Towards social safety in digital platforms: multimodal hostile content detection and explanation

Abstract

The exponential rise of user-generated content on social media has fostered connectivity but also facilitated the proliferation of hostile narratives, ranging from explicit abuse to subtle, implicit forms of hate speech and dark humor. While automated detection systems have advanced, significant gaps remain in addressing low-resource languages, multimodal complexity, and the interpretability of model decisions. This thesis proposes a comprehensive suite of frameworks to detect and understand hostile content across three modalities: text, memes, and video. In the textual domain, the thesis addresses the problem in low-resource Indic languages. First, it introduces a user-aware framework for abusive content detection in multilingual and code-mixed environments. By integrating social context features, such as user history and post affinity, with cross-lingual textual embeddings, the proposed method significantly enhances detection performance in low-resource settings. Second, to bridge the gap between black-box predictions and human reasoning, the thesis presents X-MuTeST, an explainable framework for hate speech detection in Hindi, Telugu, and English. This work contributes benchmark datasets with token-level human rationales and employs a novel training strategy that combines Large Language Model (LLM) consultation with N-gram-based explainability to improve both plausibility and faithfulness.