThumb	The path to the source’s image on Wikimedia Commons.
sourceImLicense	The licence of the source’s image.
fId	The ID of the article’s file in the NYT dataset.
aUrlId	The ID of the article as part of its URL (`http://query.nytimes.com/gst/fullpage.html?res=<HERE>`)
text	The sentence containing the VA expression (including org-mode markup).

Future Endeavours

We are very interested in analysing Vossian Antonomasia in languages other than English (especially German) and are looking for suitable corpora. In general, our aim is to better understand Vossian Antonomasia and its usage, distribution and variety.

As always, all details can be found on our project website https://vossanto.weltliteratur.net/.

Tools Mentioned in DH2020 Abstracts

2020-07-23T00:00:00+00:00

Building on our previous attempts to extract digital tools mentioned in abstracts of Digital Humanities conferences, we conducted a similar exercise for this week’s virtual DH2020. We again used our simple string-matching programme ToolXtractor and an updated list of tools in the TAPoR database.

So here is an overview of tools mentioned in more than one abstract (out of a total of 475 abstracts for all conference formats), including links to the actual texts – followed by some observations at the bottom of this blog post:

spaCy (5)

word2vec (5)

Gephi (4)

Natural Language Toolkit (4)

Voyant Tools (4)

Drupal (3)

Google Maps (3)

Tesseract (3)

ArcGIS (2)

Cytoscape (2)

MALLET (2)

Neo4j (2)

Omeka (2)

OpenStreetMap (2)

Some Acknowledgements and Observations

Programming languages were excluded in this list (for a reality check: Python is mentioned in 26 abstracts, JavaScript in 11, MySQL in 4). Also, no full-text, data or code repositories like HathiTrust, Google Books, GitHub, Dataverse or Islandora were included.
Disambiguation was needed: The stylo R package didn’t make it into the list, although the term ‘stylo’ is mentioned in two abstracts. But the second time it did not refer to the well-known stylometry library, but to another tool.
So, hm, only 14 tools that were explicitely mentioned more than once. Our guess is still that not all tools that contributed to a research project or workshop were mentioned, which makes it more difficult to understand how things were done. One reason for this could be the limited writing space for conference abstracts, but there is definitely room for improvement in terms of specifically mentioning the tools used.
The TAPoR list of tools, even if updated, still isn’t (and never will be) exhaustive, so as usual, take this with a grain of salt.

Network Modelling "The Last Days of Mankind"

2020-05-07T00:00:00+00:00

Last level. Final boss. Arguably the longest play ever written: “The Last Days of Mankind”, brought to light between 1915 and 1922 by your favourite Austrian crank: writer and journalist Karl Kraus (1874–1936).

More than 600 pages in the German original, five acts, prologue and epilogue, 220 scenes altogether documenting the apocalypse of World War I. A first complete translation to English was published at Yale University Press in 2015 (done by Fred Bridgham and Edward Timms).

As Kraus notes in the preface:

“The performance of this drama, which would take some 10 evenings in terrestrial time, is intended for a theatre on Mars. Theatre-goers on planet earth would find it unendurable.”

Now what did we do? Dialogues in the play are distributed among a total of 925 speaking instances (individual characters, voices, groups, choirs). We extracted the social network graph (based on co-occurrences of characters per scene) from the TEI version of the play. And then we took Kraus at his word and projected the network graph onto planet Mars. Resulting in our poster contribution for DHd2020 held in Paderborn in early March (one of the last conferences before coronavirus lockdown):

Fig. 1. "Visit to the Mars Theatre." (DOI:10.6084/m9.figshare.11917902)

Our visit to the Mars theatre won us the Best Poster Award, which we were very happy about. Feel free to download the hi-res version from Figshare (beware, it’s heavy). There’s also a lenghty Twitter thread about our work (in German).

Some properties of the network:

Number of nodes: 925
Number of edges: 12.425
Network density: 0,029 (very low in comparison to other plays)
Number of independent clusters: 120
Clustering coefficient: 0,976 (extremely high, corresponds to the value of other station dramas)

If you want to toy around with the network data yourself, you can download it in CSV, GEXF and GraphML format from DraCor.

Measuring the Impact of DH Conference Abstracts: The Case of DH2016

2020-02-17T00:00:00+00:00

We are still toying around with the proceedings of ADHO’s Digital Humanities conferences. This time we are interested in the measurability of the scientific impact of published abstracts.

Long story short, since no other meaningful metrics are at hand (or are there?), we have chosen Google Scholar’s citation counts. As a test case we decided to go with the abstracts from DH2016 in Kraków, because 1.) they are well covered in Google Scholar and 2.) it has been three and a half years since the conference, so that enough time has passed to actually have an impact and attract citations.

Our starting point was the fabulous DBLP, a giant computer science bibliography founded at University of Trier. Fortunately they also cover ADHO conferences since some time, so here are all contributions to DH2016 in a concise and interoperable form: https://dblp1.uni-trier.de/db/conf/dihu/dh2016.html.

Before we dive into the data, some known issues:

Google Scholar is “notoriously hard to mine”, there’s no API and access is limited (if you don’t want to deal with CAPTCHAs and “We’re sorry, but your computer network may be sending automated queries” you will have to apply some IP address magic).
Google Scholar works pretty well, but there might be some incorrectly assigned citations – since their search algorithm is proprietary, we can’t really know why things go wrong if they do.
It is also unclear what exactly is monitored by Google Scholar, so there are definitely missing citations.
When counting and ranking citations we don’t check if they are mainly self-citations.
Given the above points, please take the following with a grain of salt.

Out of 435 accepted abstracts at DH2016 (plenary lectures, panels, long and short papers, posters, pre-conference workshops and tutorials), 158 are cited at least once according to Google Scholar, as of 15 February 2020 (= 3,5 years after the conference).

Here are our resulting data sets in case you wanna explore them yourself (CSV format):

DH2016 citations according to Google Scholar (158 abstracts)
DSH Special DH2016 Issue citations according to Google Scholar (15 papers)
DH2016 follow-up papers published elsewhere (selection) (3 papers)

What follows are some rankings based on these data sets (plus some general thoughts at the bottom of this blog post):

1. Most cited abstracts from DH2016 (as of 15 February 2020)

There are 28 abstracts with at least 4 citations collected by Google Scholar:

Eetu Mäkelä, Thea Lindquist, Eero Hyvönen:
CORE - A Contextual Reader based on Linked Data – Long Paper – 15 citations
(DBLP) (Google Scholar)
Esko Ikkala, Jouni Tuominen, Eero Hyvönen:
Contextualizing Historical Places in a Gazetteer by Using Historical Maps and Linked Data – Short Paper – 11 citations
(DBLP) (Google Scholar)
Terhi Nurmikko-Fuller, Jacob Jett, Timothy W. Cole, Chris Maden, Kevin R. Page, J. Stephen Downie:
A Comparative Analysis of Bibliographic Ontologies: Implications for Digital Humanities – Short Paper – 10 citations
(DBLP) (Google Scholar)
Peer Trilcke, Frank Fischer, Mathias Göbel, Dario Kampkaspar:
Theatre Plays as ‘Small Worlds’? Network Data on the History and Typology of German Drama, 1730-1930 – Long Paper – 8 citations
(DBLP) (Google Scholar)
Aris Xanthos, Isaac Pante, Yannick Rochat, Martin Grandjean:
Visualising the Dynamics of Character Networks – Long Paper – 8 citations
(DBLP) (Google Scholar)
Manuel Burghardt, Lukas Lamm, David Lechler, Matthias Schneider, Tobias Semmelmann:
Tool-based Identification of Melodic Patterns in MusicXML Documents – Short Paper – 8 citations
(DBLP) (Google Scholar)
Fotis Jannidis, Isabella Reger, Markus Krug, Lukas Weimer, Luisa Macharowsky, Frank Puppe:
Comparison of Methods for the Identification of Main Characters in German Novels – Short Paper – 8 citations
(DBLP) (Google Scholar)
Roman Klinger, Surayya Samat Suliya, Nils Reiter:
Automatic Emotion Detection for Quantitative Literary Studies – Poster – 8 citations
(DBLP) (Google Scholar)
Maciej Maryl, Maciej Piasecki, Ksenia Mlynarczyk:
Where Close and Distant Readings Meet: Text Clustering Methods in Literary Analysis of Weblog Genres – Long Paper – 7 citations
(DBLP) (Google Scholar)
Manuel Burghardt, Michael Kao, Christian Wolff:
Beyond Shot Lengths - Using Language Data and Color Information as Additional Parameters for Quantitative Movie Analysis – Poster – 7 citations
(DBLP) (Google Scholar)
Christof Schöch, Daniel Schlör, Stefanie Popp, Annelen Brunner, Ulrike Henny, José Calvo Tello:
Straight Talk! Automatic Recognition of Direct Speech in Nineteenth-Century French Novels – Long Paper – 6 citations
(DBLP) (Google Scholar)
Johannes Hellrich, Udo Hahn:
Measuring the Dynamics of Lexico-Semantic Change Since the German Romantic Period – Short Paper – 6 citations
(DBLP) (Google Scholar)
Fahad Khan, Francesca Frontini, Federico Boschetti, Monica Monachini:
Converting the Liddell Scott Greek-English Lexicon into Linked Open Data using lemon – Short Paper – 6 citations
(DBLP) (Google Scholar)
Federico Nanni, Pablo Ruiz Fabo:
Entities as topic labels: improving topic interpretability and evaluability combining Entity Linking and Labeled LDA – Short Paper – 6 citations
(DBLP) (Google Scholar)
Susan Brown, Tanya E. Clement, Laura Mandell, Deb Verhoeven, Jacque Wernimont:
Creating Feminist Infrastructure in the Digital Humanities – Panel – 5 citations
(DBLP) (Google Scholar)
Alexander Czmiel:
Sustainable publishing - Standardization possibilities for Digital Scholarly Edition technology – Long Paper – 5 citations
(DBLP) (Google Scholar)
Isabella di Lenardo, Benoit Seguin, Frédéric Kaplan:
Visual Patterns Discovery in Large Databases of Paintings – Long Paper – 5 citations
(DBLP) (Google Scholar)
Francesca Frontini, Carmen Brando, Jean-Gabriel Ganascia:
REDEN ONLINE: Disambiguation, Linking and Visualisation of References in TEI Digital Editions – Long Paper – 5 citations
(DBLP) (Google Scholar)
Kim Jautze, Andreas van Cranenburgh, Corina Koolen:
Topic Modeling Literary Quality – Long Paper – 5 citations
(DBLP) (Google Scholar)
Mikko Tolonen, Niko Ilomäki, Hege Roivainen, Leo Lahti:
Printing in a Periphery: a Quantitative Study of Finnish Knowledge Production, 1640-1828 – Long Paper – 5 citations
(DBLP) (Google Scholar)
Peter Robinson, Barbara Bordalejo:
Textual Communities – Poster – 5 citations
(DBLP) (Google Scholar)
Alexander Dunst, Rita Hartel, Sven Hohenstein, Jochen Laubrock:
Corpus Analyses of Multimodal Narrative: The Example of Graphic Novels – Long Paper – 4 citations
(DBLP) (Google Scholar)
Maciej Eder, Jan Rybicki:
Go Set A Watchman while we Kill the Mockingbird in Cold Blood, with Cats and Other People – Long Paper – 4 citations
(DBLP) (Google Scholar)
Chao-Lin Liu:
Quantitative Analyses of Chinese Poetry of Tang and Song Dynasties: Using Changing Colors and Innovative Terms as Examples – Long Paper – 4 citations
(DBLP) (Google Scholar)
Alexandre Rigal, Dario Rodighiero, Loup Cellard:
The Trajectories Tool: Amplifying Network Visualization Complexity – Long Paper – 4 citations
(DBLP) (Google Scholar)
Valérie Beaudouin, Zeynep Pehlivan:
The Great War on the Web: the Making of Citing and Referencing by Amateurs – Short Paper – 4 citations
(DBLP) (Google Scholar)
Angelo Mario Del Grosso, Davide Albanesi, Emiliano Giovannetti, Simone Marchi:
Defining the Core Entities of an Environment for Textual Processing in Literary Computing – Poster – 4 citations
(DBLP) (Google Scholar)
Martin Reckziegel, Stefan Jänicke, Gerik Scheuermann:
CTRaCE: Canonical Text Reader and Citation Exporter – Poster – 4 citations
(DBLP) (Google Scholar)

2. Citation counts for follow-up articles published in the DSH special conference issue

EADH’s journal “Digital Scholarship in the Humanities” (DSH) published a special edition with 15 selected articles from the DH2016 conference (Volume 32, Issue suppl_2, December 2017). Here are their citation counts (titles sometimes differ from the original abstracts):

Stefan Evert, Thomas Proisl, Fotis Jannidis, Isabella Reger, Steffen Pielström, Christof Schöch, Thorsten Vitt:
Understanding and explaining Delta measures for authorship attribution – 22 citations (original abstract: 2)
(Google Scholar)
Heather Richards-Rissetto:
An iterative 3D GIS analysis of the role of visibility in ancient Maya landscapes: A case study from Copan, Honduras – 10 (2)
(Google Scholar)
Stefan Jänicke, David Joseph Wrisley:
Visualizing Mouvance: Toward a visual analysis of variant medieval text traditions – 9 (3)
(Google Scholar)
Grace Muzny, Mark Algee-Hewitt, Dan Jurafsky:
Dialogism in the novel: A computational model of the dialogic nature of narration and quotations – 8 (2)
(Google Scholar)
Elena González-Blanco, Clara Martínez Cantón, Gimena del Rio Riande, Salvador Ros, Rafael Pastor, Antonio Robles-Gómez, Agustín Caminero, María Luisa Díez Platas, Álvaro del Olmo, Miguel Urízar:
EVI-LINHD, a virtual research environment for the Spanish-speaking community – 5 (2)
(Google Scholar)
Arianna Ciula:
Digital palaeography: What is digital about it? – 4 (0)
(Google Scholar)
Katarzyna Bazarnik, Jakub Wróblewski:
First We Feel Then We Fall: James Joyce’s Finnegans Wake as an interactive video application – 4 (0)
(Google Scholar)
David L Hoover:
The microanalysis of style variation – 3 (0)
(Google Scholar)
James O Gawley, A Caitlin Diddams:
Comparing the intertextuality of multiple authors using Tesserae: A new technique for normalization – 3 (0)
(Google Scholar)
Marine Riguet, Suzanne Mpouli:
At the crossroads between the scientific and the literary discourse: Comparison as a figure of dialogism – 3 (3)
(Google Scholar)
Martijn Kleppe, Marco Otte:
Analysing and understanding news consumption patterns by tracking online user behaviour with a multimodal research design – 3 (0)
(Google Scholar)
Claire Warwick:
Beauty is truth: Multi-sensory input and the challenge of designing aesthetically pleasing digital resources – 1 (0)
(Google Scholar)
Taylor Arnold, Peter Leonard, Lauren Tilton:
Knowledge creation through recommender systems – 1 (0)
(Google Scholar)
Rui Hu, Carlos Pallán Gayol, Jean-Marc Odobez, Daniel Gatica-Perez:
Analyzing and visualizing ancient Maya hieroglyphics using shape: From computer vision to Digital Humanities – 1 (2)
(Google Scholar)
Joris J van Zundert, Tara L Andrews:
Qu’est-ce qu’un texte numérique?—A new rationale for the digital representation of text – 0 (0)
(Google Scholar)

3. Articles published elsewhere

As it is not always possible to say whether a full paper is really the elaborated version of a conference abstract, we can’t provide an exhaustive list. We only list here the three papers that go by the same titles as the original abstracts:

Melissa Terras, James Baker, James Hetherington, David Beavan, Martin Zaltz Austwick, Anne Welsh, Helen O’Neill, Will Finley, Oliver Duke-Williams, Adam Farquhar:
Enabling complex analysis of large-scale digital collections: humanities research, high-performance computing, and transforming access to British Library digital collections (DSH 33.2, June 2018) – 12 citations (original abstract: 0)
(Google Scholar)
Yuta Hashimoto, Yoichi Iikura, Yukio Hisada, SungKook Kang, Tomoyo Arisawa, Daniel Kobayashi-Better:
The Kuzushiji Project: Developing a Mobile Learning Application for Reading Early Modern Japanese Books (DHQ 11.1, 2017) – 8 (0)
(Google Scholar)
Pelle Snickars, Roger Mähler:
SpotiBot-Turing testing Spotify (DHQ 12.1, 2018) – 7 (0)
(Google Scholar)

Some Thoughts

ADHO conference abstracts do have a measurable scientific impact in the community.
Publishing a reworked conference abstract in a journal can (but doesn’t have to) increase your impact.
It would be nice to have a more stable, sustainable, citable format for ADHO conference proceedings (think DOIs maybe?).
Case in point: proceedings of DH2015 in Sydney seem to have vanished into the internet ether, there are only a bunch of unrendered XML files left somewhere, not really citable.
DOIs for books of abstracts are already a step in the right direction (see, e.g., DOI:10.5281/zenodo.2596095 for DHd2019).
Working with Google Scholar is still tedious. There are some Python libraries that are supposed to make it easier to work with it, but they don’t cover the full range of Google Scholar functions, and the proprietary mechanism can change at any time without warning and render these libraries unusable.

Anyway, so much for this little experiment. It’s a beautiful day in Moscow, time to visit the ice rink in Gorky Park! ❄️

DH Tools Mentioned in "The Programming Historian"

2020-01-17T00:00:00+00:00

This is a little follow-up to our blog post “Which DH Tools Are Actually Used in Research?” – The idea is to get a general overview of tool mentionings in the fabulous lectures of “The Programming Historian” (PH). We again used our ToolXtractor to extract tool names from PH’s 87 (English-language) lectures to date (including 7 retired ones), directly from their Markdown source files. Like last time, we used TAPoR’s database as positive list (for ToolXtractor relies on simple string matching). Any tool we failed to extract was not in TAPoR (yet), e.g., “Unity”, “Notepad++”, or “Atom” – an obvious shortcoming of this approach.

We also created a co-occurence graph (tools mentioned in the same lesson):

You can spot Python, R, Excel, GitHub and Zotero at the core of teaching in the Digital Humanities. Publishing tools are grouped on the left, network tools on the upper right side. GIS tools are on the bottom right-hand side, text analysis tools are located right above the centre, etc. (The graph was generated in a heartbeat with our tool ezlinavis and embellished with Gephi.)

No further analysis, just this overview:

jupyter-notebooks (11)

GitHub
Google Drive
JavaScript
Omeka
OpenRefine
Python
R
Ruby
Twitter
Voyant Tools
Zenodo

mapping-with-python-leaflet (9)

ArcGIS
C
GeoNames
GitHub
Google Maps
Leaflet
Nominatim
OpenStreetMap
Python

using-javascript-to-create-maps (9)

Excel
Gephi
Google Maps
GPS Visualizer
JavaScript
Leaflet
MapBox
Palladio
Zotero

corpus-analysis-with-antconc (8)

AntConc
Beautiful Soup
KWIC
Natural Language Toolkit · NLTK
OpenRefine
Python
R
Voyant Tools

geoparsing-text-with-edinburgh (7)

Excel
Firefox
GapVis
GeoNames
Google Maps
QGIS
Twitter

text-mining-with-extracted-features (7)

HT-Bookworm
Excel
GitHub
HathiTrust
Matlab
Python
R

beginners-guide-to-twitter-data (6)

Cytoscape
Excel
Gephi
Palladio
Tableau
Twitter

creating-apis-with-python-and-flask (6)

BBEdit
Firefox
JavaScript
MediaWiki
Python
Twitter

introduction-to-ffmpeg (6)

Audacity
Chrome
Excel
plot.ly
RAW
Twitter

sonification (6)

Excel
GitHub
Pandoc
Python
R
Ruby

visualizing-with-bokeh (6)

Bokeh
CartoDB
Excel
Leaflet
Python
QGIS

exploring-and-analyzing-network-data-with-python (5)

Gephi
NetworkX
Palladio
Python
R

geocoding-qgis (5)

ArcGIS
Excel
Google Maps
OpenStreetMap
QGIS

getting-started-with-github-desktop (retired) (5)

Dropbox
Google Drive
Jekyll
Pandoc
Python

getting-started-with-mysql-using-r (5)

MySQL
Python
R
RStudio
SQL Server

googlemaps-googleearth (5)

ArcGIS
Excel
Google Maps
QGIS
Twitter

introduction-to-populating-a-website-with-api-data (5)

Chrome
Firefox
GeoNames
JavaScript
Skype

introduction-to-stylometry-with-python (5)

Natural Language Toolkit · NLTK
Python
R
stylo
Zotero

json-and-jq (5)

Excel
JavaScript
Python
R
Twitter

qgis-layers (5)

ArcGIS
GDAL
OpenLayers
Python
QGIS

sustainable-authorship-in-plain-text-using-pandoc-and-markdown (5)

Google Docs
Jekyll
Pandoc
WordPress
Zotero

temporal-network-analysis-with-r (5)

Gephi
NetworkX
Python
R
RStudio

analyzing-documents-with-tfidf (4)

Bokeh
Excel
Overview
Python

extracting-illustrated-pages (4)

HathiTrust
Python
R
Tesseract

extracting-keywords (4)

Excel
OpenRefine
Python
QGIS

fetch-and-parse-data-with-openrefine (4)

Excel
Natural Language Toolkit · NLTK
OpenRefine
Python

graph-databases-and-SPARQL (retired) (4)

Excel
GeoNames
Palladio
plot.ly

intro-to-twitterbots (4)

GitHub
Python
Slack
Twitter

keywords-in-context-using-n-grams (4)

Firefox
KWIC
Python
Zotero

sentiment-analysis (4)

GitHub
Natural Language Toolkit · NLTK
Python
Twitter

transforming-xml-with-xsl (4)

Chrome
Excel
Firefox
GitHub

counting-frequencies-from-zotero-items (retired) (3)

Google Books
Python
Zotero

creating-and-viewing-html-files-with-python (3)

Firefox
Python
Zotero

data-mining-the-internet-archive (3)

Google Maps
Python
Wordle

data_wrangling_and_management_in_R (3)

Python
R
RStudio

generating-an-ordered-data-set-from-an-OCR-text-file (3)

JavaScript
Perl
Python

georeferencing-qgis (3)

GDAL
Google Maps
QGIS

geospatial-data-analysis (3)

Google Maps
plot.ly
R

gravity-model (3)

Excel
R
RStudio

installing-omeka (3)

MySQL
Omeka
WordPress

intro-to-powershell (3)

Mallet
Pandoc
Python

introduction-and-installation (3)

Beautiful Soup
Dropbox
Python

output-data-as-html-file (3)

Firefox
Python
Zotero

output-keywords-in-context-in-html-file (3)

Firefox
KWIC
Python

topic-modeling-and-mallet (3)

Excel
Mallet
Voyant Tools

automated-downloading-with-wget (2)

Python
Ruby

basic-text-processing-in-r (2)

R
RStudio

cleaning-data-with-openrefine (2)

Excel
OpenRefine

cleaning-ocrd-text-with-regular-expressions (2)

Excel
Python

correspondence-analysis-in-R (2)

R
Zenodo

creating-new-items-in-zotero (retired) (2)

Python
Zotero

dealing-with-big-data-and-network-analysis-using-neo4j (2)

Excel
Python

editing-audio-with-audacity (2)

Audacity
Soundflower

getting-started-with-markdown (2)

Pandoc
Perl

intro-to-beautiful-soup (2)

Beautiful Soup
Python

intro-to-the-zotero-api (retired) (2)

Python
Zotero

mac-installation (2)

Beautiful Soup
Python

naive-bayesian (2)

Beautiful Soup
Python

OCR-with-Tesseract-and-ScanTailor (retired) (2)

Tesseract
Zotero

preserving-your-research-data (2)

Excel
WordPress

r-basics-with-tabular-data (2)

Excel
R

transliterating (2)

Beautiful Soup
Python

understanding-regular-expressions (2)

Python
Ruby

vector-layers-qgis (2)

Google Maps
QGIS

working-with-web-pages (2)

Firefox
Python

applied-archival-downloading-with-wget (1)

Python

code-reuse-and-modularity (1)

Python

counting-frequencies (1)

Python

creating-an-omeka-exhibit (1)

Omeka

creating-mobile-augmented-reality-experiences-in-unity (1)

downloading-multiple-records-using-query-strings (1)

Python

from-html-to-list-of-words-1 (1)

Python

from-html-to-list-of-words-2 (1)

Python

installing-python-modules-pip (1)

Python

intro-to-bash (1)

Pandoc

intro-to-linked-data (1)

GeoNames

linux-installation (1)

Python

manipulating-strings-in-python (1)

Python

normalizing-data (1)

Python

research-data-with-unix (1)

Excel

up-and-running-with-omeka (1)

Omeka

viewing-html-files (1)

Firefox

windows-installation (1)

Python

working-with-text-files (1)

Python

intro-to-augmented-reality-with-unity (retired) (0)

(none)

Which DH Tools Are Actually Used in Research?

2019-12-06T00:00:00+00:00

Not that we didn’t know that, but Gephi is the most popular DH tool actually used in research work. Followed by Omeka, stylo, MALLET, Excel, D3.js, the NLTK, WordPress, Drupal, TextGrid, CollateX, GeoNames, TXM, Solr and Voyant Tools.

We know this because we did some counting. So let’s explain:

The longest-standing tool directory in the Digital Humanities, the Canadian portal TAPoR led by Geoffrey Rockwell, has around 1.500 DH tools in its database (including historic ones). We were wondering how this richness of means and utilities in our field is manifested in actual research work. To gain some first insights, we decided to extract the names of tools from TAPoR (which is easy thanks to their API) and match them with the proceedings of the largest and broadest event series in the Digital Humanities, ADHO’s annual DH conferences. The proceedings from DH2015 to DH2019 are freely available (licensed under CC BY 4.0, thx to Fabio Ciotti for giving us early access to the DH2019 data), so that was our chosen source:

DH2015, Sydney
DH2016, Kraków
DH2017, Montréal
DH2018, Ciudad de México
DH2019, Utrecht

Altogether, 238 tools were mentioned at least once in these five years, and we counted 1.498 mentions of tools altogether in all the proceedings. To extract this kind of data, we wrote a little Java tool called ToolXtractor, which you can run yourself with alternative source material if you like and also with different lists of tools. Since this is simple string matching, we had to throw out some false positives manually (not too many, though).

We set up another website providing a complete list of all mentioned tools including a link to corresponding conference abstracts for verification: “Tools mentioned in the proceedings of the annual ADHO conferences (2015–2019)”

Please also check the ranking of all tools at the very bottom of this blog post for further insights.

40 Most Used Tools

Here are the 40 tools with most mentions (an interactive version is here: https://datawrapper.dwcdn.net/L96xg/):

(Ed. We know that this line chart visually suggests a connection between tools that isn’t there. We still consider this type of display to be beneficial – especially in the interactive version – since you can compare occurrences per tool per year. By hovering over the lines you will see that, e.g., 2016 was the year of stylo and TextGrid, 2017 the year of Leaflet, 2018 the year of Omeka and 2019 the year of GeoNames and Transkribus. – Big thx to Jan Horstmann for bringing this up.)

Known Issues

Before we continue with more visualisations, some caveats:

Of course, this is not a complete list of all tools mentioned in the proceedings, only of those tools comprised in the TAPoR database (the vast majority should be covered, though).
We decided to leave programming languages like Python or JavaScript in, just like social networks (Twitter) and other things that are not DH tools in a narrower sense. Given that we published the data, it is easy to take them out to get a clearer view at actual DH tools.
We had to use a stopword list containing tools with names that are also frequent terms, like “Processing”, and we also left out one-letter programming languages, like R and C. Especially R is quite popular in our community and it would be great to compare it to its biggest contender: Python. But this would be future work. (Stopword list etc. to be found in the ToolXtractor repo.)

Btw, everyone is welcome to do their own viz, the dataset is freely available (in CSV format).

Streamgraph

This graph shows all tools and their mentions per year (but rather than staring at the graph below you should check out the interactive version of it featuring dynamic captions: https://rpubs.com/Pozdniakov/stream_dh):

Line Chart and Bar Plot (Top-10 Tools)

Bar Plot and Stacked Area (All Tools)

Ranking

All 238 tools mentioned with number of occurrences 2015–2019:

Python (125)
Twitter (82)
Gephi (60)
JavaScript (59)
Omeka (44)
GitHub (40)
HathiTrust (37)
stylo (35)
MALLET (33)
Google Books (31)
Excel (30)
MySQL (27)
D3.js (23)
Natural Language Toolkit · NLTK (23)
WordPress (20)
Drupal (19)
TextGrid (19)
CollateX (18)
GeoNames (18)
TXM (18)
Solr (17)
Voyant Tools (17)
EEBO-TCP (15)
Palladio (15)
PostgreSQL (15)
Bootstrap (14)
Leaflet (14)
OpenRefine (14)
Zotero (14)
eXist-db (13)
Google Maps (13)
Tesseract (13)
ArcGIS (12)
Music Encoding Initiative (12)
Transkribus (12)
Carto · CartoDB (11)
Juxta (10)
KWIC (10)
Perl (10)
Neatline (9)
GAMS (8)
Hypothes.is (8)
Jekyll (8)
MediaWiki (8)
Recogito (8)
Ruby · Ruby on Rails (8)
Tableau · Tableau Public (8)
AntConc (7)
Chrome (7)
Firefox (7)
Lucene (7)
OpenStreetMap (7)
Pundit (7)
Tesserae (7)
TRACER (7)
Zenodo (7)
Annotation Studio (6)
Cytoscape (6)
GATE (6)
Lexos (6)
Neo4j (6)
Photogrammar (6)
TEI Boilerplate (6)
WebLicht (6)
WorldCat (6)
CATMA (5)
Google Scholar (5)
LIWC · Linguistic Inquiry and Word Count (5)
QGIS (5)
Blacklight (4)
Bookworm (4)
Dropbox (4)
Freebase (4)
Google Ngram Viewer (4)
JGAAP (4)
nodegoat (4)
Oral History Metadata Synchronizer (4)
Protégé (4)
Textual Communities (4)
Tumblr (4)
VARD · VARD 2 (4)
Weka (4)
Wordle (4)
WordSmith (4)
Blender (3)
Commons In A Box (3)
CQPweb (3)
ELAN (3)
Google Drive (3)
NetworkX (3)
NodeXL (3)
OxGarage (3)
PhiloLine (3)
RStudio (3)
Skype (3)
TILE (3)
TileMill (3)
Tropy (3)
Versioning Machine (3)
ABBYY FineReader (2)
ANNIS (2)
Archive-It (2)
arts-humanities.net (2)
ATLAS.ti (2)
Bokeh (2)
Dataverse (2)
DiRT Directory (2)
Diva.js (2)
DSpace (2)
EATS (2)
ediarum (2)
Fedora Commons (2)
FromThePage (2)
Gamera (2)
igraph (2)
ImageJ (2)
Inkscape (2)
Islandora (2)
Joomla (2)
Map Warper (2)
Matlab (2)
MAXQDA (2)
MorphAdorner (2)
oXygen (2)
Pandoc (2)
Paper Machines (2)
PhiloLogic (2)
Protovis (2)
RelFinder (2)
RoSE (2)
Slack (2)
Spotify (2)
SylvaDB (2)
T-PEN (2)
TAPoR (2)
TRAViz (2)
TypeWright (2)
UIMA (2)
VSim (2)
Abbot (1)
Academia.edu (1)
Adobe After Effects (1)
Adobe Flash (1)
Adobe Illustrator (1)
Adobe InDesign (1)
Advene (1)
Alpheios (1)
Alveo (1)
Annotorious (1)
Anthologize (1)
Anvil (1)
AskSam (1)
Audacity (1)
AustESE (1)
Basecamp (1)
Beautiful Soup (1)
BuddyPress (1)
Chart.js (1)
Chronos Timeline (1)
COBOL (1)
Collex (1)
ColorBrewer (1)
Confluence (1)
CONTENTdm (1)
Dedoose (1)
DH Press (1)
eLaborate (1)
EndNote (1)
Evernote (1)
EVI-LINHD (1)
Exhibit 3.0 (1)
ezlinavis (1)
FAIMS Mobile Platform (1)
Finale (1)
GapVis (1)
GeoTemCo (1)
gFacet (1)
Google Docs (1)
Graphviz (1)
H-Net (1)
HyperImage (1)
HyperPo (1)
Icon (1)
ImagePlot (1)
IRaMuTeQ (1)
IsaViz (1)
jsLDA (1)
Koha (1)
KORA (1)
Lexomics (1)
LilyPond (1)
LimeSurvey (1)
LodLive (1)
Mediathread (1)
Mendeley (1)
Mukurtu CMS (1)
MuseScore (1)
music21 (1)
NeOn Toolkit (1)
NetDraw (1)
NVivo (1)
Old Maps Online (1)
OmniPage (1)
OntoViz (1)
OpenLayers (1)
Pliny (1)
Prefuse (1)
Prism (1)
Pro Tools (1)
ProcessingJS (1)
Project Quincy (1)
Prolog (1)
PyDelta (1)
RDF Gravity (1)
SARIT (1)
Scripto (1)
SemLens (1)
Serendip (1)
Serendip-o-matic (1)
Sibelius (1)
StoryMapJS (1)
Textable (1)
Textal (1)
Textexture (1)
TextRazor (1)
tFacet (1)
Transana (1)
Trello (1)
TUSTEP (1)
VexFlow (1)
Visual Browser (1)
VisualEyes (1)
VIVO (1)
Wavesurfer (1)
Wiki Map Project (1)
WordCruncher (1)
WordHoard (1)
YAGO (1)

Using webweb and netwulf with DraCor's API to Generate Interactive Literary Networks

2019-09-10T00:00:00+00:00

DraCor is the Drama Corpora platform at dracor.org which holds a growing number of collections of plays of different languages, countries and times (German, Russian, Spanish, Swedish, Roman, Greek, with more to come). All plays are encoded in XML-TEI, but you don’t have to meddle with the XML since the DraCor API makes it really easy to access structured information, e.g., network data based on the co-occurrence of characters per scene (as an example, here’s the social network extracted from Shakespeare’s “Hamlet”).

You can do all kinds of things with this network data, e.g., feed them into network visualisation tools. So, a couple of weeks ago, Twitter pointed us to two interesting Python packages, webweb and netwulf:

webweb is “a tool for creating, displaying, and sharing interactive network visualizations on the web” (DOI:10.21105/joss.01458),
netwulf is “an interactive visualization tool for networkx Graph-objects, that allows you to produce beautifully looking network visualizations” (netwulf.readthedocs.io).

It’s really easy to get started and we decided to give these two a spin, thereby also demonstrating the versatility of our DraCor API which provides network data for all TEI-encoded plays of any corpus on the fly.

We prepared a Jupyter notebook with the code ready to be executed.

If you want to play with drama networks without doing the legwork, we uploaded some ready-to-use HTML pages generated by webweb:

Pushkin: Boris Godunov
Lessing: Emilia Galotti
Russian Drama Corpus (you can choose between 190 plays in the left upper corner)
German Drama Corpus (dito, between 474 plays)

But let’s start to toy around with the other library first:

netwulf

We created a universal function (netwulf_representation in our Jupyter Notebook) and a function to retrieve the titles of all plays of a corpus (retrieve_all_plays).

Now let’s take Pushkin’s historical tragedy “Boris Godunov” (c. 1825) as an example:

# 1st network: gender groups without labels (other attributes are default)
netwulf_representation('rus', 'pushkin-boris-godunov', group='gender', show_node_labels=False)

# 2ns network: gender groups with labels
netwulf_representation('rus', 'pushkin-boris-godunov', group='gender')

# 3rd network: isGroup groups with labels
netwulf_representation('rus', 'pushkin-boris-godunov', group='isGroup')

# 4th network: gender groups with numOfSpeechActs size without labels
netwulf_representation('rus', 'pushkin-boris-godunov', group='gender', size='numOfSpeechActs', show_node_labels=False)

By the way, in order to fully regnerate the network (the radii of nodes are not stored by default by netwulf) we need to add a ‘size’ field to the nodes:

for node in network['nodes']:
    node['size'] = node['radius']

Now, as second example, let’s turn to German drama and try the usual suspect, Lessing’s “Emilia Galotti” (1772):

# 1st network: gender groups without labels
netwulf_representation('ger', 'lessing-emilia-galotti', group='gender', show_node_labels=False)  # other attributes are default (see above)

# 2nd network: isGroup groups with labels and degree sizes
netwulf_representation('ger', 'lessing-emilia-galotti', group='isGroup', size='degree')

# 3rd network: gender groups with node and link labels and numOfWords size
netwulf_representation('ger', 'lessing-emilia-galotti', group='gender', size='numOfWords', show_link_labels=True)

# 4th network: all characters with numOfScenes size with node and link labels
netwulf_representation('ger', 'lessing-emilia-galotti', size='numOfSpeechActs')

It is also possible to apply our script to all plays of a corpus (e.g., the Russian one):

playnames = retrieve_all_plays('rus')
for playname in playnames:
    netwulf_representation('rus', playname, size='weightedDegree', group='gender', show_link_labels=True, is_test=True)

With this piece of code, netwulf will launch the networks one by one. It will display each network for 5 seconds, followed by a 3-second fade-out before opening the next one. So if you are really bored, you can let hundreds of Russian drama networks pass by your eye. Our collection of currently 190 plays will steal 25 minutes of your time.

webweb

Dan Larremore’s webweb library offers a similar approach to visualise network data (see their documentation). We wrote two scripts: the first one is for representing only one play, the second one for representing several plays, for example, a whole corpus, or translations or different editions of the same play.

We created a universal function to integrate DraCor data with webweb (webweb_representation in our our Jupyter Notebook). This script is a bit easier and uses less additional parameters. That is because the library provides us an opportunity to fine-tune the settings such as ‘color nodes by’/’scale node sizes by’ interactively in the generated HTML file. Definitely an advantage since you don’t have to restart the program.

Other advantages as follows:

interactive captions where you can interactively scale node sizes (although they are limited to 5 sizes),
ability to highlight a node (but it’s not possible to highlight groups of nodes, only if these nodes have a common letter sequence in their labels), yet in netwulf we can highlight any number of nodes (need to tap twice on each of them),
choose between a collection of networks (ability to compare them on one page), dynamic networks (ability to see change over time),
ability to color nodes,
SVG output.

In our opinion, webweb also has some disadvantages in comparison to the netwulf library:

inability to zoom in and out,
no possibility to freely change parameters (e.g., colour sets, node radius, etc.) as with netwulf library,
some parameters (like the link opacity) are binary while in netwulf they are continuous, some parameters are missing (e.g., stroke width, and no ‘wiggle’ option),
inability to restore a graph (not in SVG format) in the exact same way.

Let’s look at the graphs produced by webweb. First Pushkin’s “Boris Godunov”:

# first network: gender groups with labels (size == weightedDegree)

# second network: isGroup groups with labels (size == numOfSpeechActs)

And now Lessing’s “Emilia Galotti” again:

# third network: gender groups with labels (size == weightedDegree)

# fourth network: numOfScenes groups with labels (size == strength)

As stated above, the HTML files generated by webweb embed all the JavaScript you need and you can save them as standalone pages:

"Moscow Formalism and Literary History" (De Gruyter)

2019-06-01T00:00:00+00:00

Just Out: New Translations of Early Data-Driven Formalism (Yarkho, Gasparov, Shapir)

A new volume of the “Journal of Literary Theory” was published in March 2019 (vol. 13, no. 1: “Moscow Formalism and Literary History”), a special issue prepared and edited by us comprising three translations from the Russian plus preface:

Frank Fischer, Marina Akimova, Boris Orekhov: Preface: Data-Driven Formalism (DOI:10.1515/jlt-2019-0001)
Boris I. Yarkho: Speech Distribution in Five-Act Tragedies (A Question of Classicism and Romanticism) (1935–1938, first published 1997) (DOI:10.1515/jlt-2019-0002)
Mikhail L. Gasparov: The Evolution of Russian Rhyme (1984) (DOI:10.1515/jlt-2019-0003)
Maksim I. Shapir: “For Thee There Is No Weight nor Measure”. The Possibilities and Limitations of “Exact” Methods in the Humanities (2005) (DOI:10.1515/jlt-2019-0004)

Fig. 1. Book cover and table of contents.

The idea for this volume goes back to a chat that Asya Bonch-Osmolovskaya and I had with Fotis Jannidis at the DH2015 conference in Sydney. As one of its editors, Fotis suggested to equip a special edition of “Journal of Literary Theory” with untranslated works of Russian formalists, especially quantitative literary studies. The idea was to build a bridge from early formalism to digital literary studies of today. The outcomes of the Stanford conference on “Russian Formalism & the Digital Humanities” (2015) suggested that it is not so easy to reconcile the two. That was, I would argue, due to the fact that some decisive texts were still untranslated.

Boris Yarkho

So we assembled a small team: Marina Akimova and Boris Orekhov joined as co-editors and Craig Saunders as translator. The first few pages of Yarkho’s text arrived in November 2017 with Craig admitting to “feeling a little charmed by the way he [Yarkho] writes”.

Fig. 2. Boris Yarkho (1889–1942). (source: https://urokiistorii.ru/article/52560, cropped)

Except for a draft plan of Yarkho’s “Methodology for a Precise Science of Literature” (transl. by L. M. O’Toole, publ. 1977 in vol. 4 of “Russian Poetics in Translation”) and Gasparov’s early account of “Boris Yarkho’s works on literary theory” (transl. by Michael Lavery and Marina Tarlinskaja, publ. 2016 in “Studia Metrica et Poetica”), there’s not much first-hand info on Yarkho available in English. “Speech Distribution in Five-Act Tragedies” is the first translation of an article by Yarkho and it can be expected that it will leave its mark over time.

We know from a reliable source how the manuscript was first edited in the 1990s, because one of the co-editors of our volume was directly involved. Marina received the microfilm with copies of Yarkho’s manuscripts from Gasparov and manually transcribed them into a computer, projecting the film on a wall in her flat.

Yarkho’s description of his data and experimental setup are so accurate that it was easy to implement them. DraCor, our platform for research on drama (dracor.org), automatically extracts the speech-distribution pattern in the way Yarkho described it (monologues, dialogues, three-way dialogues and so forth) for every corpus and play connected to the platform.

Here are two examples:

Fig. 3. Speech distributions for Lessing’s "Emilia Galotti" (source: dracor.org) and
Ozerov’s "Dmitry Donskoy" (source: dracor.org)

Summer in Moscow

After we had finished work on Yarkho, we tackled the remaining two articles. The main difficulty in Gasparov was to put the stress on the many examples and rhyming patterns, which were not present in the Russian original. And finally, translating Shapir proved particularly challenging. It also looks like we were the first to actually translate a Shapir article into English.

Most of the work was done by the summer of 2018 when the World Cup was being held in Russia. I remember one of our last meetings one afternoon in June when Morocco had just been eliminated from the World Cup by a Ronaldo goal. I was sitting in the Metro on my way to a cafe close to Leninka and saw many melancholic Moroccan fans with their flags trying to cheer each other up.

After we received feedback from the editors, the printing proofs were largely completed in November.

Panel Discussion at ZfL Berlin: June, 12th, 2019

Thanks to the Leibniz Center for Literary and Cultural Research (ZfL) in Berlin, we will have the opportunity to present the volume and talk about the larger context in a panel discussion on June, 12th, 2019. Marina and I will be joined on stage by ZfL’s Siarhei Biareyshik, and Zaal Andronikashvili will chair the event.

More information:
http://www.zfl-berlin.org/veranstaltungen-detail/items/moscow-formalism-and-literary-history.html

Vossian Antonomasia in "The New York Times"

2019-02-19T00:00:00+00:00

New Paper Out!

We’ve been collecting and studying Vossian antonomasia (VA, or, vossanto) for about a decade now. This stylistic device, named after Dutch humanist Gerardus Vossius, occurs when someone is called “the Jay Leno of Tel Aviv” or “the Marilyn Monroe of the Arab world” etc. etc.

We just published a new paper about this phenomenon …

Frank Fischer; Robert Jäschke: ‘The Michael Jordan of greatness’—Extracting Vossian antonomasia from two decades of The New York Times, 1987–2007. Digital Scholarship in the Humanities. 2019. (DOI:10.1093/llc/fqy087) (Preprint: arXiv:1902.06428)

… plus a website with our code and some more interesting results for you to explore:

https://vossanto.weltliteratur.net/

But let’s rewind a bit. Some years ago we published a hand-picked list with our favourite VA expressions from 2009 to 2014 (mainly from German-language media) and published what we thought would be our final article on the subject in Frankfurter Allgemeine Sonntagszeitung in December 2014.

There is a lot of research out there on all types of antonomasia, even on Vossian antonomasia (or ‘metaphorical antonomasia’, or ‘paragons’). But when we started looking, we didn’t come across any convincing automatic approach regarding the bulk extraction of VA to learn more about general patterns. All papers we found used hand-picked examples and took it from there. Of course, there are the type of websites like “The Michael Jordan of…”, which collected 1,505 Michael Jordans of something. As interesting as that is, from an extraction point of view it is trivial.

Vossian Antonomasia, Next Level

So before long we were back in the game and started hacking. We wanted to find a way to extract all (well, as many as possible 😊) occurrences of Vossian antonomasia from larger newspaper corpora, not only based on usual suspects like Michael Jordan (for what it’s worth, our research has confirmed on much more reliable grounds that he’s the ruler of the Vossian world). We wanted to find all the unexpected examples, so we fetched some newspaper corpora and our first idea was to do some POS tagging and NER and then try to match the pattern “the [name] of”. We worked with two full-text corpora, the New York Times 1987–2007 and German weekly Die Zeit 1995–2011. We presented this work at DHd2017 in Bern, but were not really content with the results. Especially NER let us down big time. But it was a start.

A couple of months later we had the idea that took our research to a new level. Thanks to a research scholarship granted to me by the Information School at the University of Sheffield, Robert and I were able to spend two weeks hacking together in May, 2017. And then, probably during one of our lunches at the Red Deer (Vegetarian Haggis ftw!), we came to talk about Wikidata and then it occurred to us that it could help us out of our NER concerns. Here are the two things we brought together:

1. We want to extract paragons, that is, well-known, infamous people. People that would definitely have a Wikipedia entry in some language, or else they wouldn’t be well-known enough to serve as source for Vossian antonomasia.
2. Wikidata inherits these famous people from all the different Wikipedia language versions and has all kinds of facts about them in a well-structured format, including nicknames and alternative spelling of names (which we would need).

Please note that this solution was really carved out for our specific research question, because Wikidata arguably holds exactly those named entities that we would need. Yet it probably wouldn’t be the best choice for other NER purposes.

The rest was pretty simple. It was now enough to use a regular expression to extract the general pattern “the … of”, and then we would match the part in-between against our newly gained persons database. And that’s what we did. The corpus at hand was the New York Times one, and whenever we initiated another extraction process (which always took some minutes), we would pay a visit to the Red Deer to have just another ginger beer.

Results

Vossian antonomasia is a rare phenomenon. Annotating a gold set of 500 randomly selected articles might not throw you even one result. So, while we cannot provide numbers for recall due to the lack of a gold standard, we can provide our precision scores, because we were able to personally assess the extraction results. Altogether we managed to extract 2,646 VA expressions from 20 years of the New York Times (1987–2007).

The paper features our most notable results, but we also provide some more things to explore on the project website:

Like mentioned above, Michael Jordan is the boss of it all, which can be witnessed in the table with the top-40 source entities for Vossian antonomasia we extracted. But Jordan is closely followed by Rodney Dangerfield, a US-American comedian infamous for the sentence “I don’t get no respect!”. I think that being the runner-up in Vossian Olympics actually might earn him some respect after all.

We also generated a gallery of the 40 top scorers by pulling their principal images from Wikidata (Property:P18):

(Btw, if you did the counting you’ll find only 39 images. One of the entities does not feature an image in Wikidata.)

Next Steps

Our research continues. We’re generally fine-graining our method because there are things we missed. And although “the … of” is arguably the most frequent pattern, we didn’t check other patterns yet, like “a … of”, or the type where the VA source is preceded by a national adjective, as in “the Vietnamese Mozart” or “the German Donald Rumsfeld”. We’re on it. And we’re open for collaborations with other scholars. Because one other thing we wanna do in the future is to broaden our research to corpora in other languages than English with richer morphology, which will add a layer of complexity.

Oh, if you find interesting Vossian antonomasia, feel free to ping me on Twitter (@umblaetterer) if you like.

"Russian Poetics in Translation" (10 Volumes, 1975–1983)

2018-08-24T00:00:00+00:00

Of course, the interest in Russian poetics is not new. What is new, however, is a growing interest within the Digital Humanities in transforming formalist and structuralist theory into algorithms and add a quantitative angle to it. An important milestone in this respect was the international symposium in Stanford in April 2015, “The Legacy of Russian Formalism and the Rise of the Digital Humanities”.

In Russia, the core texts are very present in everyday university life. Beyond Russian borders, though, it might look a bit different. Although there are a number of bibliographies of translations, it is difficult to keep track of what’s there and what’s not. Since we are currently preparing some new translations ourselves, we have tried to gain some kind of overview. In the course of doing so, we came across the excellent collection “Russian Poetics in Translation”. The 10 volumes, edited by L. M. O’Toole and Ann Shukman, were published between 1975 and 1983 (here’s an advertisement from 1982). Voilà, the front pages of the first five:

The collection is not available in a lot of libraries, it seems. If you’re lucky, you find one or two volumes in your nearest library, but probably not the whole collection. This is why a shout-out is due to Niederösterreichische Landesbibliothek (NÖLB) in St. Pölten which holds almost the entire set, donated by Joseph P. Strelka, a former professor at SUNY Albany. Thus, we didn’t have to go far to get hold of most of the volumes. The rest we gathered worldwide with the help of friends, and before long we were proud owners of (copies of) all volumes.

Now, in our experience, even online it is hard to throw a glance at the tables of contents per volume. Having said that, the volumes seem to be available as e-books via Amazon since 2013 (incl. the “look inside” option), yet we didn’t check these versions and are unsure about how they were converted and how readable/quotable they ultimately are (if you have any experience with them, feel free to tell us).

In any case, to convey the richness of this collection, we decided to create a bibliography of all articles in all 10 volumes and publish it on this web page. We also created a Zotero group in which we listed all articles. We hope this will inspire those of you who have never heard of this collection before.

* * *

Russian Poetics in Translation (10 Volumes, 1975–1983)

Table of contents of all volumes.

Vol. 1. Generating the Literary Text (1975, 77pp.)

Yu. K. Shcheglov; A. K. Zholkovskii: Towards A “Theme – (Expression Devices) – Text” Model of Literary Structure [1971]. Translated by L. M. O’Toole. (pp. 3–50)
Yu. K. Shcheglov: Towards a Description of Detective Story Structure [1968]. Translated by L. M. O’Toole. (pp. 51–77)

Vol. 2. Poetry and Prose (1976, 84pp.)

Yu. I. Levin: The Logic of Metaphor: Synthesis, Semantics, Transformations [1965/1969]. Translated by Christopher English. (pp. 5–21)
V. N. Toporov: The Semiotics of Prediction in Suetonius’s “The Lives of the Twelve Caesars” [1965]. Translated by Halina Willetts. (pp. 22–33)
Vyach. Vs. Ivanov: The Structure of Khlebnikov’s Poem “Menya pronosyat na slonovykh …” [1967]. Translated by Ann Shukman. (pp. 34–49)
Yu. K. Shcheglov: Some Features of the Structure of Ovid’s “Metamorphoses” [1962]. Translated by Julian Graffy. (pp. 50–65)
Yu. M. Lotman: The Structure of Ideas in Pushkin’s Poem “Andzhelo” [1973]. Translated by Ann Shukman. (pp. 66–84)

Vol. 3. General Semiotics (1976, 81pp.)

Yu. M. Lotman: The Modelling Significance of the Concepts ‘End’ and ‘Beginning’ in Artistic Texts [1966]. Translated by Wendy Rosslyn. (pp. 7–11)
G. Permyakov: The Logico-Semiotic Level of Proverbs and Sayings (Towards a Classification of the Genre) [1967]. Translated by Doris Bradbury. (pp. 12–32)
P. G. Bogatyrev: Stage Setting, Artistic Space and Time in the Folk Theatre [1968]. Translated by L. M. O’Toole. (pp. 33–37)
V. N. Toporov: On the Cosmological Origins of Early Historical Descriptions [1973]. Translated by Christopher English. (pp. 38–81)

Vol. 4. Formalist Theory (1977, 108pp.)

Brief Biographical and Bibliographical Notes on Leading Formalists. Translated by Ann Shukman. (pp. 1–12)
A Contextual Glossary of Formalist Terminology. Translated by Ann Shukman and L. M. O’Toole. (pp. 13–48)
Yu. Tynyanov; R. Jakobson: Problems of Research in Literature and Language [1928]. Translated by L. M. O’Toole. (pp. 49–51)
B. Yarkho: Methodology for a Precise Science of Literature (Draft Plan) [1936/1969] [Excerpts]. Translated by L. M. O’Toole. (pp. 52–70)
S. Bernshtein: Aesthetic Presuppositions for a Theory of Declamation [1927]. Translated by Ann Shukman. (pp. 71–89)
Osip Brik: The So-Called Formal Method [1923]. Translated by Ann Shukman. (pp. 90–91)
V. Shklovsky: In Defence of the Sociological Method [1927]. Translated by Ann Shukman. (pp. 92–99)
Ann Shukman: Russian Formalism: A Bibliography of Translations and Commentaries (Works in English, French, German and Italian). (pp. 100–108)

Vol. 5. Formalism: History, Comparison, Genre (1978, 93pp.)

Boris Eikhenbaum: Some Principles of Literary History: The Study of Lermontov [1924]. Translated by Christopher Pike. (pp. 1–8)
Yu. N. Tynyanov: Tyutchev and Heine [1929]. Translated by Christopher Pike. (pp. 9–19)
Yu. N. Tynyanov: Plot and Story-Line in the Cinema [1926]. Translated by Ann Shukman. (pp. 20–21)
Yu. N. Tynyanov: Preface to “The Problem of Verse Semantics” [1923]. Translated by Ann Shukman. (pp. 22–23)
V. V. Vinogradov: The Tasks Facing Stylistics [1923]. Translated by Ann Shukman. (pp. 24–29)
Olga Freidenberg: Three Plots or the Semantics of One: Shakespeare’s “The Taming of the Shrew” [1925/1936]. Translated by Ann Shukman and Halina Willetts. (pp. 30–51)
Boris Tomashevsky: Literary Genres [I. Dramatic Genres; II. Lyric Poetry; III. Narrative Genres] [1928]. Translated by L. M. O’Toole. (pp. 52–93)

Vol. 6. Dramatic Structure; Poetic and Cognitive Semantics (1979, 96pp.)

Yu. K. Scheglov: The Poetics of Molière’s Comedies [1977]. Part One: The Structural Invariants. Translated by L. M. O’Toole. Part Two: The Structure of One Comedy: M. de Pourceaugnac. Translated by N. F. C. Owen. (pp. 1–83)
Yu. M. Lotman: Culture as Collective Intellect and the Problems of Artificial Intelligence. Translated by Ann Shukman. (pp. 84–96) (PDF)

Vol. 7. Metre, Rhythm, Stanza, Rhyme (1980, 105pp.)

M. L. Gasparov: Quantitative Methods in Russian Metrics: Achievements and Prospects [1974]. Translated by G. S. Smith. (pp. 1–19)
K. F. Taranovsky: The Rhythmical Structure of Russian Binary Metres [1971]. Translated by G. S. Smith. (pp. 20–30)
M. L. Gasparov: Light and Heavy Verse Lines [1977]. Translated by G. S. Smith. (pp. 31–44)
M. G. Tarlinskaya: The Problem of English Versification: A Re-Examination [1976]. Translated by G. S. Smith. (pp. 45–60)
M. L. Gasparov: Towards an Analysis of Russian Inexact Rhyme [1975]. Translated by G. S. Smith. (pp. 61–75)
K. D. Vishnevsky: The Law of Rhythmical Correspondence in Heterogeneous Stanzas [1976]. Translated by G. S. Smith. (pp. 76–85)
Yu. M. Lotman: The Natural Language/Metre Inter-Relationship in the Mechanism of Verse [1974]. Translated by G. S. Smith. (pp. 86–89)
References and Bibliography. (pp. 90–105)

Vol. 8. Film Theory and General Semiotics (1981, 107pp.)

V. V. Ivanov: Functions and Categories of Film Language [1971/1975]. Translated by Stephen Rudy. (pp. 1–35)
Yu. M. Lotman: On the Language of Animated Cartoons [1979]. Translated by Ruth Sobel. (pp. 36–39)
A. K. Zholkovsky: Generative Poetics in the Writings of Eisenstein [1969]. Translated by L. M. O’Toole. (pp. 40–61)
A. K. Zholkovsky: Pushkin’s “Poetic World” [1976]. Translated by L. M. O’Toole. (pp. 62–107)

Vol. 9. The Poetics of Cinema (1982, 126pp.)

K. Shutko: Preface. Translated by Richard Taylor. (pp. 1–4)
B. Eikhenbaum: Problems of Cine-Stylistics. Translated by Richard Sherwood. (pp. 5–31)
Yury Tynyanov: The Fundamentals of Cinema. Translated by L. M. O’Toole. (pp. 32–54)
B. Kazansky: The Nature of Cinema. Translated by Joe Andrew. (pp. 55–86)
V. Shklovsky: Poetry and Prose in the Cinema. Translated by Richard Taylor. (pp. 87–89)
A. Piotrovsky: Towards a Theory of Film Genres. Translated by Richard Taylor. (pp. 90–106)
E. Mikhailov; A. Moskvin: The Cameraman’s Part in Making a Film. Translated by Ann Shukman. (pp. 107–118)

Vol. 10. Bakhtin School Papers (1983, 156pp.)

Ann Shukman: Introduction. (pp. 1–4)
V. N. Voloshinov [M. M. Bakhtin]: Discourse in Life and Discourse in Poetry. Questions of Sociological Poetics. Translated by John Richmond. (pp. 5–30)
V. N. Voloshinov [M. M. Bakhtin]: The Latest Trends in Linguistic Thought in the West. Translated by Noel Owen. (pp. 31–49)
P. N. Medvedev: The Formal (Morphological) Method or Scholarly Salieri-ism. Translated by Ann Shukman. (pp. 51–65)
P. N. Medvedev: Sociologism Without Sociology (On the Methodological Works of P. N. Sakulin). Translated by C. R. Pike. (pp. 67–74)
P. N. Medvedev [M. M. Bakhtin]: The Immediate Tasks Facing Literary-Historical Science. Translated by C. R. Pike. (pp. 75–91)
V. N. Voloshinov [M. M. Bakhtin]: Literary Stylistics 1: What is Language? Translated by Noel Owen. (pp. 93–113)
V. N. Voloshinov [M. M. Bakhtin]: Literary Stylistics 2: The Construction of the Utterance. Translated by Noel Owen. (pp. 114–138)
V. N. Voloshinov [M. M. Bakhtin]: Literary Stylistics 3: The Word and its Social Function. Translated by Joe Andrew. (pp. 139–152)
Glossary of Key Terms. (pp. 153–155)
Bibliography. (p. 156)

Underwater Basket Weaving: Russian Writers and Their Professions According to Wikidata

2018-08-20T00:00:00+00:00

1. Introduction

1.1. TL;DR

This is a little experiment which we conducted in a limited amount of time (one week). We wanted to find out how the range of jobs practised by Russian writers throughout history is represented in Wikidata. Since Wikidata is primarily filled in an automated fashion, we wanted to look for systematic errors in that process. However, we also dared to interpret the data on Russian authors at least a little, even if at this point it is neither complete nor specifically curated within Wikidata. As an appetiser, have a look at fig. 3:

Fig. 3. The 16 most popular professions among Russian writers between 1700 and 2000 (in 50-year spans).

No surprise there, writers generally tended to (also) be journalists, translators, literary critics. But now onto some details of our exercise…

1.2. “Underwater Basket Weaving”

Originally, according to Wikipedia, “underwater basket weaving” is an idiom “referring in a negative way to supposedly useless or absurd college or university courses”, or it can serve as an “intentionally humorous generic answer to questions about an academic degree”. As we were working on our project, we encountered a writer whose occupation was marked as “basket weaving”. In search for a good Russian translation we’ve encountered the idiom mentioned above. It seemed so funny and to the point that we named our project after it. During Soviet times, many famous writers were not officially recognised as such. Hence, they had to do some pseudo or meaningless work, something along the lines of “underwater basket weaving”.

1.3. Russian Writers’ Professions and Languages

The original goal of our project was to check the variety of jobs held by Russian writers. This is of interest since for the most part of Russian history it was hard to make a living off of just being a writer. Next to that question, and since we were experimenting with Wikidata, we also thought it would be interesting to see what languages except Russian the writers were able to speak. This could be interesting as an indicator of (foreign) cultural influences on Russian literary history.

1.4. Using Wikidata to Answer Scientific Questions

Using Wikidata as data source is fun, but problematic. In order to draw conclusions from analysing this data, you have to understand how data enters Wikidata and how it is maintained. Although human users can still alter single facts just like on any other project of the Wikimedia family, the interesting part of the Wikidata experience is that most of the processes are automated. So instead of correcting single facts, it is far more interesting to look for systematic errors.

While there are heavily curated subsets within Wikidata (like the one on Holocaust-era ghettos), our chosen subset of Russian writers is what it is at the moment and hasn’t seen any concerted action regarding its curation (at least not yet). However, we did dare to interpret the data on Russian authors at least a little, even if they are at this point neither complete nor specifically curated in Wikidata.

1.5. Ways to Automatically Improve Wikidata Using Bots

After collecting systematic errors, we also looked into ways to automatise the process using bots. Since we didn’t have a lot of time, we didn’t succeed to actually apply a bot, but we experimented a little and plan to explore this further, maybe in shape of another course work in the coming semester.

2. Description of Data Source

2.1. Wikidata

Wikidata is a free and open knowledge base that can be read and edited by both humans and machines. Wikidata acts as central storage point for structured data of all projects of the Wikimedia family (but is not limited to that). Each entity in Wikidata has a label (a name, for example), an identifier and several statements (such as “occupation”, “day of birth”, etc. for people).

2.2. Russian Wikipedia for Fact-Checking

When we spotted something suspicious in Wikidata, we cross-checked this information with the Russian Wikipedia. There were a lot of misattributions and systematic mistranslations, so we checked the original articles in Russian to find the source for this kind of mistakes.

2.3. Pywikibot Library

As Wikidata provides the instruments for automated editing, we set out to fix some Wikidata mistakes using a bot. There is a special Python library for that, Pywikibot, which can be used to create a bot. It can be launched in different ways among which we chose to host it on the Wikimedia labs environment using PAWS: A Web Shell.

3. Workflow

3.1. Querying Wikidata

We used Wikidata to identify Russian writers and their respective occupations, languages spoken, gender and year of death. To be able to look into temporal distribution, we chose the year of death as marker because we thought this was closer to their active time as writers than their birth date. The obvious limitation that comes with this approach is that writers still alive or ones lacking a year of death were not included into the analysis. But this way we were also closer to the canon of Russian literature, making us a bit more independent of Wikipedia’s “contemporary bias”.

3.2. Dataset Creation

Now, how did we find our set of Russian writers? We included people labeled as writers in Wikidata. They had to be either Russian by their nationality, have Russian as a native language or have to lived in the Russian Empire, the USSR or the Russian Federation. We excluded all people who did not feature Russian among their languages.

3.3. Systematic Problems in Wikidata

As we reviewed our raw data, we noticed a couple of errors and tried to classify them. We also made a table of all the mistakes we’ve encountered in order to then check whether they could be fixed by a bot (a link to our GitHub is below).

3.4. Attempt to Create a Wikibot

As Wikidata is an open resource with an enormous amount of data, using bots is actually dangerous. So, there are some mandatory stages you have to go through when creating a bot. We set up a regular Wikimedia account. To clarify that it is a bot you just have to mention the word “bot” at the end of username. To be approved to work on Wikidata your bot has to make several operations which will be assessed. If everything is good, the bot gets approved. When writing your bot script you can test it on the Wikidata:Sandbox.

Ultimately, we did not get to the stage of approval this time around because of struggling with several problems in the preparation process. Details are described below.

3.5. Manual Corrections

Due to the limited amount of time, we did not get to the point were we could correct systematic errors. Since our dataset is very limited, we manually corrected all mistakes spotted, and then repeated our queries.

3.6. JSON Unification in Python

Our queries initially had every writer occur as many times as he had professions plus spoken languages. To make the JSON more convenient, we wrote a Python script to transform raw data into a dictionary with unique names under which was stored a multi-layered dictionary of occupations, known languages, gender and year of death.

3.7. Visualisation in Python

The created dictionary was used to visualise information with by help of the following Python libraries: pandas, seaborn and matplotlib. At first, the data was reorganised into tables in order to create pandas dataframes. Thereafter, the seaborn graphs were compiled based on the aforementioned dataframes. Visual aspects were set up such as labels of the axes, captions, colour palettes, fonts, etc. Finally, all the charts were saved and displayed via matplotlib library.

4. Wikidata Errors

4.1. Mistranslations

Most of the problems we faced in our research were systematic, for example, mistranslations. It is quite obvious that Yanka Kupala was not an “imagemaker”. The error occurred due to automatic translation. In the Russian WP he was marked as “публицист” (transcribed “publicist” = “opinion journalist”), but the automatic translated the publicist back to “имиджмейкер” (“imagemaker”).

4.2. Misclassification Due to Mistranslation

In many cases there was a generic name instead of the actual job title (mostly due to mistranslation). For example, according to Wikidata, Bulgakov is speculative fiction and Gennady Malakhov is urine therapy. Such mistakes can be found easily, though. If as an occupation there is an entity which is not an ‘instance of profession’ but rather ‘instance of genre’ or something similar, it is an error. Yet in some cases there is no easy way to derive the original mistranslated job title as sometimes the mistranslation is too vague, such as ‘Orthodox Christianity’, making it harder to find out what the original job title was.

4.3. Redundancies

The last systematic problem we came across was the replication or synonymous labelling of entities such as “police officer” (Q361593 and Q384593) or ethnologist (Q1371378) versus ethnographer (Q12347522). Yet sometimes it is not quite clear whether articles should be merged. It is not obvious whether “official” (“должностное лицо”, Q599151), “official” (“чиновник”, Q382887) and “civil servant” (“государственный служащий”, Q212238) convey the same meaning.

4.4. Wikibot for Error Correction

Bots can be useful in solving some of the problems described above, especially in cases of misclassification. This involves the automated changing of statements, which is more complicated than just changing labels like for mistranslations.

However, the main problem we encountered was connected to bot testing within the sandbox. The structure of the sandbox is very different from the actual one, the two do not contain any identical entities, meaning there are no pages of Russian writers in the sandbox equal to the ones in the original Wikidata. In order to test our bot we had to create like 10 to 20 pages, imitating the actual Wikidata pages with all the properties we need, and only then try to work with our bot on this material. In the short amount of time that we had to prepare our project, we did not succeed to do so, but this is where we would have to continue next time around.

5. Results and Tentative Interpretations

We identified 8160 Russian writers in Wikidata meeting our eligibility criteria. Not all the fields were filled in for all of them, so those who lacked some pieces of information were excluded from the corresponding analysis. For example, we couldn’t put writers without year of death in our timeline. The amount of data on those who already have attributed a year of death can be seen in figure 1. As expected, the period best covered is the 20th century. The amount of data for the 19th century was significantly lower, and the first half of the 21st century appears to be under-represented, too, for obvious reasons.

Fig. 1. Amount of Russian writers in Wikidata.

5.1. Occupations

5.1.1. General

Figure 2 shows the distribution of different occupations of Russian writers. Unsurprisingly, the most popular professions are those where the ability to write well is a prerequisite: journalism, translation, literary criticism.

Fig. 2. Percentage of different occupations among Russian authors.

5.1.2. By Time

Figure 3 only shows the professions that had more than 5 people assigned to them in each period. We witness the decline of linguistics, philosophy, criticism, translation, diplomacy, history, education over the centuries, possibly also due to a diversifying landscape of professions, and the rise of journalism.

Fig. 3. The 16 most popular professions among Russian writers.

5.1.3. By Gender

Figure 4 shows that women writers had fewer professional options in past centuries, mostly limited to “work from home” options. They worked as interpreters, linguists, philologists, artists or were salonnières. An exception to this rule are the actresses among women writers.

Fig. 4. Popularity of professions among male and female writers.

5.1.4. Just a Writer (by Time & Gender)

Figure 5 shows the percentage of writers only marked as writers on Wikidata without additional occupations. The situation remains quite stable considering men, whereas the percentage of women falls from 100% to approximately 70% by the beginning of 19th century and then gradually declines to 45–50%. We tend to explain this by the fact that women gradually entered the labour market. Towards the 21st century, there is a 50% chance that being a writer means that a person also has an additional profession, according to Wikidata.

Fig. 5. Percentage of writers whose only profession in Wikidata is writer.

5.2. Languages Spoken

5.2.1. General

Figure 6 shows the distribution of additional languages Russian writers spoke or wrote in. There are two distinct types of languages: native regional languages spoken in territories of the former Russian Empire and the USSR (Ukrainian, Bashkir, Yiddish, etc.) and secondary European languages (French, German, English, etc.).

Fig. 6. Popularity of additional languages mastered by Russian writers.

5.2.2. By Gender

Figure 7 uncovers some interesting tendencies. Women writers in our data set mastered more European languages, such as English, French, German, and Italian, than male writers did. Women also knew better than men some smaller, less well-represented languages, like Armyan, Buryat, Hungarian, Saami.

Fig. 7. Languages spoken by Russian writers, divided by gender.

5.2.3. By Time

The distribution of languages spoken by Russian writers by time (figure 8) can reflect cultural influences of given periods. Between 1800–1850, the most popular language was German, which might be due to the German-orientated culture established by Peter the Great. During the second half of the 19th century (1850–1900), the most popular language was French, which corresponds to the big influence that French culture had on Russian culture at the time. At the beginning of the 20th century (1900–1950), the most popular languages were Ukrainian, French and Yiddish. During the second half of the 20th century (1950–2000), English joined the parade. In the most contemporary period, the most popular additional languages are French, Ukrainian and English. It is worth noting that the gap between popular and unpopular languages decreased over time, possibly an effect of globalisation, leading to a mix of contemporary cultures and languages exerting an influence on Russian literary life.

Fig. 8. Languages spoken by Russian writers, divided by time.

6. Appendix

All files, Python code, images, Wikidata search queries, etc. can be found on our GitHub.

Network Analysis of the 9th Season of "Curb Your Enthusiasm"

2018-01-01T00:00:00+00:00

2017 brought us the ninth season of “Curb Your Enthusiasm” (six whole years after season eight). I wanted to do a quick network analysis of the new season to check what we can learn from such a formalisation. To those of you who have watched it, this list of all ten episodes might serve as a quick recap, as “Curb” episode titles are usually good summaries, too:

9x01: Foisted!
9x02: The Pickle Gambit
9x03: A Disturbance in the Kitchen
9x04: Running with the Bulls
9x05: Thank You for Your Service
9x06: The Accidental Text on Purpose
9x07: Namaste
9x08: Never Wait for Seconds!
9x09: The Shucker
9x10: Fatwa!

Extraction Method and Visualisation

To link two characters to each other I used the same approach as in our research on social networks in dramatic texts: “Two characters interact with one another if they perform a speech act within the same segment of a drama (usually a ‘scene’).” (quoted from these slides; just replace ‘drama’ by ‘TV series’)

I segmented all 10 episodes by hand (414 segments across all ten episodes) and put them in the very simple format that our own tool ezlinavis parses to calculate network relations. And here’s the result after Gephi’ing it up a bit (licensed under CC BY 4.0):

No surprise there, “Curb” by all means is a Larry-centric show. There is hardly any scene without him, and each character of the show is only defined by his or her relation to Larry. We don’t know what Richard Lewis is doing or what Jeff and Susie are talking about when Larry’s not around. A notable exception in the above graph can be spotted in the left upper corner, Morsi’s visits to characters of older seasons of the show in episode 9x08.

While the Larry-centricity should be no news to watchers of the show, it is another thing to see this visualised for clarity.

Let’s look at some network-analytical metrics to get some more insights. The whole graph has 152 nodes and 427 edges. Most of the nodes are individual characters (with proper or generic names), some of them are groups of people (like the Male Hotel Guests trying to open a pickle jar in episode 9x02 or the Bus People in episode 9x07).

The network density is 0.037, a very low value, which has to do with the fact that the show is not interested in establishing relations between all its characters. Like mentioned above, all characters only come to life in their relation to Larry. The average path length is 2.197, not really a very meaningful value either for this network graph. It just means that, hypothetically, every character can reach any other character of the network via Larry, in no more than two steps, with a few exceptions.

More interesting are some node-related metrics, let’s have a look at the 10 characters with the highest degree:

Character	Degree
Larry	132
Jeff	36
Susie	29
Leon	27
Funkhouser	20
Lin	19
Sign Language Interpreter	19
Victor	17
Ted	16
Cheryl	15

So out of 151 possible connections to other characters in the network, Larry directly relates to 132 of them. A bit surprising is the high-ranking position of the Sign Language Interpreter who only appears in the last episode 9x10 (spot her in the bottom right corner of the network visualisation). Her appearance was a bit difficult to formalise. Her interpretation for the deaf is recognised by many of the main characters, which is why I established a link in each of these cases. Her degree value is high up also because she is present at two different events with two different audiences, the rehearsal for the musical and Sammi’s wedding.

But there’s a way to rank characters by their actual presence throughout the whole season. Let’s check their weighted degree (or, ‘strength’). The top ten looks a tad different here:

Character	Weighted Degree
Larry	573
Jeff	234
Susie	153
Leon	115
Funkhouser	73
Ted	69
Cheryl	68
Lin	67
Bridget	66
Richard	53

The Sign Language Interpreter as well as Sammi’s husband-to-be Victor fall out of this ranking, while Bridget (Larry’s girlfriend during some of the episodes) and his old pal Richard Lewis enter the picture. Ted and Cheryl, who are now an item, also rank higher, showing that they actually did occur on different occasions during the season, also if they didn’t play a central part.

Mrs. Templeton and Other Intricacies

Segmenting the episodes and identifying and placing all characters can be challenging. In general, I tried to use the cast lists on IMDb to find the right names for all characters, especially if they aren’t given actual names (but are listed under generic terms like “Hotel Day Manager” or “Receptionist”). If known characters appear under different names (Larry as “Buck Dancer” and Leon as “Chappie Johnson” in episode 9x02), I ignored that. In some few cases I had to introduce group names like “Male Hotel Guests”, because the characters didn’t act as distinguishable individuals but as part of a group.

And then there’s Mrs. Templeton, wife of Larry’s therapist Dr. Templeton in episode 9x04. We never see her talking at all, but I interpreted her wordless snubbing of Larry as speech act, which is why she appears in the graph. All in all, I tried to be as objective as possible. Feel free to have a look at the segmentation files.

Data and Software Used

Speaking of which, I opened a repo on GitHub for this small project, see here, where you can find the formalisation for each episode. After formalising, I threw the files into ezlinavis and saved the resulting network data in an CSV file, which I then opened in Gephi to calculate and visualise everything you’ve seen above.

As for “Curb”, word has it that we won’t have to wait another six years for the next season, production for season ten is said to start next spring.

A Giant 1890 Flowchart of Foreign Influences on German Literature (Spanning 12 Centuries)

2016-11-03T00:00:00+00:00

Exploring the influence that authors have (or might have) exerted on other authors is among the standard procedures pursued by scholars of (Comparative) Literary Studies. Provided the right type of data, these kinds of relations can be easily visualised in a network graph with arrows pointing in myriad directions: X influenced Y and, possibly, vice versa. What we have here, though, is a collection of unidirectional pointers which – in the form of streams of movements, events, works, and authors of world literature – exert their influence on one particular, ever-broader river: German literature, its manifold protagonists and epochs.

It is Cäsar Flaischlen’s “Graphische Litteratur-Tafel” (Graphic Literature Table) we’re talking about, a 58×86.5-cm poster published in 1890, and a flamboyantly beautiful one:

We came across this table just a couple of weeks ago, and quickly purchased a copy via ZVAB.com, scanned it, OCRed the 8-column preface, encoded both the text and the graphic table in TEI and put up a website to present the data, at litteratur-tafel.weltliteratur.net (still in alpha!). Simultaneously, we had another copy scanned by the Digitisation Centre of Göttingen’s University and State Library; you can download their hi-res version here (PDF format, 4.5 MB).

The Author

How did this chart come about? Cäsar Flaischlen (1864–1920) was not much of a practising literary scholar, nor an academic. The dissertation he wrote (on the Enlightenment playwright Otto von Gemmingen, with a nod to Diderot) was not well-received by his colleagues (one of which said it was “ziemlich nachlässig und schleuderhaft”, i.e., “quite sloppy and negligent”). Flaischlen’s dissertation was published in 1890, the same year in which his “Graphic Table of Literature” saw the light of day. Then he left academia to continue writing (dialect) poetry, novels and plays, while working as an editor for arts and literary magazines to pay the bills. (There’s an article on him in “Deutsches Literatur-Lexikon. Das 20. Jahrhundert”, vol. 9, 2006.)

Context I: Positivism

Flaischlen does not mention any of the sources from which he compiled his graphic table. The canon of authors and other entities he mentions is not in the least controversial and was part of the literary-historical consensus taught at universities at the time. Flaischlen’s impressive visualisation calls upon the power (and shortcomings, too) of positivist thinking (there are remarks on the outdatedness of Flaischlen’s approach in this 2012 introduction to Comparative Literature).

Positivism is largely seen in a negative light today, but one of the aspects behind it has been vividly discussed recently in the attempt to reinvent the Humanities along the lines of the natural sciences. “The natural sciences ride triumphantly in the chariot of victory to which we are all chained”, wrote Wilhelm Scherer, a central representative of positivism in the 1870s and 1880s. Trying to establish Literary Studies as an “exact science” was by no means unusual, as we can see in another quote by Shakespeare scholar Wilhelm Wetz who, in the 1890s, wanted to prove that Literary Studies, too, “can rise to the ranks of the exact sciences” (p. VII).

At the same time, Flaischlen’s flowchart stands in opposition to a popular positivist undercurrent of his time, biographism. Instead of looking painstakingly for possibly undiscovered details in an author’s life according to the biographistic paradigm, he dabbles in a kind of macroanalysis, an abstract model of the big picture, the organic connection.

Another aspect of Flaischlen’s chart is how he deals with the question of crafting a history of literature. The first half of the 19th century had a more philosophical approach. Literary history was understood as self-realisation of a mind, of an idea. This changed in the second half of the century. Erich Schmidt (who had studied under Scherer) described the new paradigm like this (in his inaugural address at the University of Vienna in 1880):

“Litteraturgeschichte soll ein Stück Entwicklungsgeschichte des geistigen Lebens eines Volkes mit vergleichenden Ausblicken auf die anderen Nationallitteraturen sein. Sie erkennt das Sein aus dem Werden und untersucht wie die neuere Naturwissenschaft Vererbung und Anpassung und wieder Vererbung und so fort in fester Kette.” (p. 491)

This approach seeks to understand a national literature in its interconnectedness to the literature of other countries. In the same article, Schmidt continues:

“Wie steht man zum Ausland? Der Begriff der Nationallitteratur duldet gleichwohl keinen engherzigen Schutzzoll; im geistigen Leben sind wir freihändlerisch. Aber ist Selbständigkeit oder Unselbständigkeit, größere Receptivität oder Productivität, wahre oder falsche Aneignung sichtbar, und wie hat die deutsche Litteratur sich allmählich zu universalistischer Antheilnahme emporgearbeitet?” (p. 493)

By and large, that is the point of Flaischlen’s flowchart.

At the time, one of the leading ideas about the history of literature was Scherer’s “Wellentheorie” (wave theory). This theory assumes that literature blooms and decays periodically, thus adopting the metaphorical form of a wave (or waves), peaking roughly every 600 years. Supposedly, the last heyday of German literature happened at around 1800. So according to the theory, contemporary literature (i.e., around 1890) would find itself in a period of decay, something that will not have found Flaischlen’s approval. He was an adherent to the naturalist movement of his time and a poet and novelist himself. His river model of German literature can thus be seen as a polemical remark about Scherer’s theory. Although the river narrows after 1200 (maintaining accordance with Scherer’s view), it becomes ever broader even after 1800, foreign influences stream into it, and the watery body of Germanophone literature absorbs, accumulates, amalgamates everything.

Context II: Visualising Time

Flaischlen was not the first to come up with the idea of representing a timeline as a river. This visualisation metaphor goes back to Austrian historiographer Friedrich Strass (1766–1845) who in 1804 published his highly influential “Strom der Zeiten” (“Stream of Time”). A hi-res scan of this chart was made by the Munich Digitisation Center (MDZ).

Rosenberg and Grafton discuss “Der Strom der Zeiten” in their fabulous compendium “Cartographies of Time” (2010, pp. 143–149). And they quote William Bell, the English translator of Strass’s impressive chart, who (in 1810) emphasises the qualities of the river metaphor:

“However natural it may be to assist the perceptive faculty, in its assumption of abstract time, by the idea of a line, (…) it is astonishing that (…) the image of a Stream should not have presented itself to any one (…). The expressions of gliding, and rolling on; or of the rapid current, applied to time, are equally familiar to us with those of long and short. Neither does it require any great discernment to trace (…) in the rise and fall of empire, an allusion to the source of a river, and to the increasing rapidity of its currents, in proportion with the declivity of their channels towards the engulfing ocean. Nay, this metaphor (…) gives greater liveliness to the ideas, and impresses events more forcibly upon the mind, than the stiff regularity of the straight line. Its diversified power, likewise, of separating the various currents into subordinate branches, or of uniting them into one vast ocean of power (…) tends to render the idea by its beauty more attractive, by its simplicity more perspicuous, and by its resemblance more consistent.” (pp. 143, 147; original quote here)

As a positivist approach, Flaischlen’s flowchart is non-controversial. The only thing that could appear controversial is the impact (breadth) of each of the influxes. He himself stresses in the last paragraph of his preface “that the breadth of the main river is not mathematically calculated”. His chart is not based on “data”; it is not yet part of those 19th-century movements that started to use numerical statistics to transform “knowledge” into “data knowledge” (“Datenwissen”, cf. Andreas Bernard: Das totale Archiv, Merkur, 2016). But his “Graphische Litteratur-Tafel” is an inspiring predecessor of our attempts today to use “graphs, maps, trees” to visualise and explore literary data.

Bibliography

Bernard, Andreas: Das totale Archiv. Zur Funktion des Nicht-Wissens in der digitalen Kultur. Merkur, vol. 801 (February, 2016), pp. 5–17. (merkur-zeitschrift.de)

Flaischlen, Cäsar: Graphische Litteratur-Tafel. Die deutsche Litteratur und der Einfluß fremder Litteraturen auf ihren Verlauf von Beginn einer schriftlichen Überlieferung an bis heute in graphischer Darstellung. Stuttgart, Göschen, 1890. (2. Tausend: Stuttgart, Göschen, 1890; 3. Tausend: Berlin, Behr, 1890)

Minor, Jakob: [Review of Flaischlen’s Dissertation on Otto von Gemmingen.] Anzeiger für deutsches Alterthum und deutsche Litteratur, XVII, 2 (April 1891), pp. 147–149. (digizeitschriften.de)

Nebrig, Alexander: Vergleichen als Wissenschaft: Zur Fachgeschichte. Evi Zemanek, Alexander Nebrig (eds.): Komparatistik. Berlin, Akademie Verlag, 2012, pp. 35–36. (books.google.com)

Rosenberg, Daniel; Grafton, Anthony: Cartographies of Time. Princeton Architectural Press, New York, 2010, pp. 143–149.

Schmidt, Erich: Wege und Ziele der deutschen Literaturgeschichte. Eine Antrittsvorlesung. Charakteristiken. Vol. I. Berlin, Weidmann, 1886, pp. 480–498. (archive.org)

Weschenfelder, Anke: Flaischlen, Cäsar (Otto Hugo). Deutsches Literatur-Lexikon. Das 20. Jahrhundert. Vol. 9. Zurich/Munich, K.G. Saur, 2006, cols. 33–37. (books.google.com)

Wetz, Wilhelm: Shakespeare vom Standpunkte der vergleichenden Litteraturgeschichte. Vol. I: Die Menschen in Shakespeare Dramen. Hamburg, Haendke & Lehmkuhl, 1897. (archive.org)

Thanks a lot …

… to Kurt Ubelhoer for copyediting this post!!!1!

Mistranslated Dates in Literature

2016-09-29T00:00:00+00:00

(This posting is about a phenomenon that I can only partly explain. It seems to be more than just a couple of random incidents. Maybe somebody else has some hints?) (Update Oct 3, 2016: Jürgen Hermes closely examined the “Count of Monte Cristo” example and explains some of the “mistranslations”, scroll to the bottom of the page for more info.)

When translating numbers and dates, strange things happen at times. Some of them are still explicable, one of which is the following. Take “Othello”, Act I, beginning of Scene 3 (Wikisource), just look at the numbers:

DUKE OF VENICE
There is no composition in these news
That gives them credit.

FIRST SENATOR
Indeed, they are disproportion’d;
My letters say a hundred and seven galleys.

DUKE OF VENICE
And mine, a hundred and forty.

SECOND SENATOR
And mine, two hundred:
(…)

Here’s how Wolf Graf Baudissin translated the passage into German (Gutenberg-DE):

HERZOG
In diesen Briefen fehlt Zusammenhang,
Der sie glaubwürdig machte.

ERSTER SENATOR
Jawohl, sie weichen voneinander ab;
Mein Schreiben nennt mir hundertsechs Galeeren.

HERZOG
Und meines hundertvierzig.

ZWEITER SENATOR
Meins zweihundert.
(…)

Ok, none of the characters seems to know the exact number of ships anyway, so why not translate 106 instead of 107? Voss also translated 106 (Google Books), while Wieland sticked to the 107 (Wikisource).

As insinuated before, that one’s easy to solve. Wieland chose to translate “Othello” in prose, the other two wanted to stick to Shakespeare’s blank verse. So for metrical reasons, they had to get rid of the second syllable of “sieben”, and how do you do that? Make it a “sechs”. Things like that happen a lot in metrical translations. (And I have to thank Thomas Efer for this and many other examples.)

So far, so good. Let’s now move from wrong numbers to wrong dates.

Strange Case of Dr Jekyll and Mr Hyde

At the beginning of chapter nine of R. L. Stevenson’s infamous novella, Dr Lanyon receives a letter from Jekyll. Lanyon says he received it on the 9th of January and that he had dined with Jekyll the day before, so obviously, the letter must have been written on the 8th or 9th of January. Yet it says in all editions of Stevenson’s text: “10th December, 18—” (Wikisource). This inconsistency has received some coverage (like here) and, for one, has been corrected in the Spanish translation which is also on Wikisource, the letter there dates from “9 de enero de 18…”. The German translation on Gutenberg-DE chose to omit the dating to mitigate the temporal inconsistency.

So these changes are still utterly explainable. But there are other cases.

The Count of Monte Cristo

I’m aware that the doctrines of literary translations changed over the centuries, and this is really a research field of its own. But the oddities I want to talk about go beyond that. Let’s have a look at the opening sentence of “Le Comte de Monte-Christo” by Alexandre Dumas:

Le 24 février 1815, la vigie de Notre-Dame de la Garde signala le trois-mâts le Pharaon, venant de Smyrne, Trieste et Naples. (Wikisource)

The Spanish translation goes along with this, no problem:

El 24 de febrero de 1815, el vigía de Nuestra Señora de la Guarda dio la señal de que se hallaba a la vista el bergantín El Faraón procedente de Esmirna, Trieste y Nápoles. (Wikisource)

Yet the German translation by Max Pannwitz, hosted on Gutenberg-DE, starts with this sentence:

Am 25. Februar 1815 fuhr der Dreimaster Pharao langsam und wie zögernd in den Hafen von Marseille.

This is in line with an earlier translation by August Zoller (see Google Books).

And the Polish translation on Wikisource has this:

W słonecznym i bardzo ciepłym dniu 27 lutego 1815 r. dano z wieży kościoła Notre Dame de la Garde, w Marsylji, hasło, zapowiadające powrót z podróży do Smyrny, Tryjestu i Neapolu — trójżaglowca “Faraon”.

Now what? 24th, 25th, 27th of February? I don’t see any reason to change the date of the arrival of the ship.

This is also not the fault of buggy OCR like in the case of the (non-existant) underreprentation of the 11th day of the month in the Google Ngram corpus.

Anna Karenina

Let’s check on Tolstoy, “Anna Karenina”, Part 3, Chapter 23, which in its original shape looks like this (Wikisource):

В понедельник было обычное заседание комиссии 2-го июня.

Yet the German translation hosted on Gutenberg-DE has this:

Montags war die gewöhnliche Sitzung der Kommission vom zweiten Juli.

2nd of June becomes 2nd of July. Which could be an honest and simple mistake, because of the similarity of the month names.

Let’s stay with the same book and jump to Part 4, Chapter 6, first sentence:

Алексей Александрович одержал блестящую победу в заседании комиссии семнадцатого августа, (…). (Wikisource)

The Spanish translation on Wikisource has this:

Karenin obtuvo una brillante victoria en la sesión celebrada por la Comisión el 1 de agosto, (…)

17th of August becomes 1st of August. Explanation could be simple again: Maybe the translator just forgot a number here?

This is also not a Julian/Gregorian calendar issue. The Julian calendar was 12 days behind in the 19th century, which doesn’t explain the difference between the two dates.

A Descent into the Maelström

Next example, Poe’s “Descent into the Maelström”, where it says:

It was on the tenth day of July, 18—, a day which the people of this part of the world will never forget—for it was one in which blew the most terrible hurricane that ever came out of the heavens. (Wikisource)

A Spanish translation goes like this:

Dentro de pocos días se cumplirán tres años desde que sucedió lo que voy a relataros. Era el 10 de agosto de 18—, día que la gente de este lado del mundo jamás olvidará, porque se desató el huracán más formidable que jamás envió el cielo. (Wikisource)

July? August!

Sherlock Holmes

Let’s do some Sherlock Holmes. The Hound of the Baskervilles!

Dr. Mortimer drew a folded newspaper out of his pocket.
“Now, Mr. Holmes, we will give you something a little more recent. This is the Devon County Chronicle of May 14th of this year. (…)” (gutenberg.org)

The same passage in Spanish translation:

El doctor Mortimer se sacó del bolsillo un periódico doblado.
–Ahora, señor Holmes, voy a leerle una noticia un poco más reciente, publicada en el Devon County Chronicle del 14 de junio de este año. (…) (biblioteca.org.ar)

May? June! – Another shift of one month when translating from English to Spanish, maybe a pattern? Well, maybe not. Let’s have some Stendhal:

The Charterhouse of Parma

Le 7 mars 1815, les dames étaient de retour, depuis l’avant-veille, d’un charmant petit voyage de Milan ; elles se promenaient dans la belle allée de platanes, récemment prolongée sur l’extrême bord du lac. (Wikisource)

Which, in a Spanish translation, looks like this:

El 7 de mayo de 1815 hacía dos días que las señoras habían vuelto de un precioso viajecito a Milán; estaban paseándose por la hermosa avenida de plátanos que había sido prolongada hacía poco tiempo hasta el borde mismo del lago. (edu.mec.gub.uy)

March? May!

Jules Verne

Le 29 mai de cette année-là, un berger surveillait son troupeau à la lisière d’un plateau verdoyant, au pied du Retyezat, qui domine une vallée fertile, boisée d’arbres à tiges droites, enrichie de belles cultures. (Wikisource)

The same passage in Spanish:

El 19 de mayo de aquel año, un pastor apacentaba su rebaño a la orilla de un verde prado, al pie del Retyezat, que domina un valle fértil, cubierto de árboles de rama-je recto y enriquecido con bellas plantaciones. (biblioteca.org.ar)

29th? 19th!

Explanation(s)?

For all we know, this last example could be another honest mistake (mistyping a number, happens). Maybe all these examples can be traced back to simple mistakes that go unnoticed by the common reader (unless you’re part of a research project whose purpose it is to extract dates from literary corpora :-)… Anyway, we got more findings like this. You’ll find some examples in TIWOLI, our app for Android and iOS. We didn’t dare to correct wrong dates, we wanted to preserve them as they were. After all, that’s how every reader of these elder translations encountered them.

Many of the translations are from the 19th or early 20th century when transations didn’t yet meet our idea of a proper, authentic translation. For example, the first German translation of Huxley’s “Brave New World” changed the location from “London” to “Berlin” (which was authorised by the author, by the way; but this wouldn’t pull through nowadays, I guess).

Some sort of weird adaptation could be the reason for some of the above-cited examples. But I really don’t know. The question is: Are all these examples plain mistakes? Or are there reasons for the way those dates were mistranslated?

(For hints, drop me a line on Twitter, via mail, or use the comment field below.)

Update

On October 3rd, Jürgen Hermes published a great follow-up to this post (in German): “Zum Tag der Daten-Einheit”. In which he interprets the “mistranslations” relating to “The Count of Monte Cristo” as “belated copyediting” by German and Polish translators. Especially the Polish translator applied some effort to repair the flawed timeline of the original novel. The same effort might be behind some of the other mistranslations, too. The example from “Jekyll and Hyde” was also pointing into that direction.

Introducing TIWOLI ("Today in World Literature"), a Calendar App of Fictional Events

2016-09-07T00:00:00+00:00

This could become a whole new series: Fancy by-products of Digital Humanities research projects…

Short version: Download TIWOLI for Android here. Download TIWOLI for iOS here.

Update (Oct 29, 2016): The iOS version is finally available in the App Store. The article was updated accordingly.

And now let’s jump right into it and name seven famous days in world literature:

The day of James Joyce’s “Ulysses”? (Of course, Bloomsday, Lá Bloom, June 16!)

That is an easy start, but the world beyond this über-day of world literature is trickier. Most people will know about the fictional events we’re going to itemise, but won’t typically have the exact dates in mind. Let’s check:

On which day did Robinson Crusoe set foot on his island? (September 30) On which day did he eventually return to England, after 35 years of absence? (June 11)
When is the deadline for Phileas Fogg’s travel around the world in eighty days? (December 21)
What’s Tristram Shandy’s birthday? (November 5) – Funny enough, there still seem to be people drinking to Tristram Shandy’s health.
When did Goethe’s “Young Werther” commit suicide? (December 21)
On what day did Charlie visit the Chocolate Factory? (February 1) – Especially interesting to note here that in the 1971 movie, the golden ticket refers to “the first day of October” for whatever reason.

So this is the subject of our blogpost, (in)famous days in world-literary fiction.

How did this come about? At the annual ADHO Digital Humanities conference 2015 (DH2015) in Sydney, we delivered a talk called “When does (German) literature take place?” (slides) In order to answer that unusual question, we extracted exact dates and months from a corpus of 2,700+ works of German-language fiction from 1510 to the 1940s. Our results suggested that “the marvellous month of May” (Heinrich Heine) is preeminent in our corpus of German fiction. There was already enough evidence to claim that May is the favourite month of poets (at least in the northern hemisphere), and Heine is just one example. But May’s preeminence in prose was a real finding and there are a bunch of explanations, but this (along with many other insights yielded by this whole date-extraction-from-fiction business) shall happen in the corresponding research paper, not here.

As a by-product (and no more than that), we coded a native Android app to showcase the more exciting and self-explanatory of our results. We called our app TIWOLI (short for “Today in World Literature”) and made it available via the Google Play Store. This was followed by the iOS version, done by our dear friend Thomas Bögel.

Now what TIWOLI does is give you the opportunity to zigzag through an entire world-literary year, a calendar filled with fictional happenings. Here some screenshots with some of the examples from above:

We’re currently curating three corpora for the app, in English, Spanish and German. In some cases, you can even switch between translations of one and the same passage (not in all cases, though, the contents of our corpora are very different per language). This is a 1st-of-January example from Balzac’s novel “Eugénie Grandet”:

The interface was also translated into the three languages (it will fall back to English if your system language cannot be matched). Here are the three versions of our welcome screen, plus the English version of the settings page:

If you like you can set the alarm so it informs you about fictional events that “happened” on the current day.

It was not all that easy to fill up a whole year with suitable quotes for every day. To replenish the English corpus with nice enough quotes we are very grateful to Mark Algee-Hewitt and the Stanford Literary Lab who helped us out with their corpus of English-language fiction. Let’s give you some more:

For each corpus, the app contains quotes from translated world-literary works, too, because after all, weltliteratur is the raison d’être of this blog:

The dates from the Russian 19th-century novels point to the Julian calendar, of course (which was 12 days behind the Gregorian back then). Some more translated literature:

For every quote we provide a link to the full-text source (gutenberg.org, gutenberg.net.au, archive.org, books.google.com, …). Epitexts and links/metadata are generated automatically. The portraits/pictures were not chosen by us, we’re relying on the ‘principal image’ provided by Wikidata, a great way to easily enrich an app. Let’s give you the Defoe examples from above, accompanied by Poe (for the sake of rhyming names):

Some more German literature in translation, including the aforementioned destiny of storm-and-stressy Werther:

The first two quotes are from Fontane’s “Effi Briest”, in many respects the righteous German counterpart to “Madame Bovary”. The second quote alludes to Effi’s and Instetten’s wedding which coincides with the national day of Germany (not a good omen, given the tragic outcome of the novel 😇).

Four final examples from the English corpus:

The last quote, from “Jekyll & Hyde”, is interesting insofar as the given date can be regarded a mistake by the author that went into several print editions (as discussed here, for example).

We also harvested some newer books as you can see in the Auster example below (the picture is not the best, but somebody might upload a better one to Wikimedia Commons someday). There’s also one more Brontë quote and two nice quotes from the time-travel novel by Bellamy (where times and dates are the actual subject):

Now let’s sneak a peek at the Spanish corpus. You can easily switch to the other corpora at any point from within the app.

We can also follow the chronology of Borges’ cuento “La muerte y la brújula” (“Death and the Compass”):

Some concluding Spanish examples:

On to the German-language corpus:

The first quotation, from Georg Büchner’s novella fragment “Lenz”, is probably one of the most famous first sentences. On the second screenshot, Schnitzler’s “Leutnant Gustl” is pondering his destiny on 4th of April, resulting in not killing himself. On the third one, the government of the Martian States takes control over our planet, because we earthlings are not capable of maintaining peace (the author of the passage, Kurd Lasswitz, can be regarded a pioneer of German sci-fi lit). The fourth quote is especially interesting, since the “27th of July” is also discussed in a “meta” way, as a combination of a printed number and a word, which serves as intellectual nourishment for the imprisoned intellectual (if you didn’t install the app yet, check the English translation of this passage on Google Books).

We already featured Theodor Fontane with some paragraphs translated to English. But what is it with the Prussian author and the 3rd of October? It is an eventful day in many of his novels, and we’re not sure if anyone has looked into that yet:

Fitfully, we included some quotations from contemporary fiction. The given sources contain page numbers as they point to print editions of the works:

(source, p. 287)
(source, p. 275)
(source, p. 116)
(source, p. 31)

Some more:

(source, p. 11)
(source, pp. 72)
(source, p. 144)

The first quote from Arno Schmidt’s 1972 novel “The School for Atheists” points to a future date, the 7th of October, 2014. When we traversed this point in time two years ago, at least one major newspaper published an article about this long-imminent event (German weekly “Die Zeit”: “The story starts now.”).

How about Jean Paul, the greatest of all May lovers:

And last not least, some Christmas quotes:

(source)
(source, p. 204)
(source)
(source, p. 582)

So much for our app, TIWOLI. Also if it wasn’t part of the original research, it helped to sharpen our notion of dates and month names in fictional texts. Literature as a genre has an inclination to waive exact dates and prefers temporal settings like “at the beginning of November”, “It was a dark and stormy night”, etc. But sometimes exact dates do occur, of course. And if they do, they have a tendency to mean something. (Unless, of course, they occur in epistolary, adventure or historical novels which are usually full of exact dates. We tried to exclude them from the app and just used some as fill-ins. Goethe’s “Werther”, all Jules Verne novels and Benito Pérez Galdós proved especially useful. 😊)

Before we conclude this admittedly BuzzFeed-like über-post with far too many pics we have two cliffhangers for you:

Our friends at the University of Cologne (Spinfo/CCeH) are currently building a web interface with our data. Prototype looks very promising.
There’s an easter egg somewhere in the app. Whoever finds it first is our hero/ine!

Contributors to World Literature – Identifying Writers in Wikipedia, Part II

2016-08-19T00:00:00+00:00

In the first post of this 2-part series we introduced six features of Wikipedia pages on writers that might help us figure out if those pages are, in fact, about (literary) writers. Today we want to discuss which of these features are actually suited to reliably identify writers in Wikipedia. Our goal is to build a set of contributors to world literature that we can further investigate (and keep updated, as Wikipedia evolves). Instead of providing a conclusive answer here, this post will concentrate on observations that could eventually help to implement an approach for the automated identification of literary writers in Wikipedia.

First Sentence

This feature is available throughout all language versions.
The automated analysis of sentences requires some sophisticated natural language processing which should also be available for different languages. The quality of such approaches highly depends on the chosen language.
The results are quite ambiguous, since even within one language version there is no standardised way to describe writers.
DBpedia has implemented a framework for fact extraction which could potentially extract this information.

At first sight, a “Bibliography” section seems to be the prevalent way to list original works of an author in the English Wikipedia. But not all writers have one. Also, the name of the section is not consistent (e.g., in the article on John Milton it is called “Works”). “Bibliography” can also refer to a list of secondary literature (see the article on Gabriel García Márquez).
Albeit its ambiguity, finding the “Bibliography” headline in the wiki source code is easy. Finding equivalent headings in other language versions would require some effort, though.
It is unclear if and how this feature is used in other language versions.
The section structure of articles is currently not provided by DBpedia.

Occupation Property

This property often contains several (ambiguous) values without further distinctions (“writer”, “novelist”, “poet”, etc.).
Values of this property are highly language-dependent, it is rather difficult to make use of it in a cross-language environment.
Not all language editions permit infoboxes for writers (among others, the German edition).
If available, the infobox properties are extracted by DBpedia. They can be found in the infobox properties dataset and in the mapping-based properties dataset. The latest dataset contains this information in the infobox properties dataset:

<http://dbpedia.org/resource/John_Irving> <http://dbpedia.org/property/occupation> "Novelist"@en .
<http://dbpedia.org/resource/John_Irving> <http://dbpedia.org/property/occupation> "Screenwriter"@en .

List of Works

In our example page on John Irving, this feature was implemented using a specific John Irving template. Other popular/canonised writers also have that kind of template, yet in general, this feature is not very widespread.
There are categories for such templates, like Category:Novelist navigational boxes, making it easy to check which writers have such a template.
Apparently, there are only two other language editions with such a template, Romanian and Farsi.

Writer Template

Not all language editions have or use such template. E.g., the German Wikipedia does not permit the writer template and, thus, its page on Johann Wolfgang von Goethe lacks an infobox.
Each page can be assigned to exactly one template, which means that some persons do not have a writer template, although we’d clearly like to have them in our set. This includes Franz Kafka (neutral person template) and Umberto Eco (tied to a philosopher template).
The association of Wikipedia pages to templates has been extracted by DBpedia in the instance types dataset. The value is provided in the type property. For John Irving, it contains the following information:

<http://dbpedia.org/resource/John_Irving> <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/ontology/Writer> .

Conclusion

Summarising the above findings in a table, we get the following result:

feature	language editions	usage within language editions	DBpedia dataset	effort	ambiguity
first sentence	all	all	-	high	high
table of contents	all	some	-	low	medium
occupation property	some	many	infobox properties	medium	high
list of works	few	few	-	low	low
categories	all	many	article categories and skos categories	high	high
writer template	some	many	instance types	low	low

Given the large variety of properties the different features have, it is quite difficult to devise an approach to identify writers on Wikipedia that works across different language versions. This 2-part blog post was a preliminary introduction to get a sense of the problem. We proposed and applied a solution that basically works across different language versions in a still unpublished paper (on which we will keep you posted) and will introduce a pragmatic and manageable set that can be used for a variety of purposes in one of our next blog posts.

Contributors to World Literature – Identifying Writers in Wikipedia, Part I

2016-08-18T00:00:00+00:00

The first challenge we face when trying to analyse the representation of world literature on Wikipedia is the automatic identification of literary writers and works across all 280+ language versions. Our last posts have shown how to extract works that are possible candidates for a Wikipedia-inherent world-literary canon (Wikidata-based and DBpedia-based). In this 2-part post we focus on the extraction of writers, i.e., possible contributors to world literature.

For a start, let us have a detailed look at the Wikipedia page of John Irving:

This web page has several features indicating that John Irving is a writer we want to have in our set. Let’s introduce and discuss six of them:

The first sentence says that John Irving is a novelist and screenwriter:
The table of contents lists a bibliography:
The infobox on the right has an occupation property with values “Novelist” and “Screenwriter”:
Scrolling down, the John Irving template assembles a list of works by the author:
Scrolling down further, we find a list of categories, among them 20th-century American novelists, 21st-century American novelists, American feminist writers, and American male screenwriters:
In addition, by looking at the wiki source code of the page we realise that the infobox uses the writer template:

To leverage one of these features to automatically identify John Irving (and others, of course) as a writer, we have to answer these questions:

Which of the information is universally available (i.e., in different language versions)?
Which of the information is available through DBpedia or Wikidata?
How widespread is the information used?
How easy is it to actually use the information to identify writers?

We will discuss these questions tomorrow in the 2nd part of this blog post.

DBpedia and World Literature

2016-05-30T00:00:00+00:00

After our blog post “Wikidata Meets World Literature” some might have mumbled into their tea cups, “ok yeah, but what about DBpedia?” Accordingly, let us add some technological diversity to our initial experiment. DBpedia is a great project and follows its very own approach, it is also a few years older than Wikidata. More on the differences between the two projects can be found on Quora.

So, DBpedia and World Literature. Let’s kick off by translating/simplifying the original SPARQL query to match the DBpedia ontology:

SELECT ?s ?desc ?authorlabel
WHERE {
  ?s rdf:type dbo:Book .
  ?s dbo:author ?author
  OPTIONAL {
    ?s rdfs:label ?desc FILTER (lang(?desc) = "en").
  }
  OPTIONAL {
    ?author rdfs:label ?authorlabel FILTER (lang(?authorlabel) = "en").
  }
}

This query returns a list of resources that have been assigned the class Book in the DBpedia ontology, along with their authors. (Yes, just click, this will auto-execute the query and show you the results.)

In our original post we used the number of Wikipedia language versions per book to rank them, with “One Thousand and One Nights” taking away all the glory (for the time being, that is). We can try something similar in DBpedia by counting the number of labels per book. Each Wikipedia language edition from which a resource was extracted by DBpedia has its label stored using the rdfs:label property. This is our query:

SELECT ?s ?desc ?authorlabel (COUNT(DISTINCT ?label) as ?labelcount)
WHERE {
  ?s rdf:type dbo:Book .
  ?s rdfs:label ?label .
  ?s dbo:author ?author
  OPTIONAL {
    ?s rdfs:label ?desc FILTER (lang(?desc) = "en").
  }
  OPTIONAL {
    ?author rdfs:label ?authorlabel FILTER (lang(?authorlabel) = "en").
  }
} GROUP BY ?s ?desc ?authorlabel ORDER BY DESC(?labelcount)

Unfortunately, the output is a bit disappointing:

s	desc	authorlabel	labelcount
:The_Adventures_of_Tom_Sawyer	“The Adventures of Tom Sawyer”@en	“Mark Twain”@en	12
:Strange_Case_of_Dr_Jekyll_and_Mr_Hyde	“Strange Case of Dr Jekyll and Mr Hyde”@en	“Robert Louis Stevenson”@en	12
:Fifty_Shades_of_Grey	“Fifty Shades of Grey”@en	“E. L. James”@en	12
:Gray’s_Anatomy	“Gray’s Anatomy”@en	“Henry Gray”@en	12
:Sense_and_Sensibility	“Sense and Sensibility”@en	“Jane Austen”@en	12
:A_Brief_History_of_Time	“A Brief History of Time”@en	“Stephen Hawking”@en	12
:Mein_Kampf	“Mein Kampf”@en	“Adolf Hitler”@en	12
:Crime_and_Punishment	“Crime and Punishment”@en	“Fyodor Dostoyevsky”@en	12
:Catching_Fire	“Catching Fire”@en	“Suzanne Collins”@en	12
:David_Copperfield	“David Copperfield”@en	“Charles Dickens”@en	12
:Mansfield_Park	“Mansfield Park”@en	“Jane Austen”@en	12
:Les_Liaisons_dangereuses	“Les Liaisons dangereuses”@en	“Pierre Choderlos de Laclos”@en	12
:Murder_on_the_Orient_Express	“Murder on the Orient Express”@en	“Agatha Christie”@en	12
:The_Sign_of_the_Four	“The Sign of the Four”@en	“Arthur Conan Doyle”@en	12
:The_Republic_(Plato)	“The Republic (Plato)”@en	“Plato”@en	12
:The_Communist_Manifesto	“The Communist Manifesto”@en	“Friedrich Engels”@en	12
:From_the_Earth_to_the_Moon	“From the Earth to the Moon”@en	“Jules Verne”@en	12
:Buddenbrooks	“Buddenbrooks”@en	“Thomas Mann”@en	12
:The_Brothers_Karamazov	“The Brothers Karamazov”@en	“Fyodor Dostoyevsky”@en	12
:The_Trial	“The Trial”@en	“Franz Kafka”@en	12

Seems that books have been extracted from only 12 language editions (ar, de, en, es, fr, it, ja, nl, pl, pt, ru, zh). Or, only these 12 languages feature a page template for books. Or … well, let’s stop speculating and have a look: DBpedia features a page with statistics about the extracted data and we can see in the “Cross-Language Instance Overlap” table that there are, for instance, 268 books appearing in 16 language editions. However, not all datasets from all language versions are available at the public SPARQL endpoint. This list shows that currently a “labels” dataset has been loaded for exactly the 12 language versions we mentioned above. For the same reason, other properties like dbo:abstract or owl:sameAs that could be linked to the number of language editions show the same behaviour. (For an overview of potential properties take a look at the DBpedia entry on “The Adventures of Tom Sawyer”.)

When looking for alternatives we found the PageRank dataset by Andreas Thalhammer. Fortunately, it is deployed on the official DBpedia SPARQL endpoint and so, instead of counting the number of language editions, we can easily use the PageRank of each page within the English Wikipedia as a measure of importance:

PREFIX vrank:<http://purl.org/voc/vrank#>

SELECT ?s (SAMPLE(?desc) AS ?label) (GROUP_CONCAT(?authorlabel, ', ') AS ?author) (MAX(?v) AS ?rank)
FROM <http://dbpedia.org>
FROM <http://people.aifb.kit.edu/ath/#DBpedia_PageRank>
WHERE {
  ?s rdf:type dbo:Book .
  ?s vrank:hasRank/vrank:rankValue ?v .
  ?s dbo:author ?author
  OPTIONAL {
    ?s rdfs:label ?desc FILTER (lang(?desc) = "en").
  }
  OPTIONAL {
    ?author rdfs:label ?authorlabel FILTER (lang(?authorlabel) = "en").
  }
} GROUP BY ?s ORDER BY DESC(?rank)

We modified the query so it groups authors of multi-author books into one row. Despite some Wikipedia-related distortions, the result is much more meaningful:

s	label	author	rank
:The_World_Factbook	“The World Factbook”@en	“Central Intelligence Agency”	146.277
:Systema_Naturae	“Systema Naturae”@en	“Carl Linnaeus”	68.6522
:Natural_History_(Pliny)	“Natural History (Pliny)”@en	“Pliny the Elder”	64.9683
:On_the_Origin_of_Species	“On the Origin of Species”@en	“Charles Darwin”	56.7624
:The_Rolling_Stone_Album_Guide	“The Rolling Stone Album Guide”@en	“Anthony DeCurtis, Dave Marsh”	47.9486
:Don_Quixote	“Don Quixote”@en	“Miguel de Cervantes”	45.3332
:All_Music_Guide_to_Jazz	“All Music Guide to Jazz”@en	“Vladimir Bogdanov (editor), Stephen Thomas Erlewine”	44.641
:Alice’s_Adventures_in_Wonderland	“Alice’s Adventures in Wonderland”@en	“Lewis Carroll”	42.4669
:All_Music_Guide_to_the_Blues	“All Music Guide to the Blues”@en	“Vladimir Bogdanov, Stephen Thomas Erlewine”	40.9857
:Records_of_the_Grand_Historian	“Records of the Grand Historian”@en	“Sima Qian”	40.262
:Nineteen_Eighty-Four	“Nineteen Eighty-Four”@en	“George Orwell”	39.9243
:The_Republic_(Plato)	“The Republic (Plato)”@en	“Plato”	38.9124
:The_Wealth_of_Nations	“The Wealth of Nations”@en	“Adam Smith”	37.4529
:Euclid’s_Elements	“Euclid’s Elements”@en	“Euclid”	36.0581
:Paradise_Lost	“Paradise Lost”@en	“John Milton”	32.8596
:Moby-Dick	“Moby-Dick”@en	“Herman Melville”	32.632
:Encyclopædia_Iranica	“Encyclopædia Iranica”@en	“Ehsan Yarshater”	30.9694
:Dracula	“Dracula”@en	“Bram Stoker”	29.6592
:Histories_(Herodotus)	“Histories (Herodotus)”@en	“Herodotus”	29.4831
:The_Hobbit	“The Hobbit”@en	“J. R. R. Tolkien”	29.4576

These are the top 20 books, ranked by their PageRank within the English Wikipedia. One thing still blocking the view a bit are the popular encyclopedic works, of course. The many, many in-links earned by the “World Factbook” or the “Rolling Stone Album Guide” will not speak for their unrivalled literary quality, supposedly. Sorting out the literary works from this set of (all kinds of) books would be the obvious next step.

Empirical Data on Over-Length Books

2016-05-25T00:00:00+00:00

(Little note upfront: The technical part of this posting involves Docker, RDF→XSLT→JSON, Elasticsearch and Kibana. More on this at the bottom of this page or without further ado on our GitHub repo. Our solution is the result of nothing more than a 4-hour hackathon, so don’t expect anything polished.)

As a preliminary for a study of over-length books (novels, primarily) we wanted to gather more empirical data. There is still this “List of longest novels” in the English Wikipedia, but this list is problematic, because it’s obviously canon-driven and uses completely different measures to define the extent of books. We also gathered some evidence on our own, a list with novels of more than a thousand pages, on another blog (the explanatory text there is in German, watch out). Our list is sorted chronologically, since this is one aspect we have to take into account when planning to pen something like a “History of the Over-Length Novel”.

Using the number of pages as measurement is, of course, part of the problem. It would be better to work with number of letters or words per work, but this is not (yet?) part of bibliographic metadata. So, number of pages. How do we access them, large-scale?

German National Library Enters the Stage

Let’s start with this panoramic view of the German Library in Leipzig, predecessor and now part of the German National Library (DNB) – please also note the book towers on the left side:

Source: Wikimedia Commons (CC BY-SA 3.0)

Really nice, yes, but it’s another kind of picture of the actual German National Library that we’re going to show you, one that is much more to the point. When we came across the data-service page of the German National Library, we were almost enthusiastic. They offer different kinds of sets in different formats (all licenced under CC0 1.0!), from which we chose the “DNBTitel.rdf.gz” one comprising the records for all books/items archived at the DNB library (the dump was generated on March 10, 2016, and is 1,5 GB in size, which makes for an uncompressed 21,3 GB). There’s no SPARQL endpoint (yet?) by which users could query the catalogue data directly (but you can register for access via OAI and SRU). So, as said before, we decided to download the RDF file and started to build our own query thing.

Now, the other picture of the German National Library we wanted to show you is this:

This is the matrix view of the RDF file we downloaded, a 3D scatter plot of triples generated with HDT-it!. Each predicate there has a different colour. In this front view, subjects meet objects, but we can see a pink predicate gleaming through, we’re looking at ISBD element P1053 (“has extent”), the thing we’ll be looking at in the following:

Some Results

Btw, we started this little project as part of a spontaneous 4-hour hackathon last July, at the R&D department of the Göttingen State and University Library. Now, just about a year later, we decided to wrap it up a bit and show you how we made it work.

First and foremost, you have to be aware of the history of the German National Library to know what you can expect from any query result. The gist of it is: They were a bit late compared to other national libraries in Europe and began collecting “all German and German-language publications from 1913, foreign publications about Germany, translations of German works, and the works of German-speaking emigrants published abroad between 1933 and 1945” (quoting the official “About us” page).

Before we bore you with how we did it, let’s go for some results. As a proof of concept, let’s see which authors are the ones with the most books in the catalogue (dcterms:creator), let’s generate a top 25:

In the bar chart above, the authors are identified by their GND records. Very well then, let’s resolve them:

DNB Identifier	Number of Records	Author
118540238	5792	Goethe, Johann Wolfgang von
118617443	3881	Steiner, Rudolf
118577166	3402	Mann, Thomas
11855042X	3336	Hesse, Hermann
11856515X	3227	Konsalik, Heinz G.
118637479	2995	Zweig, Stefan
118542257	2840	Grimm, Jacob
118578537	2732	Marx, Karl
118542265	2721	Grimm, Wilhelm
118530380	2504	Engels, Friedrich
12002179X	2478	Schaal, Eric
118607626	2405	Schiller, Friedrich
118618725	2133	Storm, Theodor
118514768	2132	Brecht, Bertolt
118559230	2100	Kafka, Franz
118559206	2086	Kästner, Erich
118818651	2080	May, Karl
118613723	2070	Shakespeare, William
11856109X	1828	Keller, Gottfried
118512676	1818	Böll, Heinrich
118601024	1794	Rilke, Rainer Maria
118587943	1781	Nietzsche, Friedrich
118534262	1755	Fontane, Theodor
118533436	1677	Fischer, Marie Louise
118520628	1674	Christie, Agatha

Looks plausible, in a way. And interesting enough for an interpretation (which, for the time being, we won’t deliver). One thing becomes clearer now, though, we should really talk about “items”, rather than “books”. Photographer Eric Schaal, for example, didn’t get into this top 25 by writing more than two thousand books. To be honest, we made a top 25 just to get at least some women into this board of men, ranking 24th and 25th. And stating the obvious, Goethe also didn’t write almost six thousand books since 1913, what really puts weight on the authors is the substantial number of re-editions, of course.

Number of Books/Items in the Catalogue

We’ve got 11.373.862 items altogether (some didn’t make it into the Elasticsearch index since we didn’t really address error handling or validation; the regexps in our XSLT weren’t perfect either, things we can improve next time we’re not high-speed hackathoning). Anyway, in 5.874.504 cases we successfully parsed the isbd:P1053 element (= ”has extent”) into a usable number of pages, summing up to a total number of 969.846.170 pages. The max number of pages in this set is 2.711.111, which is obviously the result of a metadata apocalypse, somebody must have slipped on the 1 key (this is the book that’s said to have more than two million pages) (ed. 05/27/16: n° of pages has been corrected today, they notified us via Twitter, nice!).

You can glean from our XSLT file that we’re only using extent information if we found a number succeeded by “ S.” (= ”pages”) in the isbd:P1053 string. So we’re not using at least 5.499.358 items out of the 11.373.862. Either they had no information on the book extent/number of pages or we didn’t parse it because we just used a very basic pattern. But with this simple method we still managed to cover 51,65% of all the books/items stored in the German National Library. We can sure improve our data extraction, but for the time being we’re good with what we have. After all, we’re still speaking about almost six million book records.

Now onto some more meaningful stuff. Let’s take five major publishers and compare them just by looking at the extent of their books. We held a little powwow and, full of intentional bias, chose Aufbau, Eichborn, Hanser, Rowohlt, Suhrkamp. Let’s have a look at the number of items per publisher in the catalogue:

This is an interesting perspective, also if this comparison makes not much sense. For all we know, there could be thousands of re-editions involved. Well, okay.

Average Number of Pages per Book per Publisher

So obviously, the average Suhrkamp book beats the average Rowohlt book beats the average Eichborn book when considering the number of pages. This comparison is intriguing, but it shouldn’t be taken too literally, we’re dealing with a certain amount of incorrect metadata as we’ll see in the next set of lists.

Longest Books per Publisher

This brings us a bit closer to our goal. But these lists also show the problems of erroneous data and the need for additional metadata. If we want to look into over-length novels, we’ll obviously need another indicator, one of which is not provided by the current DNB dataset. And once again, the lists clarify that it’s more appropriate to speak of “item” than of “book”, but now let’s start with the rankings:

Aufbau

DNB Identifier	Number of Pages	Author: Title (Year)
988488205	1359 S.	Vikram Chandra: Der Pate von Bombay (2009)
1001932447	1291 S.	Lew Tolstoi: Krieg und Frieden (2010)
576696420	1286 S.	Alexej Tolstoi: Der Leidensweg (2. Aufl., 1955)
576696439	1286 S.	Alexej Tolstoi: Der Leidensweg (3. Aufl., 1959)
1011565994	1243 S.	Hans Fallada: Wolf unter Wölfen (2011)
98848823X	1227 S.	Lew Tolstoi: Anna Karenina (2008)
945188846	1211 S.	Friedrich Gorenstein: Der Platz (1995)
988488272	1200 S.	Fjodor Dostojewski: Die Brüder Karamasow (2008)
949346470	1183 S.	Lew Tolstoi: Anna Karenina (1996)
451896025	1174 S.	G. W. F. Hegel: Ästhetik (1955)

Eichborn

DNB Identifier	Number of Pages	Author: Title (Year)
946561486	~~1814 S.~~	(wrong number of pages)
967526825	1222 S.	Leo Tolstoi: Krieg und Frieden (2003)
950298603	1081 S.	Rolf Vollmann: Die wunderbaren Falschmünzer (1997)
988571005	991 S.	Daniel Schwartz: Schnee in Samarkand (2008)
979687187	954 S.	Paul Verhaeghen: Omega minor (2006)
974540919	855 S.	David M. Crowe: Oskar Schindler (2005)
979691044	852 S.	Laurence Sterne: Leben und Ansichten von Tristram Shandy, Gentleman (2006)
840181604	841 S.	Fred Denger: Der grosse Boss (1984)
860929477	841 S.	Fred Denger: Der grosse Boss (6. Aufl., 1985)
870140876	841 S.	Fred Denger: Der grosse Boss (5. Aufl., 1985)

Hanser

DNB Identifier	Number of Pages	Author: Title (Year)
1022146394	~~4587 S.~~	(wrong number of pages)
99886398X	1810 S.	Walter Doberenz; Thomas Gewinnus: Visual C# 2010 (2010)
98401098X	1806 S.	Walter Doberenz; Thomas Gewinnus: Borland Delphi 7 (2007)
998863955	1802 S.	Walter Doberenz; Thomas Gewinnus: Visual Basic 2010 (2010)
760043035	1672 S.	Joseph von Eichendorff: Werke (4. Aufl., 1971)
451062442	1606 S.	Joseph von Eichendorff: Werke (2. Aufl., 1959)
451062817	1590 S.	Joseph von Eichendorff: Werke (1. Aufl., 1955)
453424740	1511 S.	Eduard Mörike: Sämtliche Werke (3. Aufl., 1964)
780204875	1511 S.	Eduard Mörike: Sämtliche Werke (5. Aufl., 1976)
970961294	1469 S.	Uwe Bünning; Jörg Krause: Windows XP Professional (3. Aufl., 2004)

Rowohlt

DNB Identifier	Number of Pages	Author: Title (Year)
944325807	2253 S.	Klaus Harpprecht: Thomas Mann (1995)
945394659	2253 S.	Klaus Harpprecht: Thomas Mann (16.–30. Tsd., 1995)
967713358	2026 S.	Karl Corino: Robert Musil (2003)
101781905X	1723 S.	Péter Nádas: Parallelgeschichten (2012)
1028105657	1723 S.	Péter Nádas: Parallelgeschichten (Taschenbuch, 2013)
1008548022	1719 S.	Rolf Hochhuth: Essayistische Prosa und Gedichte (2011)
575594950	1671 S.	Robert Musil: Der Mann ohne Eigenschaften (1952)
961281588	1642 S.	Rolf Hochhuth: Alle Erzählungen, Gedichte und Romane (2001)
457661054	1632 S.	Robert Musil: Der Mann ohne Eigenschaften (1970)
575594969	1632 S.	Robert Musil: Der Mann ohne Eigenschaften (23.–29. Tsd., 1960)

Suhrkamp

DNB Identifier	Number of Pages	Author: Title (Year)
946102384	~~3980 S.~~	(wrong number of pages)
945262094	~~2909 S.~~	(wrong number of pages)
991420225	2569 S.	Amos Oz: Die Romane (2009)
988814668	2085 S.	E. M. Cioran: Werke (2008)
986493635	1909 S.	Marguerite Duras: Die Romane (2008)
986531766	1840 S.	Thomas Bernhard: Die Romane (2008)
991398939	1838 S.	Hermann Hesse: Die Erzählungen und Märchen (2009)
988840758	1782 S.	Max Frisch: Romane, Erzählungen, Tagebücher (2008)
998413925	1782 S.	Bertolt Brecht: Prosa (2013)
1008349852	1735 S.	Alejo Carpentier: Die Romane (2011)

We won’t comment these lists now, although they make for some interesting discussions. You can obviously do much more with what we built here and we’ll certainly get back to this later. So let’s close this blog post with a short note on how we built this bridge from the freely available DNB catalogue data to the results shown above.

German National Library Goes Elasticsearch

The somewhat weird original idea we had when initiating the hackathon was this: We wanted to know how much the German National Library weighs on books, and we wanted to find out by the number of pages of all the books it stores which we then would have multiplied by the average weight of a book page. Well, you can do the maths yourself now, you can find the total number of pages we counted above and then extrapolate, don’t forget to divide this by two, it will be a good enough approximation.

What saves us time now is that we already described our mechanism on GitHub where we also provide all the info you need to rebuild our machine. We basically reorganised the whole thing as a Docker project, which will create a container running Elasticsearch/Kibana. The repo also features shell scripts for downloading the current version of the German National Library title catalogue. Some selected data fields from every book in that catalogue are then transformed into JSON and pushed to the Elasticsearch instance. After that you will be able to query the DNB catalogue data with Elasticsearch to create nice outputs with Kibana. As should be clear from the text above, the data fields we’re focusing on are mainly the number of pages per book and some book metadata (author, title, year, publisher, etc.) for identification.

That’s it for now. And lest we forget, special thanks to Max Brodhun and Carsten Thiel for writing the XSLT and helping with the shell scripting, “as quick as boiled asparagus”, so to speak, it was only because of them that we could go on with what we actually wanted to do. The next Uludağ is on us! ;)

Fefe Research Institute

2016-05-18T00:00:00+00:00

Ok, the headline is a joke, of course. But let’s start at the beginning. There are three things German computer scientists/IT guys must not miss: their Heise Online newsticker, their xkcd, and their daily Fefe.

Fefe is the nom de guerre of Felix von Leitner, a C programmer (check dietlibc), IT security adept and blogger, running his own speaker’s corner “Fefes Blog” since 2005. This blog is notorious for its “cut the crap” non-layout and his author for his brash comments on incidents in the broad field of IT security and their political implications, especially regarding the industry of mass surveillance. His reach is enormous and will surely beat that of many established newspapers (according to this 2011 post, anyway; newer numbers would be helpful, though).

We don’t know if pressing F5 at least three times a day in the browser tab reserved for “Fefes Blog” makes us part of the fan base. Either way, we couldn’t abstain from sneaking a peek behind the curtain and wanted to analyse some traits of Fefe’s characteristic style and tone. They are, in fact, so characteristic that he was even mimicked by the ultimate German satire magazine “Titanic”, which sure counts for something.

Before we get to it: There have been other attempts to analyse Fefe’s language, just take the inspiring 2012 blog post “Darüber lacht Fefe” by Joachim Scharloth. What we’re going to do is, in fact, something very basic, we’re primarily looking for word (and URL) frequencies and n-grams. For that purpose, we wrote a little Python script which does the following three things:

download the whole blog (easy on bandwidth, because, one: “Fefes Blog” really features the lightest HTML code imaginable, and two: time.sleep(random.randint(1, 10)) while scraping)
parse HTML files: extract date, post identifier and the actual post and put them into a single 3-column TSV file (this is really some of the cleanest data you can work with, what a feast!)
start analysis

After step two, the TSV file contained all blog entries from the beginning (March, 2005) to Mid-May, 2016, which is the cut-off time for our dataset.

Main Sources

These are the most frequent top-level domains linked to in “Fefes Blog”:

Rank	TLD	Frequency
1	spiegel.de	4042
2	fefe.de	3316
3	heise.de	2673
4	wikipedia.org	1319
5	bbc.co.uk	1102
6	tagesschau.de	1096
7	guardian.co.uk	1043
8	youtube.com	1010
9	nytimes.com	692
10	twitter.com	624
11	tagesspiegel.de	531
12	yahoo.com	489
13	rian.ru	485
14	zeit.de	475
15	faz.net	468
16	reuters.com	443
17	taz.de	408
18	sueddeutsche.de	400
19	cnn.com	376
20	washingtonpost.com	364

So this list shows Fefe’s main sources. Now, it’s part of his irony to urge his readers – in the subheader of the blog – to send him “fancy conspiracy links”. This, at least for some, leaves room for irritation and Fefe kind of clarifies this in the FAQ: “Why do you write ‘conspiracy links’ if your content is all normal news?” Reply: “Yes.”

Countries Mentioned

This list is a bit half-baked, since we’re only counting exact hits and only added synonyms for two countries (GB, USA). Wales, Scotland and England were, too, mapped to Great Britain. All a bit hasty, but anyway:

Rank	Country	Frequency
1	USA	1918
2	Deutschland	1506
3	Israel	997
4	Iran	927
5	Großbritannien	708
6	China	657
7	Irak	474
8	Afghanistan	354
9	Griechenland	298
10	Ukraine	294
11	Frankreich	287
12	Türkei	279
13	Japan	253
14	Syrien	231
15	Schweiz	222
16	Österreich	209
17	Polen	177
18	Pakistan	172
19	Italien	167
20	Schweden	158

We matched our dataset against the list of countries provided by datendieter.de. (Btw, it’s very probable that Daten-Dieter is a close friend of MS-DOS-Manfred, BIOS-Bernhard, Hardware-Hanspeter and Lötkolben-Ludwig, but that’s a whole different story.)

Countries Mentioned Over Time

We can also add time as a component, so we let Gnuplot draw a line chart showing Fefe’s interest in the 12 most-mentioned countries over time:

This chart we can actually try to read out loud a bit: The Snowden year, 2013, marks the beginning of an increasing coverage of the USA and Great Britain. In 2014, Russia and Ukraine are peaking, for obvious reasons. One more interesting thing is the slow but steady decline of interest in Israel and Iran since 2005. Greece is starting to be covered in late 2009 with the beginning of the Greek government-debt crisis. So as can be expected from the list of his main sources seen above, “Fefes Blog” more or less mirrors mainstream media coverage.

Acronyms

Well, acronyms, or words written in capitel letters:

Rank	Acronym/Word	Frequency
1	USA	2332
2	CDU	1213
3	US	927
4	EU	926
5	NSA	800
6	SPD	723
7	WTF	556
8	BND	541
9	FDP	486
10	BKA	484
11	CIA	467
12	CCC	393
13	DAS	366
14	OK	341
15	FBI	313
16	DIE	277
17	BESTEN	253
18	IV	250
19	CSU	232
20	NIE	217

3-Grams

Let’s now look at some n-grams (we used AntConc for this). Attention, the lists were curated by us and only show the frequencies for self-contained phrases, which in our opinion are contributing to the typical Fefe sound.

Rank	Frequency	Phrase
1	863	in den usa
3	387	lacher des tages
28	209	das ehemalige nachrichtenmagazin
49	173	bug des tages
56	166	die amis haben

Runner-up among the 3-grams is “das ist ja”, and then we have “das ist ein” and “was für eine” ranking 4th and 5th. Sure, all these high-frequent 3-grams add to Fefe’s stylometric fingerprint, but since we’re not (yet) doing stylometry here, we decided to curate the n-gram lists a bit to not overthrow you with endless lists of boring syntagmas.

4-Grams

Rank	Frequency	Phrase
3	214	kommt ihr nie drauf
10	115	stellt sich raus dass
13	109	oh und wo wir
38	75	kennt ihr den schon
39	75	was für eine farce
49	68	wie arsch auf eimer
61	61	einmal mit profis arbeiten
67	60	wir werden alle störben
69	59	wer hätte das gedacht

5-Grams

Rank	Frequency	Phrase
21	62	habt ihr das auch gehört
42	45	bei uns ist kernkraft sicher
50	40	bei uns ist atomkraft sicher
58	37	was kann da schon passieren
81	30	was kann da schon schiefgehen

6-Grams

Rank	Frequency	Phrase
2	89	oh und wo wir gerade bei
7	65	die polizei dein freund und helfer
10	60	na dann ist ja alles gut
13	52	da weiß man was man hat
17	41	update mir mailt gerade jemand dass
25	35	das geht ja mal gar nicht

7-Grams

The 7-grams are simply a class of their own and prove to be real Fefe earworms, don’t they?

Rank	Frequency	Phrase
1	85	also damit konnte ja wohl niemand rechnen
2	76	die besten der besten der besten sir
3	63	das wird euch jetzt sicher genau so
4	52	kann man sich gar nicht ausdenken sowas
5	48	aus der beliebten kategorie bei uns ist
6	38	man sich mal auf der zunge zergehen
12	33	aus der beliebten reihe bei uns ist
13	33	aus der beliebten serie bei uns ist
15	31	wo kämen wir da auch hin wenn
24	25	beste demokratie die man für geld kaufen
30	24	da fühlt man sich doch gleich viel
33	21	gar nicht so viel fressen wie man

Fun Part I: Given Names of German IT Guys

When Fefe posts a link, hint or story that someone sent him, he gives due credit, the pattern being “(Danke, [name].)” So we also looked into 2-grams starting with “danke”, and while the result was somehow predictable, it is still quite funny, isn’t it? It not only sheds a light on who Fefe’s fiercest audience is, but also shows the opulence of names given to German boys in the 1970ies and 1980ies.

Rank	Frequency	2-Gram/Name
1	237	danke mathias
2	220	danke frank
3	183	danke christian
4	176	danke thomas
5	164	danke stefan
6	135	danke andreas
7	135	danke michael
8	106	danke martin
9	106	danke peter
10	100	danke daniel
11	100	danke klaus
12	95	danke matthias
13	94	danke florian
14	94	danke jan
15	91	danke rop
16	84	danke jens
17	73	danke timo
18	73	danke tobias
19	65	danke sebastian
20	64	danke markus
21	59	danke alexander
22	55	danke johannes
23	55	danke jörg
24	54	danke kris
25	49	danke ralf
26	45	danke lutz
27	45	danke sven
28	43	danke julian
29	42	danke christoph
30	39	danke philipp
31	39	danke stephan
32	37	danke gerry
33	37	danke hans
34	36	danke simon
35	35	danke bernd

Somebody suggested to put the names in a word cloud, but we hate word clouds. :)

If you happen to know some of Fefe’s friends and colleagues, you can surely guess who is behind some of the credits pinned to a name (disclosure: we, the authors of this article, have been thanked at least twice, too, if we recall correctly).

Fun Part II: Using Markov Chains to Generate Fefe Texts

And now, this: Let’s generate some pseudo random Fefe text by using Markov chains. For this experiment we quickly forked dellis23’s markov.py script and adjusted it, the chain size we worked with is 3 and gives us results like this:

Ziegen fickt”, erklärte drastisch US-Präsident Harry S. Truman einem nachdenklichen Kongressabgeordneten. “Wenn er aus und das Ende des Tages: Die Belegschaft sagt, dass die Behörden so 2009 herum angefangen, missliebige Mitbürgern nach allen Regeln der Kunst auseinandernehmen. Ich weiß, welches T-Shirt ich ab jetzt haben, wo sie den Hartz IV kürzen will, stürmen noch schnell und lautlos miterledigen. Keine weiteren Fragen. Die Mühlen der Full-Disclosure-Fraktion bei Sicherheitslücken: Der Jeep-Hack neulich wurde noch der Presse. Erstens: Nachtsicht- und GPS-Geräte kaufen -> 3 Jahre Haft vorgeschlagen. (Danke, Marcel) Die Russen sind schlauer als der losredete, war das in der Türkei.

This is far from being anything worthwhile and the mere result of toying around with this juicy corpus. But alright, let’s generate another one:

Mit Terrorismus hat das mit Cyberangriffen deutlich anders aus, Stichwort Lawful Interception. Wieder was gelernt, diesmal über amerikanische Studenten: Ich habe ehrlich gesagt nicht so aus, dass der vor Gericht erstreiten muss, kann ich nur mit Adblocker nicht zu uns, denn in westlichen Demokratien wie der öffentliche GNU-CVS-Server, furchtbar überlastet ist, und daher muss auch sein Lebensunterhalt direkt davon abhängt, dass er jetzt alles seine Ordnung” um. Denn die Labels daran zugrunde gehen werden. Francesco-Parisi-Universität Styrum — endlich mal was tun diesmal! Sonst lassen die Regierung mit den Nazis damals funktionieren. Zweitens: Wir merken es oft nicht gelöscht würden.

The code is quite simple and can sure be optimised in many ways, but for today the Fefe Research Institute pulls the plug.

Distant Reading with Foucault?

2016-05-06T00:00:00+00:00

Half a year ago, November 14, 2015, we participated in a workshop at the Institute of Science and Art in Vienna, entitled “Distant Reading and Discourse Analysis”. Our talk, which went under the headline “Distant Reading with Foucault?”, promised to give thoughts on “the practice of distant reading” and ponder potential “operationalisations of Foucauldian discourse analysis”.

A revised version of the talk was published last week as a featured article in the brilliant foucaultblog, and here comes the caveat, it’s in German: “Fernlesen mit Foucault?” So, the raison d’être of this blog post is to give you a short summary of the article.

What we tried to do in the first place was to define what Distant Reading is, or rather: what it was in the past 15 years since Moretti coined the term in his essay Conjectures on World Literature. We considered five possible answers. Distant Reading, was it …

… “a joke”?
… a polemical term, a buzzword?
… a computer-based method for the analysis of literature?
… Moretti’s attempt towards a ‘canon-critical’ large-scale literary historiography?
… a failed or at least terminated project?

All of these possible answers had their grains of truth in them (and are discussed thoroughly in the original German version of our article). However, what Distant Reading certainly did not provide in the past one and a half decade, was a reliable methodology. Scholars using this term usually thought it enough to reference Moretti. At the same time, one of Moretti’s main epistemic moves seems to have been that of following analogies, as has been suggested by scholars such as Christopher Prendergast (2005) or Katja Mellmann (2009), something Moretti is not unaware of given that he himself, for one, discussed the problematic analogy between World-Systems Theory and a “World-Literary System” (in More Conjectures).

‘Old’ and ‘New’ Distant Reading

When rereading Moretti for this talk we were kind of surprised to find that Distant Reading in its original form has nothing at all to do with the practices of the Digital Humanities. It is hard to find Moretti talking about technological implications, he never mentions standards, protocols, scripts, databases and all the nerve-racking little problems you encounter when trying to squeeze literary data for meaningful findings. There was no freely available corpus or code or documentation that made his theses reproducible.

In Vienna, while strolling up and down Berg-Gasse and climbing the notorious Strudlhofstiege a.k.a. Stiedlhufstroge, we were discussing what our quasi-obituary for Old Distant Reading meant for our own research. Because at the very same time we were preparing a data-driven poster for the annual Digital Humanities conference of the German-speaking countries which took place in March, 2016. Our poster set out to be a “Distant-Reading Showcase: 200 Years of German-Language Drama at a Glance” (you can find it on figshare).

We wrote about the making of the poster in our DLINA project blog ([Digital] Literary Network Analysis), where we also give examples of what you can “distantly read” when looking at this obscure bulk of network graphs. What is important here is that our own project confronted us with the question: What should a renewed Distant Reading be like? While we shouldn’t cling to the term, we can state that Distant Reading (or Macroanalysis, or whatever you call it) should be reproducible, which includes all the above-mentioned aspects: freely available corpus, code, documentation, data, something that can cost months of additional work, but something we regard as essential as the eventual presentation of the results.

And Now, Foucault

The clarification on what Distant Reading meant or means ate up the better part of our talk, but it was not too late to check in with Foucault. So, could there be a methodologically clean ‘Distant Reading with Foucault’?

It is obvious that traditional, semantically rich concepts balk – almost programmatically – at their operationalisation in contexts of quantitative, formalised research. This concerns numerous hermeneutic ideas regarding the ‘deeper’ meaning of a text or text element. However, there is no reason why the Digital Humanities shouldn’t utter ideas for possible operationalisations. For the most part, this will result in a genuinely different perception of a subject, or, to paraphrase Moretti: At the end, these terms will probably have their place within completely different theories.

This becomes clear when looking at “the elementary unit of discourse”, the ‘statement’ (‘énoncé’). Foucault’s definition is perimetric, he repeatedly stresses that a ‘statement’ is something quite different from a string of characters (that could be located in a corpus based on definable rules). Instead, considering his extensive explanations in The Archaeology of Knowledge, a ‘statement’ is, to a large extent, context-relative, its determination, it seems, is less a positivist than a hermeneutical act. ‘Hermeneutical’ insofar as the discourse analyst has to “at least superficially understand the meaning of statements” to be able to classify them as such (quoting Philipp Sarasin 2014, p. 66; our translation).

Therefore, in light of the apparently very low operational potential of Foucault’s discourse analysis, we should ask ourselves whether it wouldn’t perhaps make more sense to go another way. Instead of thinking about how Foucault’s theory design could be operationalised, we could, for starters – following Moretti’s slogan “Forget programs and visions” (PDF) – view and discuss the results provided by well-established techniques of text analysis. In fact, decades ago we saw the emergence of a field of discourse analysis relying on corpus-linguistic/lexicometric approaches, just think of the pioneering work on an Analyse automatique du discours (1969) by Michel Pêcheux.

Also Moretti (in a Pamphlet from his post-Distant-Reading phase, so to speak) simply applied corpus-linguistic methods to describe something like the transformation of economic discourse based on an analysis of the annual reports of the World Bank (PDF). Sure, this is not an examination of ‘statements’ in a Foucauldian sense. Instead, Moretti is looking for most frequent words, collocations or the frequency of certain parts of speech and grammatical constructions, in other words, he and his co-author Dominique Pestre are looking for linguistic, not discourse-analytical ‘units’. But maybe, based on this kind of data, a new (and necessarily different) discourse analysis could be established, a revamped implementation of Foucault’s project, which could – unlike most corpus-linguistic approaches – continue what appears to us the key signature of Foucault’s “work on the discourses”: research as a practice of a critical science. ▣

Concluding Remarks

Given the workshop character of the event, our talk was little more than a first tentative approach to the question raised in the title. Sure enough, the workshop did spur some nice discussions that will certainly be continued. All talks of the workshop will soon be part of a separate ‘issue’ on the foucaultblog, edited by Simon Ganahl and Maurice Erb.

Wikidata Meets World Literature

2016-04-25T00:00:00+00:00

You can currently access and contribute to 281 active Wikipedia language versions. What follows is a top-25 list of ‘books’ ranked by the number of articles dedicated to these ‘books’ in different Wikipedia language editions (results first, explanation thereafter):

Wikidata item	Title	Authorlabel	Linkcount
Q8258	One Thousand and One Nights	anonymous	116
Q15228	The Lord of the Rings	J. R. R. Tolkien	115
Q480	Don Quixote	Miguel de Cervantes	113
Q25338	Le Petit Prince	Antoine de Saint-Exupéry	110
Q40591	The Communist Manifesto	Karl Marx/Friedrich Engels	108
Q459842	Book of Mormon	Joseph Smith	106
Q208460	Nineteen Eighty-Four	George Orwell	102
Q8251	The Art of War	Sun Tzu	95
Q92640	Alice’s Adventures in Wonderland	Lewis Carroll	93
Q8279	Shahnameh	Ferdowsi	93
Q74287	The Hobbit	J. R. R. Tolkien	91
Q43361	Harry Potter and the Philosopher’s Stone	J. K. Rowling	86
Q1396889	Animal Farm	George Orwell	85
Q8265	Dream of the Red Chamber	Cao Xueqin	82
Q41675	Guinness World Records	Craig Glenday	80
Q48244	Mein Kampf	Adolf Hitler	80
Q8269	The Tale of Genji	Murasaki Shikibu	79
Q165318	Crime and Punishment	Fyodor Dostoyevsky	77
Q47209	Harry Potter and the Chamber of Secrets	J. K. Rowling	77
Q47598	Harry Potter and the Prisoner of Azkaban	J. K. Rowling	75
Q46887	Harry Potter and the Half-Blood Prince	J. K. Rowling	75
Q80817	Harry Potter and the Order of the Phoenix	J. K. Rowling	75
Q123397	The Republic	Plato	74
Q170583	Pride and Prejudice	Jane Austen	74
Q147787	Anna Karenina	Leo Tolstoy	73

Sports commentator proclaims: Legendary Arabian story collection first, the Lord of the Rings a close runner-up, Spanish hack Cervantes completing the podium! But let’s curb our enthusiasm for this quirky race for a moment. You cannot be cautious enough when interpreting this kind of lists/rankings. There are a hundred things you’d have to involve when doing this.

Our ranking reveals nothing on the nature or quality of, let’s say, these 110 language versions of “Le Petit Prince”. We could deal with a one-sentence article in 50 Wikipedias for all we know. Understanding the genesis of Wikipedia and measuring its contents is a research field of its own. And Wikidata, although building on Wikipedia content, of course, is yet another thing whose characteristics you’d have to take into account before jumping to any rash conclusions.

Let’s start slowly. What this ranking suggests to show is a list of 25 ‘books’ (we’ll come back to this term later) that are of whatsoever importance across many countries/cultures/language communities. It is not at all a proper basis to redefine the term ‘world literature’ for the digital age. But it does include the idea of ‘world’ as part of ‘world literature’, meaning ‘books’ that transcend national and language borders, also if their literary value may vary considerably. In any case, before starting to interpret anything, you’ll always have to communicate how exactly you generated this kind of results. So …

… as an example of how the Wikidata Query Service can facilitate this type of quantitative analysis, we wrote a SPARQL query that orders ‘instances of’ ‘books’ with an author by ‘site link’. This means that if a work of literature has the statement ‘instance of book’ in Wikidata, and it has a Wikipedia page, it can be ranked by the number of Wikipedias that have a page for it.

This is easily reproducible for you at your own machines, just head over to http://query.wikidata.org and copy & paste this in the upper box, then hit “Execute” (and while the query is being processed, throw a glance at Alan Liu’s criticism of the concentration “on pushing the ‘execute’ button” in the Digital Humanities):

prefix schema: <http://schema.org/>
prefix wd: <http://www.wikidata.org/entity/>
prefix wdt: <http://www.wikidata.org/prop/direct/>

SELECT ?s ?desc ?authorlabel (COUNT(DISTINCT ?sitelink) as ?linkcount)
WHERE {
  ?s wdt:P31 wd:Q571 .
  ?sitelink schema:about ?s .
  ?s wdt:P50 ?author
  OPTIONAL {
    ?s rdfs:label ?desc filter (lang(?desc) = "en").
  }
  OPTIONAL {
    ?author rdfs:label ?authorlabel filter (lang(?authorlabel) = "en").
  }
} GROUP BY ?s ?desc ?authorlabel ORDER BY DESC(?linkcount)

Did it work for you?

By the way, changing the object criteria from ‘book’ to ‘literary work’ (wd:Q7725634) returns a different result, that includes books of the Bible (that are not classified as ‘instance of book’). Running it without the author, also yields a slightly different result, as there are a few works without an author that are very widely translated (like, for example, today’s winner, “One Thousand and One Nights”).

The query that yielded the above ranking was run today (April 25, 2016) at 08:25 CEST. If you save the results of your query (like we did, in CSV format, which can be done directly in Wikidata’s Query Service) and compare it to the same query conducted at a later date, you would have a useful rudiment for the analysis of how certain works gain or lose importance in the Wikipedian universe. Something we might come back to later. Until then, happy SPARQL’ing!

This is weltliteratur.net

2016-04-23T00:00:00+00:00

“Welcome! Welcome! Welcome!”
– John Oliver

Now this is, believe it or not, the sky over Göttingen #OnThisDay, April 23, 2016, after just some minimal photoshopping (turning light grey into summery light blue):

Göttingen (and Hanover) were the places where we first started to think about a blog called weltliteratur.net, a neat and plushy think tank where we could publish some of our musings on Digital Humanities-related things that might or might not be or become part of a bigger research project.

Now that some of us are scattered to the four winds (to Potsdam, Sheffield, Moscow, etc.) we’re planning to use this site as a place to continue discussing and exploring ideas emanating from our daily DH’ing. Let’s see how it goes.

Today is World Book Day, thanks to Shakespeare and Cervantes, who both died exactly 400 years ago, well, Shakespeare some days later, thanks to the Julian calendar. Either way, a suitable day to start off with something called weltliteratur.net, which also alludes to Goethe’s notion that “the epoch of world literature is at hand”, as uttered by him in 1827.

Our first non-meta article will be delivered on Monday, something about Wikidata meeting World Literature, a small blogpost on how you can try to grasp what this fuzzy corpus called world literature is, according to Wikipedia/Wikidata.

Oh, one other USP of weltliteratur.net is the very unofficial DH Soundtrack, a growing list of current songs we listen to before, after and, of course: while conducting our research. So if some of our entries seem to make no sense to you, try switching on our official musikbett to ponder a supposed deeper meaning.

This, by and large, is it.

On behalf of everybody else,
Frank – Mathias – Robert

weltliteratur.net

DraCor Platform Update (July 2025)

💡 A Smarter, Cleaner API

🖥️ A Refined Frontend for Exploration

🧾 TEI Schema 1.0: A Solid Foundation

📜 Comprehensive Documentation

📅 Save the Date: DraCor Summit 2025

🧭 Follow the Drama

How to Identify Key Passages in Literary Works? Using Algorithms and Machine Learning!

Context

Corpus

Automatic Identification of Quotations

QuidEx – Visualization and Exploration

Summary and Outlook

Streamlining the DraCor API

Again, What Is DraCor?

DraCor API 1.0

Legacy API

pydracor and rdracor

Other Updates

The Team Behind the Update …

Reverse Engineering "Frembdsch", A Fantasy-Language Radio Play by Dagmara Kraus and Marc Matter

Vossian Antonomasia Extraction Using Pre-Trained Language Models

Automatic Extraction of Network Data From Amazon Prime Videos (Using "1917" As an Example)

Conclusion

Quantification of Scholarly Articles on German Drama

Context

Data Sources

Implementation

Results per Play

Results per Author

DraCor Summer Update (August 2021)

How Did We Get Here?

API

Ways of Interaction

Ted Underwood's "Distant Horizons": Reviews and Replications

REVIEWS

REPLICATIONS OF CODE AND/OR DATA

The First of May in German Literature

Vossian Antonomasia, Next Level

Interactive Timeline for Data Exploration

Automatic Detection of Vossian Antonomasia

Modifiers

JSON Data Dump

Future Endeavours

Tools Mentioned in DH2020 Abstracts

spaCy (5)

word2vec (5)

Gephi (4)

Natural Language Toolkit (4)

Voyant Tools (4)

Drupal (3)

Google Maps (3)

Tesseract (3)

ArcGIS (2)

Cytoscape (2)

MALLET (2)

Neo4j (2)

Omeka (2)

OpenStreetMap (2)

Some Acknowledgements and Observations

Network Modelling "The Last Days of Mankind"

Measuring the Impact of DH Conference Abstracts: The Case of DH2016

1. Most cited abstracts from DH2016 (as of 15 February 2020)

2. Citation counts for follow-up articles published in the DSH special conference issue

3. Articles published elsewhere

Some Thoughts

DH Tools Mentioned in "The Programming Historian"

jupyter-notebooks (11)

building-static-sites-with-jekyll-github-pages (9)

mapping-with-python-leaflet (9)

using-javascript-to-create-maps (9)

corpus-analysis-with-antconc (8)

creating-network-diagrams-from-historical-sources (8)

geoparsing-text-with-edinburgh (7)

text-mining-with-extracted-features (7)

beginners-guide-to-twitter-data (6)

creating-apis-with-python-and-flask (6)

introduction-to-ffmpeg (6)

sonification (6)