Dial-QP: A Multi-tasking and Keyword-Guided Approach for Enhancing Conversational Query Production.

Recent works have utilized a dynamic process to seek up-to-date knowledge from search engines, yielding promising performance improvement in the field of knowledge-grounded response generation models. This pipeline consists of two stages: a Conversational Query Production (CQP) stage to acquire knowledge and a Response Generation to generate the final response. In this paper, we focus on the CQP task, which aims to condense the dialogue context into a concise query that is then fed to search engines. Previous studies have treated the problem as a generative task and always generated queries solely based on the dialogue context in an end-to-end manner. However, such a straightforward approach suffers from the followings: 1) the CQP model is hard to determine whether generating a query is necessary since not all scenarios require retrieval knowledge; 2) redundant content in the dialogue context can make it difficult for the model to recognize the key information that reflects the user’s concerns, particularly in multi-turn conversations. To address these challenges, we propose a novel BART-based Dial-QP for the CQP task, which decomposes the CQP task into Query Classification and Query Production tasks and then introduces a multitasking training paradigm to improve the decision-making process of No Query. Meanwhile, a Keyword Recognition stage helps the model focus on the key information. Extensive experiments and analyses on two CQP datasets DuSinc and WoI have demonstrated the effectiveness of Dial-QP. Our code is available at: https://github.com/Y-NLP/Chatbots/tree/main/NLPCC2023_Dial-QP.