A STEP towards Interpretable Multi-Hop Reasoning:Bridge Phrase Identification and Query Expansion

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 4552–4560 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 4552 A STEP towards Interpretable Multi-Hop Reasoning: Bridge Phrase Identification and Query Expansion Fan Luo, Mihai Surdeanu University of Arizona Tucson, AZ, USA fanluo@email.arizona.edu, msurdeanu@email.arizona.edu Abstract We propose an unsupervised method for the identification of bridge phrases in multi-hop question answering (QA). Our method constructs a graph of noun phrases from the question and the available context, and applies the Steiner tree algorithm to identify the minimal sub-graph that connects all question phrases. Nodes in the sub-graph that bridge loosely-connected or disjoint subsets of question phrases due to low-strength semantic relations are extracted as bridge phrases. The identified bridge phrases are then used to expand the query based on the initial question, helping in increasing the relevance of evidence that has little lexical overlap or semantic relation with the question. Through an evaluation on HotpotQA(Yang et al., 2018), a popular dataset for multi-hop QA, we show that our method yields: (a) improved evidence retrieval, (b) improved QA performance when using the retrieved sentences; and (c) effective and faithful explanations when answers are provided. 1. Introduction Multi-hop question answering (QA) requires synthe- sizing information across multiple documents or para- graphs to infer the correct answer. Information retrieval (IR) techniques are commonly used to narrow down the search space from millions of web pages to a small set of relevant documents or paragraphs. However, simply matching the semantics of the question and context is not sufficient to answer a multi-hop question. Consid- ering the example in Figure 1. For the question “What is the name of the executive producer of the film that has a score composed by Jerry Goldsmith?”, it is nec- essary to first find the film that ‘has a score composed by Jerry Goldsmith’. And then find out the ‘name of the executive producer’ of the film. Phrases such as ‘Alien’ that connect these supporting evidences in multi-hop QA are called bridge phrases. Identifying bridge phrases remains one of the challeng- ing problems for multi-hop QA, especially when ex- plainability is a concern. In this work, we propose an unsupervised method for bridge phrase identification, and show that it improves all downstream components in a multi-hop QA system. In particular, the contribu- tions of this paper are: (1) We introduce a new strategy, for the identification of bridge phrases for multi-hop QA by dynamically con- necting scattered information pieces from the question and the available context, organizing them as a noun phrases graph, modeling the bridge phrases identifica- tion task as a Steiner tree problem (Hartmanis, 1982) 1, and then applying Takahashi and others (1980)’s algo- rithm to identify the minimal sub-graph that connects all question phrases, in which the Steiner points are ex- tracted as bridge phrases. We call our method STEP (Steiner Tree-based Expansion Phrase identification). 1https://en.wikipedia.org/wiki/ Steiner_tree_problem Question: What is the name of the executive producer of the film …

Content truncated to 500 words