CORE: Cooperative Training of Retriever-Reranker for Effective Dialogue Response Selection

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3102–3114 July 9-14, 2023 ©2023 Association for Computational Linguistics CORE: Cooperative Training of Retriever-Reranker for Effective Dialogue Response Selection Chongyang Tao1, Jiazhan Feng2, Tao Shen3, Chang Liu2, Juntao Li4 Xiubo Geng1, Daxin Jiang1∗ 1Microsoft, Beijing, China 2Peking University, Beijing, China 3University of Technology Sydney, Sydney, Australia 4Soochow University, Suzhou, China 1{chotao,xigeng,djiang}@microsoft.com 3shentao@uts.edu.cn 2{fengjz,changliu}@pku.edu.cn 4ljt@suda.edu.cn Abstract Establishing retrieval-based dialogue systems that can select appropriate responses from the pre-built index has gained increasing attention. Recent common practice is to construct a two- stage pipeline with a fast retriever (e.g., bi- encoder) for first-stage recall followed by a smart response reranker (e.g., cross-encoder) for precise ranking. However, existing stud- ies either optimize the retriever and reranker in independent ways, or distill the knowledge from a pre-trained reranker into the retriever in an asynchronous way, leading to sub-optimal performance of both modules. Thus, an open question remains about how to train them for a better combination of the best of both worlds. To this end, we present a cooperative training of the response retriever and the reranker whose parameters are dynamically optimized by the ground-truth labels as well as list-wise super- vision signals from each other. As a result, the two modules can learn from each other and evolve together throughout the training. Exper- imental results on two benchmarks demonstrate the superiority of our method. 1 Introduction The development of a smart human-computer con- versation system has been a longstanding objective in the field of artificial intelligence. Recent years have seen an increase in interest in constructing dialogue systems through data-driven approaches, leveraging advancements in deep learning tech- niques (Vaswani et al., 2017; Devlin et al., 2019). With the help of information retrieval (IR) tech- niques to select an appropriate response from a pre-built index (Lowe et al., 2015; Whang et al., 2020), or text generation techniques to synthesize a ∗Corresponding author. response (Zhang et al., 2019), existing neural mod- els are now capable of providing natural replies to user queries. In this paper, we concentrate on retrieval-based dialogue systems (Lowe et al., 2015; Boussaha et al., 2019; Yu et al., 2021; Su et al., 2021), which can deliver smooth and informative responses, and have powered industrial applica- tions (Shum et al., 2018; Ram et al., 2018). Retrieval-based dialogue systems usually fol- low the retrieval-reranking paradigm (Wang et al., 2013; Li et al., 2017), i.e., two-stage retrieval model, where the model first retrieves a bundle of response candidates from a pre-built index by a fast retriever and then selects an appropriate one with a more sophisticated yet costly response reranker. Specifically, as for the retriever, early methods based on hand-crafted features (Robertson et al., 2004; Qiu et al., 2017) (e.g., BM25) for fast re- trieval, however suffering from vocabulary mis- match problem, especially in context-to-response retrieval. A recent trend is resorting to deep neural model to represent text as dense embeddings in latent semantic space, which is …

Content truncated to 500 words