HateMirage: An Explainable Multi-Dimensional Dataset for Decoding Faux Hate and Subtle Online Abuse

1. Problem Statement

The paper addresses a critical gap in online safety research: the detection and explanation of Faux Hate.

• The Challenge: Traditional hate speech detection focuses on overt toxicity (slurs, profanity). However, a growing form of harmful content, termed "Faux Hate," embeds prejudice within misinformation, deceptive narratives, or fabricated claims.
• The Limitation: Existing datasets (e.g., HateXplain, HARE) primarily capture surface-level toxicity or provide token-level rationales. They fail to capture the causal link between a debunked fake claim and the resulting hate speech.
• The Need: There is a lack of resources that explain why a comment is harmful by connecting the underlying misinformation to the specific target, the author's intent, and the broader social implications. Current models struggle with the contextual reasoning required to identify these subtle, indirect attacks.

2. Methodology

The authors propose a novel dataset and a benchmarking framework to address these issues.

A. Dataset Construction (HateMirage)

• Source Material: The dataset is built upon a curated set of 4,530 user comments collected from YouTube videos associated with widely debunked misinformation claims.
• Data Pipeline:
  1. Claim Identification: Widely debunked fake claims were sourced from fact-checking websites (e.g., AltNews, FactChecker).
  2. Comment Extraction: Related YouTube videos were identified, and comments were scraped. The data includes English and code-mixed Hindi-English comments.
  3. Filtering: Comments were labeled for Hate (0/1) and Fake (0/1) using a GPT-4 model augmented with Retrieval-Augmented Generation (RAG) and human-in-the-loop validation. Only "Faux Hate" instances (Fake=1, Hate=1) were retained.
  4. Translation: Code-mixed comments were translated to English to ensure consistency in downstream processing.
• Annotation Schema: Unlike binary classification, each comment is annotated with three interpretable dimensions:
  1. Target: Who is being attacked (explicit or implicit entities).
  2. Intent: The underlying motivation (e.g., vilification, delegitimization, polarization).
  3. Implication: The potential social impact (e.g., fueling anger, eroding trust, normalizing profiling).
• Quality Control: A human-in-the-loop process validated 10–15% of the data, checking for factual consistency, semantic coherence, and linguistic clarity.

B. Benchmarking Framework

• Task: Structured explanation generation (predicting Target, Intent, and Implication).
• Models Evaluated: A range of open-source models from 1B to 8B parameters, including:
  • SLMs: LLaMA-3.2 (1B, 3B), Qwen-2.5 (1.5B, 3B), Phi-3-128k (3B).
  • LLMs: Mistral-v0.3-7B, LLaMA-3.1-8B.
• Experimental Settings:
  • Zero-Shot: Models rely solely on pre-trained knowledge.
  • RAG-Based: Models are provided with retrieved evidence from fact-checking databases to ground their reasoning.
• Evaluation Metrics:
  • Sentence-BERT (SBERT) Similarity: To measure semantic alignment.
  • ROUGE-L F1: To measure lexical and structural overlap.

3. Key Contributions

1. HateMirage Dataset: The first dataset specifically designed for Faux Hate, containing 4,530 comments with multi-dimensional structured explanations (Target, Intent, Implication). It uniquely integrates hate speech and misinformation analysis.
2. Multi-Dimensional Explanation Framework: Moves beyond simple classification to provide granular reasoning about the mechanism of harm, linking deceptive narratives to specific social consequences.
3. Comprehensive Benchmark: Establishes baselines for explanation generation across various model sizes and architectures, highlighting the difficulty of reasoning in this domain.
4. Interdisciplinary Scope: Unifies hate speech detection and misinformation research, addressing the intersection where fake narratives fuel hate.

4. Experimental Results & Analysis

• Model Performance:
  • Phi-3-128k-3B demonstrated superior performance in identifying Targets and Intents in both Zero-Shot and RAG settings. This suggests that smaller models trained on reasoning-focused data can outperform larger models in factual and role-focused tasks.
  • Mistral-v0.3-7B performed best on the Implication dimension, which requires abstract reasoning about societal effects.
  • RAG Impact: The inclusion of retrieved fact-checking evidence generally improved grounding, particularly for the "Target" and "Intent" fields, though gains varied by model architecture.
• Error Analysis:
  • Models often relied on surface-level cues (e.g., prominent nouns) rather than context, leading to incorrect Target identification in sarcastic or indirect comments.
  • Intent explanations were often generic (e.g., "spreading hate") rather than capturing nuanced motives like political positioning or racialized mockery.
  • Implication was the most challenging dimension, with models failing to predict specific social consequences (e.g., scapegoating) compared to broad generalizations.
• Human Evaluation:
  • GPT-4 generated explanations showed high semantic similarity (SBERT) to human annotations for Target (74.8%) but lower scores for Intent (70.2%) and Implication (57.8%).
  • This confirms that while models can identify entities, reasoning about complex causal consequences remains a significant challenge.

5. Significance and Future Work

• Significance: HateMirage advances the field of Explainable AI (XAI) in NLP by making implicit harm explicit. It provides a tool for researchers and moderators to understand how misinformation triggers hate, rather than just flagging the hate itself. This is crucial for developing trustworthy, interpretable moderation systems.
• Limitations: The dataset relies on synthetic explanations generated by GPT-4 (though validated by humans), which may lack the full emotional nuance of human interpretation. The data is currently limited to YouTube comments on international news channels.
• Future Directions:
  • Extending the dataset to the multimodal domain (memes, visual narratives).
  • Developing faithfulness metrics to evaluate the quality of reasoning beyond lexical similarity.
  • Creating automated evaluators for multi-dimensional explanations to scale up research.

In conclusion, HateMirage establishes a new standard for studying the interplay between misinformation and hate speech, emphasizing that effective detection requires deep contextual reasoning and multi-dimensional explainability rather than simple pattern matching.