Single Sequence Prediction over Reasoning Graphs for Multi-hop QA

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 11466–11481 July 9-14, 2023 ©2023 Association for Computational Linguistics Single Sequence Prediction over Reasoning Graphs for Multi-hop QA Gowtham Ramesh∗, Makesh Sreedhar∗, and Junjie Hu University of Wisconsin-Madison {gramesh4,msreedhar,junjie.hu}@wisc.edu Abstract Recent generative approaches for multi-hop question answering (QA) utilize the fusion-in- decoder method (Izacard and Grave, 2021) to generate a single sequence output which in- cludes both a final answer and a reasoning path taken to arrive at that answer, such as passage titles and key facts from those passages. While such models can lead to better interpretability and high quantitative scores, they often have difficulty accurately identifying the passages corresponding to key entities in the context, re- sulting in incorrect passage hops and a lack of faithfulness in the reasoning path. To ad- dress this, we propose a single-sequence pre- diction method over a local reasoning graph (SEQGRAPH)1 that integrates a graph structure connecting key entities in each context passage to relevant subsequent passages for each ques- tion. We use a graph neural network to encode this graph structure and fuse the resulting rep- resentations into the entity representations of the model. Our experiments show significant improvements in answer exact-match/F1 scores and faithfulness of grounding in the reasoning path on the HotpotQA dataset and achieve state- of-the-art numbers on the Musique dataset with only up to a 4% increase in model parameters. 1 Introduction Multi-hop Question Answering (QA) involves rea- soning over multiple passages and understanding the relationships between those pieces of informa- tion to answer a question. Compared with single- hop QA, which often extracts answers from a single passage, multi-hop QA is more challenging as it requires a model to determine the relevant facts from multiple passages and connect those facts for reasoning to infer the final answer. To tackle multi-hop QA, recent works have inves- tigated large pretrained generative models (Lewis ∗Equal contribution 1Code/Models will be released at https://github.com/ gowtham1997/SeqGraph "An American Werewolf in Paris was a partial sequel to the comedy film starring whom?" An American Werewolf in Paris [f1] It follows the general concept of, and is a loose sequel to, John Landis' 1981 film "An American Werewolf in London". David Naughton #1 David Naughton #2 An American Werewolf in London [f1] An American Werewolf in London is a 1981 horror comedy film. [f2] It was written and directed by John Landis and starring David Naughton, Jenny Agutter and Griffin Dunne. [f1] Naughton first became widely known as a result of his 4-year stint (1977–1981) singing and using appearances to promote Dr Pepper. [f1] David Walsh Naughton (born February 13, 1951) is an American actor and singer known for his starring roles in the 1981 horror film, " An American Werewolf in London", and the 1980 Disney comedy, "Midnight Madness". Question: Reasoning Path: [title-1] An American Werewolf in Paris [facts-1] [f1] [title-2] An American Werewolf in London [facts-2] [f2] [answer] David Naughton, Jenny Agutter and Griffin Dunne …

Content truncated to 500 words