GLOBAL DIGITAL HUMANITIES

TEXT-MINING MULTILINGUAL LITERARY CORPORA

Computational Criticism has a language problem

• Despite the ability of text mining techniques to analyze hundreds of thousands of texts, there still exists no reliable way to compare statistics between languages

​

• From a literary perspective, this means that phenomena that take place across nations, or between languages, are never fully explored through computation

​

• For example, the Enlightenment, Modernism, and the Renaissance are all aesthetic, political, and (for our purposes) literary movements that include objects in a number of (European) languages

​

• Current no digital humanities project is capable of exploring any phenomena in a global context: world literature remains impossible

Anglo-Centrism

• Most corpora of texts are in English: this means that most computational critical work is also done in English creating a global disparity of both resources and methods

​

• Because most corpora are in English, most methods are developed primarily for English language texts, creating difficulties with NLP models and particularly non-English character sets.

​

• Roopika Risam (2016) argues that non-English Digital Humanities (particularly the Global South) suffers as a result of this deficit

​

• For example, even the French Bibliotheque Nationale corpus numbers 120,000 volumes; the Literary Lab collective English corpora alone numbers over 540,000 values

Anglo-Centrism: Hathi Trust

Global Digital Humanities: Aligning Results

• This project seeks to find ways of aligning quantitative results across multi-lingual corpora

​

• While individual words are irreducible in their given languages, the summary statistics generated by quantitative literary criticism can be made comparable

​

• But we lack methods for aligning these results and making them compatible across corpora

Alignment is NOT translation

• The act of translation does violence to the text by substituting words and concepts that may not be identical, or even meaningfully similar

​

• Rybicki (2012) shows how translator signals in single author corpora can derail a stylometric analysis for authorship

​

• Despite the improvements in computational translation, it is impossible to automate even a base level translation such that meaning, let alone the textual nuances, are retained

Alignment is NOT translation

• The translation makes the text brutal, replacing words and phrases that are not necessarily identical or similar in meaning

​

• Rybicki (2012) explains how interpreter signals in one author's authors can hamper the analysis of the standard measurement of composition

​

• Despite improvements in the translation of calculations, it is not even possible to automate the basic translation, so the value, including text colors, is retained.

Global Collaborations

Over the past several years, the Literary Lab has joined with other Digital Humanities organizations to explore multi-lingual projects:

​

• The unexpected turn: narrative analysis of early 20^th century short stories in newspapers and journals in Greek and English with Anastasia Natsina (University of Crete, Greece)

​

• Sententious Sentences in English and German with Fotis Jannidis (Universität Würzburg, Germany)

​

• Literature/Littérature with Alexandre Gefen and Marianne Reboul (CNRS, France)

Multilingual Alignment: 3 Experiments

1. Network Analysis

​

• Topic Models

​

• Word Embeddings / Vector Models

Dramatic Structure

1: Networks

• Networks abstract structure from the language of the text

​

• Dramatic networks connect characters (nodes) by how many lines of dialogue they exchange (edges)

​

• These visualizations retain coherence across languages

​

• We can also extract network statistics to compare across models

Dramatic Structure

Centrality: Betweenness

Centrality: Betweenness

Amphitryon, Dryden (1690) – Betweenness Centrality

Amphitryon, Hawkesworth (1756) – Betweenness Centrality

Amphitryon, Kleist (1803) – Betweenness Centrality

Network Density of Amphitryon (Three Editions)

Topicity Between Languages

2: Topic Models

TWO TOPICS FROM A TOPIC MODEL OF �~200 WORKS OF SUSPENSE FICTION

Topic Model: The Ambassadors

Epistemological words

​

question

learning

understanding

secret

prove

Topic Model: The Ambassadors

Space/time words

​

reached

hotel

friend

evening

room

Topic Model: The Ambassadors

Topicity: Mono-Topical Paragraph

Topicity: Bi-Topical Paragraph

Topicity in Dickens vs Goethe

Topicity in Dickens vs Goethe: scaled

Topicity in Dickens vs Goethe

Dickens Mean Topicity: 2.55 (Scaled); 2.02 (Unscaled)

​

Goethe Mean Topicity: 1.08 (Scaled); 1.04 (Unscaled)

Confounding Effects:

​

1. Language baseline: does English have a higher topic density than German at all scales?
2. Period: Dickens is writing different kinds of novels than Goethe

​

A set of protocols or methods for aligning these results would allow us to interpret the difference.

History of Literature / Histoire de Littérature

3: Vector Models

CURRENT CORPUS (FRENCH)

• 1840 titles of literary critic (417 perfectly digitized, the rest not so well)
• 758 authors
• Between 1707 and 1949
• Texts chosen from a larger database of literary critics currently digitized
• All metadata verified and curated (which means many difficult decisions …)
• 419 978 033 tokens
• 29 000+ occurrences of “littérature”, more than 58000 if we include typo and word derivatives (e g “littéraires”) ;
• 0,0063% global frequency of “littérature” in our corpus compared to 0.0035% in Google books (for 1800-1900) : two times more focused

​

CURRENT CORPUS (ENGLISH)

• Based on ProQuest database “Historical Literary Criticism”
• 500 authors (as subject); 20,129 texts
• Between 1503 and 1992
• However: dating is a problem for some works as some dates correspond to the edition, rather than the original text – 7819 (39%) have an unreliable date
• 46 504 576 tokens
• 19 896 occurrences of “literature”, more than 44000 if we include typo and word derivatives (e g “literary”) ;

​

Corpus Statistics

Corpus Statistics

Corpus Statistics

CORPORA KNOWN BIASES

• Mainly XIXth century for practical reasons
• Some critical/pedagogical editions of older work and many extracts (French)
• Critical excerpts from texts (English) / Only books and no press (French)
• English corpus is soley author focused: no texts on belles lettres in general
• Too many books on “critique dramatique”
• Discrete data : means no linear curves but scatter plot and tendencies curves

​

Global Vectors for Word Representation: GloVe

(Penning, Socher and Manning, 2014)

​

Log bi-linear model with a weighted least-squares objective

​

Less computationally expensive than the neural nets associated with word2vec but similar results at high token counts (n>5M)

​

Allows for distance scores via cosine similarity as well as vector math on results

Vector Models

All vectors calculated to 150 dimensions using 5 token skip-grams

​

Vector model uses shared contextual similarity to assign distance:

​

Two words ”close” to each other may never appear in the same sentence.

​

Synonym/antonym relationships are captured by the model

​

Vector Models

Terms closest to “literature/littérature”

Terms closest to “literature/littérature – nationality/nationalité”

Terms closest to “literature”

Terms closest to “literature” (detail)

Terms closest to “littérature”

Terms closest to “littérature” (detail)

Nationality as a function of history

Nationalité as a function of history

Art vs Science (English)

Art vs Science (French)

Arts vs Sciences (French)

Art-Science vector model

Arts-Sciences vector model

Philsophie vs Philosophy

Decreasing terms of function (English)

Increasing literary/criticism terms (English)

Decreasing disciplinary terms (French)

Increasing professional terms (French)

Romantic-Classical vector model

Romantique-Classique vector model