Interpretable Multi-hop Question Answering

Helping computers learn to think or reason is the long-standing and challenging task of Artificial Intelligence. Question Answering (QA), an advanced form of information retrieval in the field of Natural Language Processing, aims to build a system to answer natural language questions posed by humans. Since QA tasks can be used to quantify the understanding and reasoning ability of intelligent systems, a large number of QA datasets have been developed in recent years, and significant progress has been made in this direction. However, most existing QA systems use an end-to-end model to answer natural language questions, making it difficult to demonstrate that a model has performed the desired reasoning to predict the correct answer. Especially for multi-hop QA, questions often permit reasoning shortcuts by which QA systems can directly locate the final answer by matching the question to a single sentence in the paragraph without complex reasoning. As a result, most QA systems have fallen short of the goal of teaching computers to answer multi-hop questions in an interpretable fashion using an acceptable reasoning pattern. This thesis attempts to tackle two distinct problems: i) how to build an interpretable QA system to answer a complex multi-hop question; ii) how to learn knowledge for the new QA tasks while retaining the knowledge learned on previous QA tasks. In this thesis, we focus on multi-hop QA tasks which require reasoning over multiple paragraphs and providing explanations for answers. Compared to traditional QA tasks, this setting requires a QA system not only to predict the final answer but also to provide the underlying reasoning chain, e.g., to make explainable predictions. This thesis describes several distinct approaches to tackling these problems in multi-hop QA. Since these approaches are designed to solve complex multi-hop QA tasks in an interpretable fashion, they offer the potential of helping humans to understand and trust the mechanism of the QA system. Therefore, all of the approaches described in this thesis are designed to help humans understand the performed multi-hop reasoning better and to help computers learn to think and reason better. Specifically, i) we describe a question decomposition approach based on abstract meaning representation (AMR) for multi-hop QA, which first transforms the task of decomposition of the multi-hop question into the task of segmentation of a corresponding AMR graph, and then decomposes a multi-hop question into simpler sub-questions and answers them in order; ii) we describe structured knowledge and contextual information fusion graph neural networks (GNNs) intended to mimic human step by step reasoning; these networks provide a new perspective on how to perform interpretable reasoning by GNN-based approaches and achieve explicit reasoning via asynchronous update of graph node representations using information from neighbours; iii) we describe prompt-based conservation learning for multi-hop QA, which acquires new knowledge from multi-hop QA tasks while conserving old knowledge learned on a previous component QA task, mitigating forgetting. Moreover, we also learn soft prompts to condition pre-trained language models to perform type-specific reasoning. This thesis explores several approaches to designing reliable and interpretable multi-hop QA systems, and demonstrates and evaluates these proposed QA systems over multiple open-domain multi-hop QA datasets collected from Wikipedia, with the aim of helping computers to think and reason about natural language text.

Name thesis:degree_name

PhD

Level thesis:degree_level

Doctoral

Discipline thesis:degree_discipline

Computer Science

Grantor dc:publisher

ResearchSpace@Auckland

Year dc:date.issued

2023

Author dc:creator

• Deng, Zhenyun

Advisors dc:contributor.advisor

• Witbrock, Michael
• Riddle, Patricia

Statement dc:rights

• Items in ResearchSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

Licence dc:rights.uri

• https://researchspace.auckland.ac.nz/docs/uoa-docs/rights.htm

Handle dc:identifier.uri

https://hdl.handle.net/2292/64102

OAI identifier oai:identifier

oai:researchspace.auckland.ac.nz:2292/64102