Explicit Query Rewriting for Conversational Dense Retrieval

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4725 - 4737 December 7-11, 2022 ©2022 Association for Computational Linguistics Explicit Query Rewriting for Conversational Dense Retrieval Hongjin Qian, Zhicheng Dou∗ Gaoling School of Artificial Intelligence, Renmin University of China {ian,dou}@ruc.edu.cn Abstract In a conversational search scenario, a query might be context-dependent because some words are referred to previous expressions or omitted. Previous works tackle the issue by either reformulating the query into a self- contained query (query rewriting) or learning a contextualized query embedding from the query context (context modelling). In this pa- per, we propose a model CRDR that can per- form query rewriting and context modelling in a unified framework in which the query rewrit- ing’s supervision signals further enhance the context modelling. Instead of generating a new query, CRDR only performs necessary modifi- cations on the original query, which improves both accuracy and efficiency of query rewriting. In the meantime, the query rewriting benefits the context modelling by explicitly highlighting relevant terms in the query context, which im- proves the quality of the learned contextualized query embedding. To verify the effectiveness of CRDR, we perform comprehensive experi- ments on TREC CAsT-19 and TREC CAsT-20 datasets, and the results show that our method outperforms all baseline models in terms of both quality of query rewriting and quality of context-aware ranking. 1 Introduction The recent rising of intelligent assistants triggers the transition of information retrieval systems from ad-hoc search to conversational search (Croft, 2019; Gao et al., 2020; Ren et al., 2021). Search- ing in a conversational manner provides interactive information exchange between users and the sys- tem. Such interactive format is usually achieved by multi-turn dialogues, with which the system can understand the user’s complex information needs (Dalton et al., 2020b,a). In a conversational system, a user often asks follow-up questions about something referred to ∗Corresponding author. Where is the Phoenix city? What is its population? … with 1,680,992 people (as of 2019). It is also the fifth-most populous city in … How about ( the population of ) New York? … With a 2020 population of 8,804,190 distributed over 302.6 square miles … Anaphora Ellipsis q1 q2 q3 … Phoenix is the capital and most populous city in Arizona … Figure 1: A case of Conversational Search previous dialogue context (Vakulenko et al., 2021a; Radlinski and Craswell, 2017; Ren et al., 2021). Therefore, a conversational search query is likely to be context-dependent and not self-contained. The real user intent is hidden underneath not only the current query but the whole query context. Fig- ure 1 shows an example of conversational search, from which we can find two most typical linguistic phenomenons that undermine the semantic com- pleteness of a single query in the conversation, namely anaphora and ellipsis (Vakulenko et al., 2020). Specifically, anaphora refers to the phe- nomenon of an expression that depends on an ex- pression in the previous context. For example, in Figure 1, the word “its” in q2 …

Content truncated to 500 words