Diverse and Non-redundant Answer Set Extraction on Community QA based on DPPs

Proceedings of the 28th International Conference on Computational Linguistics, pages 5309–5320 Barcelona, Spain (Online), December 8-13, 2020 5309 Diverse and Non-redundant Answer Set Extraction on Community QA based on DPPs Shogo Fujita1∗, Tomohide Shibata2 and Manabu Okumura1 1Tokyo Institute of Technology 2Yahoo Japan Corporation {fujisyo,oku}@lr.pi.titech.ac.jp tomshiba@yahoo-corp.jp Abstract In community-based question answering (CQA) platforms, it takes time for a user to get useful information from among many answers. Although one solution is an answer ranking method, the user still needs to read through the top-ranked answers carefully. This paper proposes a new task of selecting a diverse and non-redundant answer set rather than ranking the answers. Our method is based on determinantal point processes (DPPs), and it calculates the answer importance and similarity between answers by using BERT. We built a dataset focusing on a Japanese CQA site, and the experiments on this dataset demonstrated that the proposed method outperformed several baseline methods. 1 Introduction Community-based question answering (CQA) platforms have been used by many people to solve prob- lems that cannot be solved simply by searching in books or web pages. On a CQA platform, the ques- tioner posts a question and waits for others to post answers. The posted answers are viewed by not only the questioner herself but also other users. A single question may receive several to dozens, even hundreds, of answers, and it would take a lot of time to see all the answers. The questioner chooses the most helpful answer, called the best answer, and users can view this answer as a representative answer. However, there may be more than one appropriate answer to a question that asks for reasons, advice, and opinions, while there is often only one appropriate answer for a question that asks for facts, etc. For example, Table 1 shows example answers to the question “Why do birds turn their heads backwards when they sleep?”, which asks for reasons. The best answer A1 says it is to keep out the cold and A4 says it is for balance, which is also an appropriate answer. In such a case, if users view only the best answer, they miss other appropriate answers. One possible way to deal with this problem is to use ranking methods (Wang et al., 2009; Xue et al., 2008; Bian et al., 2008; Zhao et al., 2017; Tay et al., 2017; Zhang et al., 2017; Laskar et al., 2020). Since the studies use the best answers as training data, similar answers are ranked higher although the existing methods try to reduce redundancy by using the similarity between a question and an answer or between answers. As a result, users need to read through carefully the answers that are ranked higher to obtain the information they want to know. We propose a new task of selecting a diverse and non-redundant answer set from all the answers, instead of ranking the answers. We treat the set of answers as the input, from which the system determines both the appropriate number of …

Content truncated to 500 words