Diverse Multi-Answer Retrieval with Determinantal Point Processes

Proceedings of the 29th International Conference on Computational Linguistics, pages 2220–2225 October 12–17, 2022. 2220 Diverse Multi-Answer Retrieval with Determinantal Point Processes Poojitha Nandigam∗, Nikhil Rayaprolu∗, Manish Shrivastava Language Technologies Research Centre (LTRC) International Institute of Information Technology, Hyderabad, India poojitha.nandigam@research.iiit.ac.in, nikhil.rayaprolu@students.iiit.ac.in, m.shrivastava@iiit.ac.in Abstract Often questions provided to open-domain ques- tion answering systems are ambiguous. Tradi- tional QA systems that provide a single answer are incapable of answering ambiguous ques- tions since the question may be interpreted in several ways and may have multiple distinct an- swers. In this paper, we address multi-answer retrieval which entails retrieving passages that can capture majority of the diverse answers to the question. We propose a re-ranking based approach using Determinantal point processes utilizing BERT as kernels. Our method jointly considers query-passage relevance and passage- passage correlation to retrieve passages that are both query-relevant and diverse. Results demonstrate that our re-ranking technique out- performs state-of-the-art method on the Am- bigQA dataset. 1 Introduction The objective of open-domain question answering is to provide answers to queries utilising a large collection of documents from the World Wide Web, Wikipedia etc. More than 50% of questions present in a widely used open-domain QA dataset (Natural Questions Kwiatkowski et al. (2019)) comprise of ambiguous questions (Min et al. (2020)). Questions that are ambiguous may be interpreted in a number of ways and as a result, they need various answers. In this paper, we concentrate on questions with multiple distinct answers. Open domain question-answering systems are designed to generate answers from several data sources. Since similar information can be present across multiple data sources, it introduces a sig- nificant amount of redundancy. Traditional open- domain QA (Chen et al. (2017)) systems comprise of a Retriever, which retrieves passages relevant to the question. A passage retriever is primarily con- cerned with retrieving passages that are relevant *Equal contribution to the query, and it does not address redundancy in the passages during retrieval. To be able to pro- duce diverse answers to the question, the passages retrieved must be both relevant to the question and distinct from one another. After the retrieval stage, we introduce a novel re-ranking approach to handle redundant passages. As a result, the re-ranked pas- sages would capture most of the diverse answers to the question. In this paper, we investigate the multi-answer retrieval task, which entails retrieving passages that can cover the distinct answers. Re-ranking methods have been employed pre- viously to improve the question answering accu- racy significantly(Wang et al. (2019);Nogueira and Cho (2019); Min et al. (2021); Clark and Gardner (2017)). Min et al. (2021) tackles diverse multi- answer retrieval by proposing a re-ranker based on an auto-regressive framework in which each pas- sage selected is dependent on the passages chosen at a previous time step. Determinantal Point Processes (DPP) (Kulesza and Taskar (2012)) are probabilistic models that are effective at identifying diverse subsets of ele- ments from a collection while preserving quality. DPP methods have proven effective in natural lan- guage processing tasks …

Content truncated to 500 words