LLMs-Integrated Automatic Hate Speech Recognition Using Controllable Text Generation Models

TL;DR

Addressing hate speech in spoken content, the paper proposes a multitask framework that fuses an ASR encoder with an LLM decoder via a Q-Former bridge to transcribe and censor in one pass. It creates a labeled hate-speech training set through Chain-of-Thought prompts and HateXplain-based augmentation, masks hateful terms with *** during transcription, and uses curriculum learning to control hate-level exposure. The approach combines instruction tuning (via LoRA) and activation tuning (KL divergence) to maintain generalization while specializing on censorship tasks, and reports higher masking accuracy and favorable MAR/UMWER tradeoffs. The work offers a scalable path for safe, real-time speech moderation by leveraging synthetic data and integrated multimodal modeling.

Abstract

This paper proposes an automatic speech recognition (ASR) model for hate speech using large language models (LLMs). The proposed method integrates the encoder of the ASR model with the decoder of the LLMs, enabling simultaneous transcription and censorship tasks to prevent the exposure of harmful content. Instruction tuning of the LLM to mask hate-related words with specific tokens requires an annotated hate speech dataset, which is limited. We generate text samples using an LLM with the Chain-of-Thought (CoT) prompting technique guided by cultural context and examples and then convert them into speech samples using a text-to-speech (TTS) system. However, some of them contain non-hate speech samples with hate-related words, which degrades the censorship performance. This paper filters the samples which text classification models correctly label as hate content. By adjusting the threshold for the number of correct answer models, we can control the level of hate in the generated dataset, allowing us to train the LLMs through curriculum learning in a gradual manner. Experimental results show that the proposed method achieves a masking accuracy of 58.6\% for hate-related words, surpassing previous baselines. We also confirm that the curriculum training contributes to the efficiency of both transcription and censorship tasks.

Figures (2)

• Figure 1: Process for filtering candidate sentences of hate speech
• Figure 2: System archtecture of the LLM-integrated ASR model