KNOT-MCTS: An Effective Approach to Addressing Hallucinations in Generative Language Modeling for Question Answering

KNOT-MCTS: An Effective Approach to Addressing Hallucinations in Generative Language Modeling for Question Answering Chung-Wen Wu, Guan-Tang Huang, Yue-Yang He, Berlin Chen Department of Computer Science and Information Engineering National Taiwan Normal University {40947040s, 40947025s, 40947005s, berlin}@ntnu.edu.tw Abstract Contemporary large language mod- els (LLMs) have made significant advancements, capable of generating fluent conversations with humans and accomplishing various tasks such as programming and question answering (QA). Nevertheless, current LLMs are still faced with numerous challenges, including generating hallucinations, lacking the latest information, suffering from biases, and others. In this paper, we proposed a technique, Knowledge-based Navigation for Optimal Truthfulness Monte Carlo Tree Search (KNOT-MCTS), which can reduce hallucinations of LLMs by aligning semantics of responses with external knowledge during the generation process. This technique acts as a plug-and-play knowledge injection method, which does not require any training and can be applied to any (large) language model. First, snippets,we retrieve relevance knowledge incorporating them into the prompt section and subsequently fed into the decoding process. Then, during the decoding process, we utilize our semantic alignment heuristic function to guide the response generation process of LMs through the Monte Carlo Tree Search (MCTS) decoding process. In our experiments on the TruthfulQA dataset, KNOT-MCTS paired with various LMs consistently outperforms their respec- tive baselines. Our results demonstrate that KNOT-MCTS can effectively inject knowledge into various LMs to reduce hallucinations of LMs. Keywords: Monte Carlo Tree Search, Knowledge Retrieval, Knowledge Injection, Semantic Alignment 1 Introduction In this era, large language models (LLMs) have played an increasingly significant role in our lives. However, apart from scientific and humanistic knowledge, there is also a abun- dance of myths, urban legends, fake news, and other misleading information. During train- ing or task execution, there may be instances where we reference this information. Despite the convenience these powerful models bring to our lives, we still need to pay attention to the untrue responses due to the hallucination (Maynez et al., 2020; Zhang et al., 2023). In ad- dition, many researches also raise issues with biases (Sap et al., 2019; Abid et al., 2021) and imitative falsehoods (Lin et al., 2021). Mentioned by Askell et al. (2021), we ex- pect a helpful, honest, and harmless (called as ‘HHH’) AI model. There are several main improvement approaches to reduce untrue or harmful responses: • Filter the training dataset of the model to avoid it learning incorrect or misleading information. • Perform fine-tuning and adjustments on the model during task execution to ad- dress its shortcomings and prevent hallu- cinations in its responses. Therefore, in this paper we purpose a plug- and-play technique, named KNOT-MCTS, which can enhance the truthfulness and ac- curacy of the model’s responses. Without any training, KNOT-MCTS retrieve external knowledge snippets based on a question, incor- porate those snippets into a prompt, and input it into the Monte Carlo Tree Search (MCTS) decoding process. Our MCTS decoding pro- cess is driven by the incorporation of a se- The 35th Conference on Computational Linguistics and Speech Processing (ROCLING …

Content truncated to 500 words