TL;DR
This paper introduces PHATE, the first comprehensive multi-label Persian hate speech dataset with detailed annotator rationales. By providing not just labels but also the specific text spans justifying those labels, PHATE aims to advance hate speech detection research, particularly in low-resource languages, and foster more interpretable AI models.
What Problem Does This Paper Address?
The proliferation of online hate speech necessitates robust automated detection systems. However, developing these systems heavily relies on high-quality annotated datasets, which are scarce, especially for languages other than English. Existing Persian hate speech datasets are limited, often non-public, or suffer from simplistic data collection methods that can introduce bias. Furthermore, understanding why a model flags something as hate speech is crucial for building trust and improving accuracy, yet most datasets lack this explanatory information. This paper addresses the dual challenge of creating a rich, nuanced Persian hate speech dataset and incorporating annotator rationales to facilitate interpretable hate speech detection.
Methods
The authors constructed the PHATE dataset by collecting over 7,000 Persian tweets from Twitter. They employed a multi-stage approach:
- Data Collection: Instead of relying on explicit slur lists, they curated controversial and ambiguous Persian words to minimize bias. They used Twitter's Premium API to gather original tweets, excluding retweets and replies.
- Initial Classification (ChatGPT): To pre-filter the collected tweets, they used ChatGPT in a zero-shot manner to classify tweets into "normal," "hate speech," or "cannot identify." This step helped create a more balanced dataset for subsequent manual annotation.
- Manual Annotation: Two graduate students, selected for their high annotation consistency, manually annotated the tweets. They followed detailed guidelines to classify tweets as either "normal" or "hate speech."
- Hierarchical Labeling: Hate speech was further categorized into three sub-labels: "Violence," "Hate," and "Vulgar." This creates a multi-level, multi-label classification task.
- Rationale Annotation: Crucially, annotators were instructed to highlight the specific text spans within a tweet that justified their assigned hate speech label. They also identified the target of the hate speech and its identity dimension (e.g., religious, racial).
- Quality Assurance: A third annotator (one of the authors) adjudicated any disagreements between the initial two annotators to ensure high-quality gold annotations. Inter-annotator agreement metrics (Fleiss' Kappa) were calculated to validate the annotation process.
- Rationale-Assisted Fine-tuning: The authors explored a novel fine-tuning approach where the annotated rationales were used to guide the model's attention, aiming to improve performance and interpretability. This involved creating "rationale masks" and incorporating an additional loss function during training.
Datasets
The primary contribution is the PHATE dataset, a new benchmark for Persian multi-label hate speech detection. It contains over 7,000 manually annotated Persian tweets, featuring:
- Binary classification (Normal vs. Hate Speech).
- Multi-label classification for hate speech subcategories: Violence, Hate, Vulgar.
- Annotator rationales (highlighted text spans) explaining the label assignment.
- Information about the target of the hate speech.
The paper also references and discusses existing hate speech datasets, including Pars-OFF (Persian), SemEval-2021 Task 5, DOSA, HateXplain (English), and multilingual datasets like mBERT and XLM-R, to contextualize PHATE's novelty.
Key Insights and Ideas
- The Value of Rationales: The paper strongly suggests that annotator rationales are not just descriptive but can be actively used to improve model performance. Masking these rationales significantly degrades model performance, highlighting their importance for understanding and detecting hate speech.
- Low-Resource Language Challenges: The performance gap between state-of-the-art multilingual models and even human evaluators on the Persian dataset underscores the ongoing challenges in low-resource language NLP.
- Bias Mitigation in Data Collection: The authors' deliberate choice to avoid explicit slur lists and instead use ambiguous words for data collection is a key strategy to mitigate inherent biases in hate speech datasets.
- Interpretability as a Goal: PHATE is designed not just for classification accuracy but also to enable research into why models make certain predictions, aligning with the growing demand for explainable AI.
Key Results
- Inter-Annotator Agreement: High agreement scores (Kappa of 0.72 for binary classification, and F1 agreement of 0.88 for hate speech) indicate clear annotation guidelines and reliable data.
- Model Performance: ParsBERT and XLM-R performed best among the tested models, outperforming mBERT and ChatGPT (zero-shot). However, all models showed a significant performance gap compared to human evaluators.
- Rationale-Assisted Fine-tuning: This novel approach showed promising results, leading to performance improvements (e.g., increased recall for Violence) compared to standard fine-tuning, demonstrating the utility of rationales in guiding model attention.
- Human Performance: Human evaluators achieved high F1 scores across categories, serving as a benchmark, though even they showed variations, particularly in distinguishing "Vulgar" from "Hate."
Limitations and Future Work
- Platform Specificity: PHATE is derived from Twitter, meaning it may not fully represent hate speech across all online platforms or contexts.
- Language and Cultural Nuances: The dataset and results are specific to Persian language and its cultural context, limiting direct transferability without adaptation.
- Annotator Subjectivity: Despite clear guidelines, human annotation inherently carries some degree of subjectivity and potential for inconsistency.
- Scope of Hate Speech: The dataset focuses on detecting hate speech but doesn't explore the underlying causes or contextual factors that contribute to its production.
- Future Work: The authors suggest that PHATE can serve as a valuable resource for developing more interpretable models and for creating similar datasets in other languages. The rationale-assisted fine-tuning method also presents a promising avenue for future research.
Summary updated