NLP & Discourse ModerationAccepted Research Paper
Context-Aware Hate Speech Detection via Post-Comment Joint Modeling
Created a context-aware moderation framework replacing isolated text analysis with post-comment joint modeling, achieving 95% accuracy and ROC-AUC of 0.9911.
Classification Accuracy
95.0%
ROC-AUC Score
0.9911
Moderation Taxonomy
4 Action Classes
Embeddings
SBERT + DistilBERT
Abstract & Overview
Traditional automated moderation classifies isolated user comments, frequently hallucinating or missing nuanced toxicity in context-dependent discourse. We present a context-aware joint modeling framework that simultaneously embeds root posts and reply comment threads using SBERT semantic similarity, emotion distributions, and DistilBERT token probabilities, achieving 95% accuracy and 0.9911 ROC-AUC.
Problem Formulation
Automated content moderation systems frequently suffer from high false positive rates when evaluating sarcastic, slang-heavy, or context-reliant conversational comments without analyzing the root post context.
Architecture & Key Mechanisms
1Post-Comment Joint Modeling: Concatenates and co-embeds parent post context with user reply utterances.
2Semantic & Emotion Fusion: Fuses SBERT dense representations with fine-grained categorical emotion distributions.
3Actionable Moderation Taxonomy: Maps predictions directly to four explicit governance actions: Approve, Flag, Soft-Warn, and Auto-Quarantine.
Empirical Highlights
- Structured four-class reaction taxonomy mapping discourse to explicit moderation actions.
- Fused parallel features including SBERT semantic similarity, emotion distributions, and DistilBERT probabilities.
- Validated framework robustness on curated social media interaction dataset.
Technologies & Tooling
PythonSBERTDistilBERTNLPPyTorchScikit-learn