Multi-Stage Conversational Passage Retrieval: An Approach to Fusing Term Importance Estimation and Neural Query Rewriting

Abstract

Conversational search plays a vital role in conversational information seeking. As queries in information seeking dialogues are ambiguous for traditional ad hoc information retrieval (IR) systems due to the coreference and omission resolution problems inherent in natural language dialogue, resolving these ambiguities is crucial. In this article, we tackle conversational passage retrieval, an important component of conversational search, by addressing query ambiguities with query reformulation integrated into a multi-stage ad hoc IR system. Specifically, we propose two conversational query reformulation (CQR) methods: (1) term importance estimation and (2) neural query rewriting. For the former, we expand conversational queries using important terms extracted from the conversational context with frequency-based signals. For the latter, we reformulate conversational queries into natural, stand-alone, human-understandable queries with a pretrained sequence-to-sequence model. Detailed analyses of the two CQR methods are provided quantitatively and qualitatively, explaining their advantages, disadvantages, and distinct behaviors. Moreover, to leverage the strengths of both CQR methods, we propose combining their output with reciprocal rank fusion, yielding state-of-the-art retrieval effectiveness, 30% improvement in terms of NDCG@3 compared to the best submission of Text REtrieval Conference (TREC) Conversational Assistant Track (CAsT) 2019.

Conversational Information Seeking

“…Following this, we discuss open retrieval from large text corpora as part of the QA process. In particular, Section 5.2 goes beyond short answer QA to approaches performing conversational passage retrieval from open text collections including multi-stage neural ranking, for instance recently considered by Lin et al (2021). We briefly discuss long answer generation approaches in Section 5.3 including both extractive and abstractive summarization methods.…”

“…The best performing model applies diverse techniques: rewriting, expansion, and reranking in a multi-stage pipeline (Lin et al, 2020c). An approach based upon both early and late fusion of multiple expansions and rewrites across both retrieval and reranking is currently required for effectiveness (Lin et al, 2020c;Lin et al, 2021). This indicates there is opportunity for more unified approaches combining the different sub-components.…”

“…It is common to use a multi-stage cascaded architecture for OR-ConvPR tasks. One of the prototypical multi-stage systems that perform this is developed by Lin et al (2021). A core building block of their approach is Historical Query Expansion (HQE) that generates expanded queries based upon the dialogue history using a sequence-to-sequence model.…”

Text-to-Text Multi-view Learning for Passage Re-ranking

“…Many studies have adopted the standard multi-stage passage ranking pipeline in ad-hoc search [25] while using the reformulated query 𝑞 ′ as an ad-hoc query. Examples include the works of [15,21,34,36]. Specifically, an effective ConvSearch system consists of a CQR module F CQR , a first-stage retriever F RT and a second-stage passage re-ranker F RR , as follows:…”

“…Among all the ConvSearch systems, the multi-stage cascaded architecture has proven to be the most effective approach, which addresses the issue of query ambiguity in ConvSearch through the addition of a conversational query reformulation (CQR) module that employs via heuristic [35,38] or neural approaches [1,15,21,34,36,39]. Although effective, such additional modules may increase query latency and complexity, posing challenges for realworld deployment.…”