Multiview Identifiers Enhanced Generative Retrieval

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6636–6648 July 9-14, 2023 ©2023 Association for Computational Linguistics Multiview Identifiers Enhanced Generative Retrieval Yongqi Li1, Nan Yang2, Liang Wang2, Furu Wei2, Wenjie Li1 1The Hong Kong Polytechnic University 2Microsoft liyongqi0@gmail.com {nanya,wangliang,fuwei}@microsoft.com cswjli@comp.polyu.edu.hk Abstract Instead of simply matching a query to pre- existing passages, generative retrieval gener- ates identifier strings of passages as the re- trieval target. At a cost, the identifier must be distinctive enough to represent a passage. Current approaches use either a numeric ID or a text piece (such as a title or substrings) as the identifier. However, these identifiers cannot cover a passage’s content well. As such, we are motivated to propose a new type of identifier, synthetic identifiers, that are generated based on the content of a passage and could integrate contextualized information that text pieces lack. Furthermore, we simultaneously consider mul- tiview identifiers, including synthetic identi- fiers, titles, and substrings. These views of iden- tifiers complement each other and facilitate the holistic ranking of passages from multiple per- spectives. We conduct a series of experiments on three public datasets, and the results indicate that our proposed approach performs the best in generative retrieval, demonstrating its effec- tiveness and robustness. The code is released at https://github.com/liyongqi67/MINDER. 1 Introduction Text retrieval is a fundamental task in informa- tion retrieval and plays a vital role in various lan- guage systems, including search ranking (Nogueira and Cho, 2019) and open-domain question answer- ing (Chen et al., 2017). In recent years, the dual- encoder approach (Lee et al., 2019; Karpukhin et al., 2020), which encodes queries/passages into vectors and matches them via the dot-product oper- ation, has been the de-facto implementation. How- ever, this approach is limited by the embedding space bottleneck (Lee et al., 2022a) and missing fine-grained interaction (Wang et al., 2022b). An emerging alternative to the dual-encoder ap- proach is generative retrieval (De Cao et al., 2020; Tay et al., 2022; Bevilacqua et al., 2022). Gen- erative retrieval utilizes autoregressive language Query: Who is the singer of does he love you? Passage (https://en.wikipedia.org/wiki/Does_He_Love_You) "Does He Love You" is a song written by Sandy Knox and Billy Stritch, and recorded as a duet by American country music artists Reba McEntire and Linda Davis. It was released in August 1993 as the first single from Reba's album "Greatest Hits Volume Two". It is one of country music's several songs about a love triangle. "Does He Love You" was written in 1982 by Billy Stritch. ...... Multiview Identifiers Title: Does He Love You Substrings: "Does He Love You" is a song ..., recorded as a duet by American country music artists Reba McEntire and Linda Davis, ... Pseudo-queries: Who wrote the song does he love you? Who sings does he love you? When was does he love you released by reba? What is the first song in the album "Greatest Hits Volume Two" about? Relevant Figure 1: An example of multiview identifiers …

Content truncated to 500 words