Abstract
Automated content moderation often fails to distinguish between the perpetration of sexism and the reporting of it. This paper presents a framework for detecting sexism in French social media by categorizing content based on speech act theory. By distinguishing between direct/descriptive sexist acts and reported experiences (e.g., #MeToo testimonies), we aim to prevent the accidental censorship of victims. Using a new dataset of 12,000 French tweets and BERT-based models enhanced with linguistic features and generalization strategies, we demonstrate significant improvements in classification accuracy. Specifically, our best model achieves an F-score of 0.824 in binary classification, effectively "rescuing" non-sexist reports from over-moderation.
Introduction: The Blunt Instrument of Moderation
Automated content moderation is often a blunt instrument. Most current "sexism detectors" treat any message containing sexist language as a violation. However, there is a massive functional difference between a user hurling a sexist insult and a woman sharing a story about a sexist comment she received at work. If a moderation bot cannot distinguish between these, it risks "double-victimizing" women by deleting their testimonies of abuse. This research identifies a gap in existing NLP: while many models detect hate speech, few can distinguish the intent or discursive force behind the message, a challenge particularly acute in French discourse where nuanced datasets have been scarce.
Theoretical Framework: Speech Act Theory
The researchers approached the problem by categorizing sexist content into three types based on "perlocutionary force"—the effect the speech has on the listener:
- Direct Assertions: Second-person attacks (e.g., "You are a woman, you don't know football").
- Descriptive Assertions: Third-person stereotypes (e.g., "Women love cleaning").
- Reported Assertions: Sharing an experience (e.g., "My boss asked me who would cook for my husband while I'm away").
This distinction is crucial because reported speech acts as a distancing mechanism. When a user says "He said [sexist remark]," they are documenting the remark rather than endorsing it.
Dataset and Annotation
The authors introduced a new dataset of approximately 12,000 French tweets collected between 2017 and 2018 using keywords and hashtags like #balancetonporc and #payetashnek.
Manual annotation by five experts revealed that nearly 80% of tweets containing sexist content were actually "reporting" sexism. This high percentage highlights how easily a standard classifier could over-moderate these users if it fails to recognize the reporting context.
Experimental Setup: Models and Generalization
Several architectures were tested, including CNNs, LSTMs, and BERT. The standout approach involved BERT_base supplemented with:
- Linguistic Features: Incorporating emoji polarity, opinion lexicons, and "hedges" (negation/modality).
- Generalization Strategies: To help the model learn concepts rather than specific rare words, certain terms were replaced with tags (e.g., replacing specific names with
<Name_gendered>). - Cascade Classification: A two-stage process first deciding if a tweet has sexist content, then determining if it is a "report" (non-moderatable) or an "act" (moderatable).
Results and Analysis
The experimental results demonstrate the efficacy of combining contextual enrichment with generalization.
In binary classification (BIN), the best model—utilizing features and place/name generalization—achieved an accuracy of 0.837 and an F-score of 0.824. For the more complex 3-class task (Reporting vs. Non-Reporting vs. Non-Sexist), adding linguistic features and generalization strategies improved the F-score from a baseline of 0.540 to 0.655. The "Cascade" classifier proved particularly effective at "rescuing" non-sexist reports that standard classifiers would have flagged.
Conclusion and Future Work
While the proposed framework significantly improves the detection of reported sexism, challenges remain in identifying irony, satire, and complex metaphors that lack explicit "trigger words." Future work will focus on integrating irony detection and expanding the sexist vocabulary to better capture the evolving nature of online discourse. By distinguishing the "discursive force" of messages, we can create moderation tools that protect rather than silence victims of discrimination.
Summary updated