# Dense X Retrieval: What Retrieval Granularity Should We Use?

**Authors:** Tong Chen, Hongwei Wang, Sihao Chen, Wenhao Yu, Kaixin Ma, Xinran Zhao, Hongming Zhang, Dong Yu  
**arXiv:** [2312.06648](https://arxiv.org/abs/2312.06648)  
**Date:** 2023-12-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.06648v2/

## Key points

- Proposition retrieval outperforms passage or sentence-based methods in both passage retrieval accuracy and downstream QA performance
- FACTOIDWIKI is an indexed version of the English Wikipedia dump, segmented into 250 million propositions
- Proposition retrieval demonstrates higher success rates compared to sentence and passage retrieval methods within the range of 100-200 words
- Passage-level retrieval struggles with entity ambiguity and references, while sentence-based retrieval lacks context
- Proposition-based retrieval ensures each unit contains one factoid and necessary context
- Proposition-based retrieval faces challenges with questions involving multi-hop reasoning over long-range textual analysis
- Recent works on dense retrievers typically adopt a dual-encoder architecture to improve retrieval performance
- This study provides insights into the impact of different types of retrievals on dense retrievals and downstream QA tasks

## Summary

In the conclusion of their study, the authors propose the use of propositions as retrieval units for indexing a corpus in order to improve dense retrieval performance. They conducted experiments on five open-domain question answering (QA) datasets with six different dense retrievers and found that retrieval by proposition outperforms passage or sentence-based methods in both passage retrieval accuracy and downstream QA performance. They also introduced FACTOIDWIKI, an indexed version of the English Wikipedia dump, where text from 6 million pages is segmented into 250 million propositions. The authors hope that FACTOIDWIKI, along with their findings, will facilitate future research on information retrieval. The results showed that proposition retrieval demonstrates a higher success rate compared to sentence and passage retrieval methods for a fixed retrieved word budget. The most significant improvement of proposition retrieval over passage retrieval occurs within the range of 100-200 words. As the word count further increases, the recall rates of all three granularities converge since all question-relevant information is included in the retrieved text. In an error case study, it was observed that passage-level retrieval struggles with entity ambiguity and references while sentence-based retrieval faces challenges due to lack of context. On the other hand, proposition-based retrieval tackles these problems by ensuring each unit contains one factoid and necessary context is incorporated in propositions. However, proposition-based retrieval faces challenges with questions involving multi-hop reasoning over long-range textual analysis. The related work section highlights recent works on dense retrievers that typically adopt a dual-encoder architecture which has been used to improve retrieval performance in various tasks. Overall, this study provides insights into the impact of choice of different types of retrievals on dense retrievals and downstream QA tasks demonstrating promising results for using propositions as a unit for retrieving information which opens avenues for further research in information retrieval.