Tracing Antisemitic Language Through Diachronic Embedding Projections: France 1789-1914

Rocco Tripodi, Massimo Warglien, Simon Levis Sullam, Deborah PaciACL 2019

TL;DR

This paper uses diachronic word embeddings to analyze how antisemitic language evolved in France between the French Revolution and World War I. By tracking semantic shifts in words related to Jews and projecting them onto defined antisemitic discourse streams, the study reveals a significant and growing trend of antisemitism, particularly in the late 19th century, culminating around the Dreyfus Affair.

What Problem Does This Paper Address?

While word embeddings are powerful tools for capturing word meanings and have been shown to inherit human biases, this paper shifts the focus from debiasing these embeddings to using their inherent biases for historical analysis. Specifically, the authors address the challenge of quantitatively tracing the evolution of complex societal biases, like antisemitism, over time. Historians have identified distinct periods and streams of antisemitic discourse in France, but a computational method to track these shifts and their semantic impact has been lacking. This paper aims to bridge that gap by leveraging diachronic word embeddings to provide a data-driven perspective on the historical development of antisemitic language.

Methods

The core methodology involves training word embeddings on a large corpus of French texts spanning 1789-1914. The key techniques employed are:

  • Diachronic Word Embeddings: Instead of training a single model on the entire corpus, the data is divided into chronological "bins." Word embeddings (specifically, word2vec skip-gram models) are trained independently for each bin. This allows for the analysis of how word meanings and their semantic relationships change over time.
  • Local Neighborhood Measure: To quantify semantic change for specific words (like "juif" and its variants), the authors use a measure that compares the "neighborhood" of a word in one time bin to its neighborhood in the next. A higher cosine distance between these neighborhoods indicates a greater semantic shift.
  • Embedding Projections onto Semantic Axes: To measure biases, the authors define six "semantic axes" corresponding to distinct streams of antisemitic discourse identified by historians (religious, economic, socio-political, racial, conspiratorial, and ethic). These axes are constructed using pairs of antonyms related to each stream. Word vectors are then projected onto these axes. A positive projection indicates an adverse bias towards Jews within that semantic stream, while a negative projection indicates a favorable bias.

Datasets

  • Corpus: The primary dataset is a custom corpus of French books and periodicals published between 1789 and 1914, sourced from Gallica (the digital library of the Bibliothèque Nationale de France). The corpus was constructed by selecting all documents containing keywords related to Jews. It comprises 54,403 books and 245,188 periodical issues.
  • Word Embeddings: Word2vec skip-gram models were trained on this corpus, divided into 26 time bins of approximately 450 million tokens each.
  • Antonym Pairs: A public resource (http://www.synonyms-fr.com) was used to collect antonym pairs for defining the six semantic bias axes.

Key Insights and Ideas

  • Bias as a Historical Lens: The paper demonstrates that the biases embedded in language models can be a valuable tool for historical research, offering a quantitative way to track evolving societal attitudes.
  • Semantic Shifts Mirror Historical Events: The study highlights how specific historical events, such as the rise of modern antisemitism, the Dreyfus Affair, and publications like Drumont's La France juive, correlate with significant shifts in the semantic space of words related to Jews.
  • Distinct Streams of Antisemitism: The analysis confirms that antisemitism manifested across multiple semantic domains (religious, economic, etc.), with varying intensity and temporal patterns.
  • Singular vs. Plural Bias: The paper observes that the singular "juif" is more strongly associated with conspiratorial and socio-political antisemitism, suggesting a focus on individual Jewish actions or a generalized "enemy" concept. Conversely, the plural "juifs" is more linked to economic and ethic streams, implying collective behaviors or stereotypes.
  • Religious Bias as a Foundation: The religious semantic area consistently shows the highest adverse bias, suggesting that traditional religious prejudices formed a bedrock for later forms of antisemitism.

Key Results

  • Growing Adverse Bias: The analysis reveals a steady increase in adverse bias associated with words referring to Jews, particularly starting in the mid-1880s, intensifying around the Dreyfus Affair, and remaining high thereafter.
  • Peak Periods: Peaks in adverse bias were observed around the establishment of Napoleon III's Second Empire (1855-1866) and during the Dreyfus Affair.
  • Religious and Economic Dominance: The religious semantic area consistently exhibits the highest adverse bias, followed closely by economic and ethic areas.
  • Racial and Conspiratorial Rise: Racial, conspiratorial, and socio-political semantic areas show a steady adverse bias that increases significantly after 1886, correlating with the rise of racial theories and conspiracy narratives.
  • Comparative Biases:
    • "Juif" shows an entirely adverse bias, while "catholique" shows an entirely favorable bias, highlighting a stark contrast.
  • "Protestant" and "musulman" groups show adverse biases, particularly in the religious sphere, with nuances related to historical events like citizenship debates.
  • Slang terms like "youpin" exhibit exclusively adverse bias, similar to "juif," suggesting their interchangeable use in derogatory contexts.

Limitations and Future Work

  • Corpus Limitations: The reliance on the Gallica corpus means the findings are specific to the digitized French print media of the period. The OCR quality threshold (98%) might still introduce some noise.
  • Keyword Selection: The corpus was built around keywords related to Jews. While comprehensive, this approach might miss broader antisemitic discourse that doesn't explicitly use these terms.
  • Interpretation of Bias: While the paper quantifies bias, the precise historical mechanisms and causal links between specific events and semantic shifts require further qualitative historical investigation.
  • Future Work: The authors suggest further exploring the association between adverse bias and antisemitic discourse through detailed examination of historical contexts and literary texts. They also propose extending this analysis to other languages and historical periods.

Summary updated

BibTeX
@inproceedings{tripodi2019tracing,
  author = {Rocco Tripodi and Massimo Warglien and Simon Levis Sullam and Deborah Paci},
  title = {{Tracing Antisemitic Language Through Diachronic Embedding Projections: France 1789-1914}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure