Question-Answer Sentence Graph for Joint Modeling Answer Selection

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 968–979 May 2-6, 2023 ©2023 Association for Computational Linguistics Question-Answer Sentence Graph for Joint Modeling Answer Selection Roshni G. Iyer1∗, Thuy Vu2, Alessandro Moschitti2, and Yizhou Sun1 1University of California, Los Angeles; Los Angeles, CA, USA 2Amazon Alexa AI; Manhattan Beach, CA, USA {roshnigiyer, yzsun}@cs.ucla.edu, {thuyvu, amosch}@amazon.com Abstract This research studies graph-based approaches for Answer Sentence Selection (AS2), an es- sential component for retrieval-based Question Answering (QA) systems. During offline learn- ing, our model constructs a small-scale rele- vant training graph per question in an unsuper- vised manner, and integrates with Graph Neu- ral Networks. Graph nodes are question sen- tence to answer sentence pairs. We train and integrate state-of-the-art (SOTA) models for computing scores between question-question, question-answer, and answer-answer pairs, and use thresholding on relevance scores for cre- ating graph edges. Online inference is then performed to solve the AS2 task on unseen queries. Experiments on two well-known aca- demic benchmarks and a real-world dataset show that our approach consistently outper- forms SOTA QA baseline models. 1 Introduction Automated Question Answering (QA) research has received renewed attention thanks to diffusion of Virtual Assistants. For example, Google Home, Siri, and Alexa provide general information in- quiry services, while many others serve customer requests in very different application domains. Two main tasks have been widely studied: (i) Answer Sentence Selection (AS2), which, given a question and set of answer-sentence candidates, consists of selecting sentences (e.g., retrieved by a search en- gine) that correctly answer the question; and (ii) machine reading (MR), e.g., (Chen et al., 2017), which, given a question and reference text, involves finding an exact text span that answers the question. AS2 models can more efficiently target large text databases (as they originated from the TREC- QA track (Voorhees and Tice, 1999)) and there is evidence that they are currently used in personal assistants, e.g., Alexa (Matsubara et al., 2020). ∗ This work was done while the author was an intern at Amazon Alexa AI. q: Who won the 1967 NBA Championship ? score label c1: The 1967 NBA World Championship Series was the championship series of the 1966-67 National Basketball Association season and was the conclusion of the 1967 NBA Playoffs. 0.810 0 c2: This was the first championship series in 11 years without the Boston Celtics, who were defeated in the Division Finals by Philadelphia. 0.048 0 c3: The 76ers won the series over the Warriors 4-2. 0.142 1 q1: Who won the 2009 Super Bowl ? a1: The Steelers defeated the Cardinals by the score of 27–23. q2: Who won Fifa World Cup 2010 ? a2: In the final, Spain, the European champions, defeated third- time finalists the Netherlands 1–0 after extra time. q3: Who won the most NBA championships ? a3: Bill Russell won 11 championships with the Boston Celtics. Table 1: An example from WikiQA, for candidates ranked by TANDA with normalized scores and labels. Garg et al. …

Content truncated to 500 words