Fusing Semantics and LLM-Driven Emotions by Gated Attention for Implicit Hate Speech Detection

Abstract

Detecting implicit hate speech continues to be a significant challenge for natural language processing systems, as harmful intent is often concealed within polite, neutral, or coded language. Traditional text-based classifiers often struggle to identify these subtle forms of hostility because they lack the emotional context needed to interpret the underlying aggression. To address this issue, we propose the Emotion Gated Attention Fusion Network (EGAF-Net), an architecture that integrates textual semantics with psychological emotions. Unlike conventional approaches that rely on static lexicons, our method utilizes a Large Language Model (LLM) to extract a 9-dimensional affective vector directly from the text. This vector represents Plutchik's eight basic emotions along with an overall measure of arousal. Simultaneously, we employ a frozen pretrained Transformer model (RoBERTa) to capture deep linguistic semantics. To combine these two modalities, we introduce a feature-wise Gated Attention Layer that utilizes a sigmoid-driven masking mechanism to reduce modality-specific noise and dynamically adjust the scale of semantic-emotional features. In evaluations on the Implicit Hate Corpus (IHC), our model demonstrates improved discriminative capability, achieving a competitive ROC-AUC score of 0.7714 and outperforming pure-semantic baselines. However, further architectural refinements are necessary to address the low recall rates observed across all tested baselines. Additionally, qualitative error analysis indicates that our gating mechanism successfully identifies several high-risk instances of implicit hate that traditional models misclassify as safe. These findings suggest that incorporating LLM-driven emotional intelligence is a promising approach to uncovering implicit toxicity.