TL;DR
This paper investigates how sociocultural factors influence the effectiveness of automated systems designed to detect anti-LGBTQ+ hate speech on social media. The authors find that existing open-source training data, often collected using keyword searches, can lead to models that overfit on specific slurs and miss nuanced forms of hate speech, particularly across different English language varieties.
What Problem Does This Paper Address?
The proliferation of online hate speech, especially against marginalized communities like LGBTQ+ individuals, poses a significant threat, leading to digital exclusion and barriers to access. While Natural Language Processing (NLP) has seen advancements in hate speech detection, most research focuses on methodological improvements without adequately considering the sociocultural context. This is problematic because hate speech is deeply intertwined with social, cultural, and political attitudes, which vary significantly across different regions and linguistic communities. Existing open-source hate speech detection systems, often trained on data from specific contexts, may not be suitable for monitoring content in diverse global settings, potentially leading to inaccurate detection and a failure to address the full scope of the problem.
Methods
The researchers developed two multiclass classification models to detect anti-LGBTQ+ hate speech. They employed a transformer-based language modeling approach, specifically using XLM-ROBERTA, a cross-lingual model. To address class imbalance, they used techniques like up-sampling minority classes and collapsing categories. Domain adaptation was also applied to fine-tune the models for specific language varieties. The study then applied these models to georeferenced social media data (Twitter posts) from various countries to monitor anti-LGBTQ+ hate speech. They used a mixed-methods approach, combining quantitative analysis of model outputs with qualitative insights to interpret the results and understand the impact of sociocultural factors.
Datasets
The study utilized two open-source training datasets:
- MLMA (Multilingual and Multi-Aspect Hate Speech Data Set): This dataset, primarily from X (Twitter), uses a keyword search approach to identify hate speech. It includes English, French, and Arabic data.
- LTEDI (Language Technology for Equality, Diversity, and Inclusion): This dataset consists of YouTube comments related to LGBTQ+ content, manually annotated by members of the LGBTQ+ community. It includes English, Tamil, and English-Tamil data.
The researchers also collected a subset of the CGLU corpus, containing georeferenced posts/tweets from X (Twitter) in English, from both inner-circle (e.g., Australia, Canada, UK, US) and outer-circle (e.g., Ghana, India, Kenya, Malaysia, Pakistan) varieties of English.
Key Insights and Ideas
- Keyword-based training data can be a double-edged sword: While effective for identifying explicit slurs, it can lead models to overfit on these specific terms, missing more subtle or context-dependent hate speech.
- Sociocultural alignment matters: The LTEDI dataset, annotated by LGBTQ+ individuals and aligned with a South Asian context, showed better performance in identifying anti-LGBTQ+ hate speech in outer-circle English varieties compared to the MLMA dataset, which was developed with a more culturally-agnostic approach.
- Linguistic variation is crucial: Different varieties of English (inner-circle vs. outer-circle) exhibit distinct patterns of hate speech, and models trained on one may not generalize well to others.
- Qualitative insights are essential: Relying solely on quantitative metrics like F1-scores can be misleading. Understanding the why behind the predictions, through qualitative analysis of word frequencies and contextual understanding, is vital for accurate interpretation.
- "Cleaning" slurs is a form of resistance: The paper touches upon the idea that re-appropriating stigmatizing labels can be a form of resistance and identity reclamation for marginalized groups.
Key Results
- The LTEDIR model (LTEDI dataset with domain adaptation) generally predicted higher rates of anti-LGBTQ+ hate speech compared to the MLMAR model (MLMA dataset with domain adaptation).
- The MLMAR model showed a tendency to identify more anti-LGBTQ+ hate speech in inner-circle varieties of English, while the LTEDIR model was more sensitive to outer-circle varieties.
- Wordcloud analysis revealed that the MLMA training data heavily featured specific slurs (e.g., "faggot," "dyke"), but the predicted outputs showed "sleep" and "gay" as more frequent, suggesting a disconnect. The LTEDI predicted outputs, conversely, showed a higher correlation with community and religious terms, reflecting the dataset's focus.
- The study found that the growth rate of detected anti-LGBTQ+ hate speech remained relatively stable over time across different English varieties, despite fluctuations in platform usage.
Limitations and Future Work
- Invisibility of Q+ identities: The study acknowledges that the LGBTQ+ acronym may not fully encompass all gender and sexual minorities, potentially excluding intersex individuals and other expressions of gender and sexuality.
- Sociocultural bias in data collection: The MLMA's keyword search approach led to overfitting, and the LTEDI's religious subtext in training data might misinterpret certain content. The models also don't account for semantic bleaching or slur reclamation.
- Pitfalls of large language models: The paper acknowledges that PLMs can introduce cultural and linguistic biases, though domain adaptation was used to mitigate some of these.
- Class imbalance: While addressed through up-sampling, it remains a challenge.
- Future Work: The authors suggest that NLP researchers should increase the visibility of Q+ communities and address sociocultural biases in training data and large language models. They also emphasize the need for collaboration with stakeholders, including affected communities, advocates, and social media platforms, to develop effective solutions.
Summary updated