Towards Computational Comparative
Literary Studies

Adressing the Challenges of Multilingualism

Christof Schöch
(Trier University, Germany)

KEASTWEST Conference 2024
Dongguk University, Seoul, South Korea

25 May 2024

Introduction

• Before starting, I would like to thank the organizers for the invitation to speak here.
• It is a privilege and a special experience for me to be here.
  

Multilingualism and me

Trier University

Trier Center for Digital Humanities

Alliance of Digital Humanities Organizations

• Multilingualism is a topic that is really close to my heart
• And it is highly relevant in my current key roles
  • At Trier University, I use computational methods to investigate French literature, in a German context, mostly publishing in English; and Trier is at the borders to France, Luxemburg and Belgium, so in a multilingual setting.
  • At the TCDH, we have a broad portfolio of projects on materials in German, English, French and other languages, whether in the original or in translations
  • In the context of ADHO as an increasingly global organization, multilingualism plays a role in almost everything that we do.
• For this reason, I am very happy to be able to speak about multilingualism in Digital Humanities;
• More specifically, about the challenges of multilingualism when attempting to practice what could be called Digital Comparative Literature, or more precisely perhaps: “Computational Comparative Literary Studies”.
  

Computational Comparative Literary Studies?

• What could CCLS be?
  • Literary Studies
  • but: Computational (using digital data and methods)
  • and: Comparative (transnational, transmedial)
• Many challenges for conversion
  • requires multiple areas of expertise
  • significant challenges of multilingualism

• Comparative Literature is well-established, of course, and I am not here to tell you what it is;
• Computational Literary Studies are emerging (and have a tradition going back to the 1950s);
• but the intersection of these two strands of research has so far been quite limited. So the question is: what is holding us back in this area, and what can we do to address these hindrances?
• I think there are multiple factors at play:
  • mutual lack of knowledge about the other field, for sure
  • also the fact that for CCLS, one needs to acquire a very wide range of expertise
  • but one is certainly the fact that Comparative Literature deals typically with materials in multiple languages and emerging from multiple cultural and historical contexts;
  • and that Computational Literary Studies relies on data and tools that are very much biased towards single-language, English-focused applications.
• That is why I believe that the challenges of multilingualism need to be at the heart of any attempt to bring Computation and Comparison together in Literary Studies.

Three attempts at CCLS

1. Corpus Building: The Diversity Paradox
2. Data Modeling: Linked Open Data
3. Text Analysis: Multilingual Stylometry

• For this reason, I would like to discuss some of the challenges with multilingual computational literary studies and try to identify strategies to overcome them.
• In order to do this, I would like to speak about three areas where multilingualism provides challenges, and illustrate each based on a very concrete example from my own research.
• These three areas are:
  • corpus building
  • data modeling
  • text analysis

(1) Corpus Building:
The Diversity Paradox

• The first case study in multilingualism concerns corpus building
• Specifically, it concerns the process of building a multilingual collection of corpora, the “European Literary Text Collection”.

See: distant-reading.net

The COST Action ‘Distant Reading for European Literary History’

• We did this in the framework of the COST Action “Distant Reading for European Literary History”.
• This was a networking project funded by the European Union from 2017 to 2022.
• More than 200 researchers from more than 30 countries were involved
• One of the key results of the project was the publication of ELTeC:
  The European Literary Text Collection

The ‘European Literary Text Collection’ (ELTeC)

• The goal of building ELTeC was to provide a benchmark corpus that would allow computational investigations into the history of the novel across Europe
• And to do this using current best practices: openly available, encoded in XML-TEI, linguistically-annotated, and with rich and relevant metadata.
• The two tables shown here describe the various corpora included in ELTeC
• There is a lot of information here, so I’ll break it down a little
• First of all, there are 21 different languages represented in the collection (language codes on the left)
• For 12 of them, we have reached our goal of including 100 novels published between 1840 and 1920

A closer look: corpus composition in ELTeC

          English ELTeC corpus

          Romanian ELTeC corpus

• What you see here is a visualization of the corpus composition for two corpora, the English one and the Romanian one.
• The visualizations summarize the distribution of three key characteristics of each text:
  • its period of publication
  • its length (in three simple categories)
  • and the gender of its author
• Even without going into any details, it is obvious that the two corpora are very different:
  • the English corpus is very-well balanced, with respect to these criteria.
  • the Romanian corpus much less so.
• What is less visible, but also important:
  • The criteria themselves were source of challenges and a matter of debate
  • For example, the text length criteria: huge differences of what counts as a “short” or a “long” novel, depending on literary traditions in each language / nation

The Diversity Paradox

• ELTeC design goals: enable meaningful cross-language investigations
  • Balance with respect to key text characteristics
    (text length, author gender, prestige)
  • Inclusivity with respect to language-based literary traditions
• Consequence: the ‘diversity paradox’
  • If the criteria are too loose, balance is compromimsed
    (many, but invalid, corpora)
  • If the criteria are too strict, inclusivity is compromised
    (valid, but few, corpora)
  • In both cases, meaningful cross-language investigations are impossible

• Explain the key ideas of the slide first
• So what is the way out of the diversity paradox?
  • First: a bit of flexibility, using ranges of target values rather than strict quotas.
  • Second: Support less-resourced corpus builing teams (Romanian!)

Reference: Schöch et al. (2021).

(2) Data Modeling:
Linked Open Data

• The second case study is one where there is actually a brilliant solution for multilingual data and multilingual audiences alread:
• For data modeling in a multilingual context, linked open data is extremely useful.

See: mimotext.uni-trier.de/en

The project ‘Mining and Modeling Text’

• Mining and Modeling was a project funded by the regional government from 2019 to 2023
• This was an interdisciplinary project involving computer science, literary studies, legal studies, computational linguistics, and digital humanities
• The key goal of this project was to use computational methods to build a knowledge base about the French novel of the Enlightenment
• But more generally, also, to experiment with the idea of “Linked Open Data for Literary History”

More information: Schöch et al. (2022)

Linked Open Data: Simple Statements

• How does “Linked Open Data” actually work?
• There are subjects, predicates and objects
• Subjects are entities such as authors or works
• Objects are other entities, but also topics, locations, etc.
• Predicates link subject and object, like a verb
• This results in very simple statements
• The potential of LOD only unfolds with a large number of simple statements
• This seemingly simple structure is very powerful, because it is so generic
• To illustrate this power, one could liken it to chess: simple rules with lots of possibilities
• All the information we identify, using text and data mining, from our texts, is expressed in this way.
• A knowledge graph is created, which becomes part of the so-called “semantic web”

Source: Wikidata and datajournalism.com/read/longreads/the-promise-of-wikidata.

Linked Open Data: Multilingualism

• In LOD, all information is modeled semantically
• That means, each “unit of meaning” as an identifier
• And it has a lable and a description
• More precisely: many lanels and descriptions, in many languages
• So LOD and Wikidata are inherently multilingual
• That’s a big asset in a global scholarly world

MiMoText Base: Query for themes in novels

• Every piece of information we capture is identified using authority data
• When using Wikidata, this means labels and definitions are available in multiple languages
• For our own taxonomies, we have focused on French, German and English
• When performing queries, we can mobilise this multilingual potential
• The above is a query for the dominant themes in the 200 novels of our corpus;
• We can easily switch between languages for labeling the graph

Live SPARQL query: query.mimotext.uni-trier.de

(3) Text Analysis:
Multilingual Stylometry

• The third case study concerns computational methods of text analysis
• One of the oldest paradigms in this area is stylometric authorship attribution
• Essentially, the idea that the author of an anonymous text can be identified based on a comparison of the word frequency profiles of the disputed text and texts by multiple candidate authors.
• The work I present here was done in the framework of the project “Computational Literary Studies Infrastructure”
• This project is funded by Horizon 2020 from 2021 to 2025.

See: showcases.clsinfra.io

High-profile cases of stylometric authorship attribution

William Shakespeare:
Craig and Kinney (2009)

Molière and Corneille:
Cafiero and Camps (2019)

Elena Ferrante:
Tuzzi and Cortelazzo (2018)

Galbraith / Rowling:
Juola (2015)

• Some high-profile cases of authorship attribution in recent times
• Of course, the long-standing investigations of Shakespeare
• In France, there has been a “Molière-Corneille”-controversy, almost a “feud”
• Recent high-profile cases have involved best-selling pseudonomous author Elena Ferrante
• And the detective novels written by Joanne K. Rowling, the Harry Potter author, under a pseudonym.
• These are almost always single-language authorship questions.

Multilingual stylometry?

translation ↦ ↧ original	fra	eng	hun	ukr
fra	fra-fra	fra-eng	fra-hun	fra-ukr
eng	eng-fra	eng-eng	eng-hun	eng-ukr
hun	hun-fra	hun-eng	hun-hun	hun-ukr
ukr	ukr-fra	ukr-eng	ukr-hun	ukr-ukr

• Using corpora from the European Literary Text Collection (ELTeC)
• Translated entirely into the other languages using DeepL

• In our project, the question has been addressed from a multilingual perspective
• That is, we want to know: How does language infulence stylometric performance?
• But also: how does corpus composition influence it?
• To answer such questions, we need to decouple language from corpus composition
• Work with corpora in four languages: eng, fra, hun, ukr
• Using original versions and translations into all three other languages (using DeepL machine translation)
• We’ve build an interactive showcase to allow us and others to investigate this dataset.

Some first results

More information: Dudar et al. (in progress).

• There are strong differences between the languages, which in part can be explained by language
• Performance is always better in the original language (the diagonal from top left to bottom right)
• Performance drops for the translated corpora (to some extent, because the translator is always the same)
• There is a correlation between corpus composition and stylometric performance: we have developed a measure of corpus

Full interactive showcase

showcases.clsinfra.io/stylometry

Conclusion

Take-home message

• Good, multilingual corpora are rare and hard to build
• Linked Open Data is a huge opportunity for multilingual data modeling
• Text analysis is still primarily multi-lingual rather than cross-lingual
  (but multilingual LLMs are in the process of changing that)

Lessons learned

• Multilingual research is multicultural research
• ‘Computational Comparative Literary Studies’ requires multiple competencies
• Nobody can learn everything: we need interdisciplinary collaboration
• Let’s learn from each other: Computational and Comparative Literary Studies

Thank you for your kind attention!

References

Cafiero, Florian, and Jean-Baptiste Camps. 2019. “Why Molière Most Likely Did Write His Plays.” Science Advances 5 (11): eaax5489. https://doi.org/10.1126/sciadv.aax5489.

Craig, Hugh, and Arthur F. Kinney. 2009. Shakespeare, Computers, and the Mystery of Authorship. Cambridge University Press.

Dudar, Julia, Evgeniia Fileva, Artjoms Šeļa, and Christof Schöch. in progress. “Multilingual Stylometry: The Influence of Corpus Composition and Language on the Performance of Authorship Attribution Using Corpora from the European Literary Text Collection (ELTeC).” Tbc, in progress.

Juola, Patrick. 2015. “The Rowling Case: A Proposed Standard Protocol for Authorship Attribution.” Digital Scholarship in the Humanities 30 (suppl. 1): 100–113. https://doi.org/10.1093/llc/fqv040.

Schöch, Christof, Maria Hinzmann, Julia Röttgermann, Katharina Dietz, and Anne Klee. 2022. “Smart Modelling for Literary History.” International Journal of Humanities and Arts Computing 16 (1): 78–93. https://doi.org/10.3366/ijhac.2022.0278.

Schöch, Christof, Roxana Patras, Tomaž Erjavec, and Diana Santos. 2021. “Creating the European Literary Text Collection (ELTeC): Challenges and Perspectives.” Modern Languages Open, no. 1: 25. https://doi.org/10.3828/mlo.v0i0.364.

Tuzzi, Arjuna, and Michele A. Cortelazzo, eds. 2018. Drawing Elena Ferrante’s Profile: Workshop Proceedings. Padova: Padova UP.