Aiming at the problem of insufficient recognition of implicit variants by existing Chinese sensitive text detection methods, this paper proposes the IPKE-MoE framework, which consists of three parts, namely, a sensitive word variant extraction framework, a sensitive word variant knowledge enhancement layer and a mixture-of-experts (MoE) classification layer. First, sensitive word variants are precisely extracted through dynamic iterative prompt templates and the context-aware capabilities of Large Language Models (LLMs). Next, the extracted variants are used to construct a knowledge enhancement layer for sensitive word variants based on RoCBert models. Specifically, after locating variants via n-gram algorithms, variant types are mapped to embedding vectors and fused with original word vectors. Finally, a mixture-of-experts (MoE) classification layer is designed (sensitive word, sentiment, and semantic experts), which decouples the relationship between sensitive word existence and text toxicity through multiple experts. This framework effectively combines the comprehension ability of Large Language Models (LLMs) with the discriminative ability of smaller models. Our two experiments demonstrate that the sensitive word variant extraction framework based on dynamically iterated prompt templates outperforms other baseline prompt templates. The RoCBert models incorporating the sensitive word variant knowledge enhancement layer and a mixture-of-experts (MoE) classification layer achieve superior classification performance compared to other baselines.

In the interactive structure of social media platforms, the comment section has become the primary space for users to express their opinions. However, the proliferation of Internet troll comments has severely disrupted platform order and hindered the user experience. Existing governance methods mostly rely on manual intervention and traditional approaches, such as banning and deleting content. Still, these methods are inadequate in addressing troll comments’ large volume and complexity. This paper introduces an automatic response generation framework based on Large Language Models (LLMs), referred to as Human Feedback Based Response to Internet Trolls (TRHF), to address this issue. This framework utilizes multi-module semantic understanding, fine-grained classification, expert model-driven prompt design, and a reward constraint mechanism to generate responses that neutralize troll comments automatically. The framework’s performance is comprehensively evaluated using the Rational-Toxicity-Variety (RTV) metric, which assesses response effectiveness from rationality, toxicity suppression, and response diversity perspectives. Experimental results demonstrate that the TRHF framework enhances the effectiveness of intervention in troll comments and offers high real-time performance and ethical adaptability, providing a viable technological solution for intelligent governance on social media platforms.