Phrase Retrieval Learns Passage Retrieval, Too

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3661–3672 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 3661 Phrase Retrieval Learns Passage Retrieval, Too Jinhyuk Lee1,2∗ Alexander Wettig1 Danqi Chen1 Department of Computer Science, Princeton University1 Department of Computer Science and Engineering, Korea University2 {jinhyuklee,awettig,danqic}@cs.princeton.edu Abstract Dense retrieval methods have shown great promise over sparse retrieval methods in a range of NLP problems. Among them, dense phrase retrieval—the most fine-grained re- trieval unit—is appealing because phrases can be directly used as the output for question an- swering and slot filling tasks.1 In this work, we follow the intuition that retrieving phrases naturally entails retrieving larger text blocks and study whether phrase retrieval can serve as the basis for coarse-level retrieval includ- ing passages and documents. We first observe that a dense phrase-retrieval system, without any retraining, already achieves better passage retrieval accuracy (+3-5% in top-5 accuracy) compared to passage retrievers, which also helps achieve superior end-to-end QA perfor- mance with fewer passages. Then, we pro- vide an interpretation for why phrase-level su- pervision helps learn better fine-grained entail- ment compared to passage-level supervision, and also show that phrase retrieval can be im- proved to achieve competitive performance in document-retrieval tasks such as entity link- ing and knowledge-grounded dialogue. Fi- nally, we demonstrate how phrase filtering and vector quantization can reduce the size of our index by 4-10x, making dense phrase retrieval a practical and versatile solution in multi-granularity retrieval.2 1 Introduction Dense retrieval aims to retrieve relevant contexts from a large corpus, by learning dense representa- tions of queries and text segments. Recently, dense retrieval of passages (Lee et al., 2019; Karpukhin et al., 2020; Xiong et al., 2021) has been shown *This work was done when JL worked as a visiting re- search scholar at Princeton University. 1Following previous work (Seo et al., 2018, 2019), the term phrase denotes any contiguous text segment up to L words, which is not necessarily a linguistic phrase (see Section 2). 2Our code and models are available at https:// github.com/princeton-nlp/DensePhrases. ������ ���������� ��� ������� ������ ��� ��������� ������������������������������������������������ �������������������������������������������������� �������������������������������������� ������ ���������� ��� ������� ������ ��� ��������� ������������������������������������������������ �������������������������������������������������� �������������������������������������� ������������ �� �������� �������������� �������� �������������� ��� ������� ������� Figure 1: Comparison of passage representations from DPR (Karpukhin et al., 2020) and DensePhrases (Lee et al., 2021). Unlike using a single vector for each pas- sage, DensePhrases represents each passage with mul- tiple phrase vectors and the score of a passage can be computed by the maximum score of phrases within it. to outperform traditional sparse retrieval methods such as TF-IDF and BM25 in a range of knowledge- intensive NLP tasks (Petroni et al., 2021), includ- ing open-domain question answering (QA) (Chen et al., 2017), entity linking (Wu et al., 2020), and knowledge-grounded dialogue (Dinan et al., 2019). One natural design choice of these dense re- trieval methods is the retrieval unit. For instance, the dense passage retriever (DPR) (Karpukhin et al., 2020) encodes a fixed-size text block …

Content truncated to 500 words