ConvTrans: Transforming Web Search Sessions for Conversational Dense Retrieval

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2935 - 2946 December 7-11, 2022 ©2022 Association for Computational Linguistics ConvTrans: Transforming Web Search Sessions for Conversational Dense Retrieval Kelong Mao1, Zhicheng Dou1∗, Hongjin Qian1, Fengran Mo2, Xiaohua Cheng3, Zhao Cao3 1Gaoling School of Artificial Intelligence, Renmin University of China 2Université de Montréal, Québec, Canada 3Huawei Poisson Lab {mkl,dou}@ruc.edu.cn Abstract Conversational search provides users with nat- ural and convenient new search experience. Re- cently, conversational dense retrieval has shown to be a promising technique to realize conver- sational search. However, as conversational search systems have not been widely deployed, it is hard to get large-scale real conversational search sessions and relevance labels to support the training of conversational dense retrieval. To tackle this data scarcity problem, previous methods focus on developing better few-shot learning approaches or generating pseudo rele- vance labels, but the data they use for training still heavily rely on manual generation. In this paper, we present ConvTrans, a data augmentation method that can automatically transform easily-accessible web search ses- sions into conversational search sessions to fun- damentally alleviate the data scarcity problem for conversational dense retrieval. ConvTrans eliminates the gaps between these two types of sessions in terms of session quality and query form to achieve effective session transforma- tion. Extensive evaluations on two widely used conversational search benchmarks, i.e., CAsT- 19 and CAsT-20, demonstrate that the same model trained on the data generated by Con- vTrans can achieve comparable retrieval perfor- mance as it trained on high-quality but expen- sive artificial conversational search data. 1 Introduction Conversational search systems support multi-turn interaction with users through conversations to meet various information needs. Such a new search paradigm is more natural, convenient, and user- friendly, which is expected to hugely improve users’ search experience in the future (Culpepper et al., 2018). Different from traditional ad-hoc search systems, conversational search systems have to un- derstand the user’s real search intent of the current ∗Corresponding author. ConvTrans Relevance Labels WSS Sessions Original ConvTrans Intent: William Shakespeare Intent: Theme of Hamlet Web: William Shakespeare Conv: Who is William Shakespeare? Web: William Shakespeare poems Conv : Can you show some his poems? Web: When WS wrote hamlet Conv : When he wrote hamlet? Web: hamlet theme Conv : What is its theme? Search Session Example Intent: William Shakespeare poems Intent: Hamlet's finished time Conversational Session Encoder Relevance Labels Sessions Conversational Session Encoder Original v.s ConvTrans Artificial CSS Pseudo CSS train train Figure 1: Left: An example of a web search session (WSS) (orange) and a conversational search session (CSS) (blue) sharing similar search intents shift. Right: A high-level illustration of ConvTrans. turn based on the history conversation context, fac- ing a more complex search intent understanding problem (Gao et al., 2022). Recently, inspired by the success of dense re- trieval (Karpukhin et al., 2020), conversational dense retrieval has also been developed for con- versational search, whose core is to train a con- versational session encoder to directly encode …

Content truncated to 500 words