22 Mar 2024 | Zhenrui Yue, Huimin Zeng, Yimeng Lu, Lanyu Shang, Yang Zhang, Dong Wang
The paper addresses the issue of online misinformation by proposing a novel framework called Retrieval Augmented Response Generation (RARG) to generate counter-misinformation responses. RARG consists of two main stages: evidence collection and response generation. In the evidence collection stage, a two-stage retrieval pipeline is designed to efficiently retrieve and rerank scientific evidence from over 1 million academic articles. In the response generation stage, large language models (LLMs) are aligned with human feedback through reinforcement learning (RLHF) to generate polite and factual responses based on the collected evidence. The proposed reward function maximizes the utilization of retrieved evidence while maintaining text quality. Extensive experiments on COVID-19 misinformation datasets demonstrate that RARG outperforms existing methods in generating high-quality, evidence-based counter-responses. The framework's effectiveness is validated through both in-domain and cross-domain evaluations, showing its ability to handle diverse misinformation scenarios.The paper addresses the issue of online misinformation by proposing a novel framework called Retrieval Augmented Response Generation (RARG) to generate counter-misinformation responses. RARG consists of two main stages: evidence collection and response generation. In the evidence collection stage, a two-stage retrieval pipeline is designed to efficiently retrieve and rerank scientific evidence from over 1 million academic articles. In the response generation stage, large language models (LLMs) are aligned with human feedback through reinforcement learning (RLHF) to generate polite and factual responses based on the collected evidence. The proposed reward function maximizes the utilization of retrieved evidence while maintaining text quality. Extensive experiments on COVID-19 misinformation datasets demonstrate that RARG outperforms existing methods in generating high-quality, evidence-based counter-responses. The framework's effectiveness is validated through both in-domain and cross-domain evaluations, showing its ability to handle diverse misinformation scenarios.