GenDec: A Generative Question-decomposition method for Multi-hop Question-answering

Anonymous

Abstract: Multi-hop QA involves step-by-step reasoning to answer complex questions and find multiple relevant supporting facts. Previous question-decomposition research on multi-hop QA has shown that performance can be boosted by first decomposing questions into simpler, single-hop sub-questions (QD), and then answering them one by one in a specific order. However, such decomposition often leads to error propagation during QA: 1) incorrect QD leads to wrong QA results; 2) wrong answers to a previous sub-question compromise the next sub-question. In this work, we propose GenDec, a generative QD-based model for multi-hop QA from the perspective of explainable QA by generating independent and complete sub-questions based on incorporating supporting facts. This approach first introduces sub-questions in retrieving relevant passages at each hop and fuses features of sub-questions into QA reasoning, which enables it to provide an explainable reasoning process for its answers. We evaluate GenDec by comparing it with existing QD-based and other strong QA models and the results show GenDec outperforms all QD-based multi-hop QA models for answer spans on the HotpotQA and 2WikihopMultiHopQA datasets. We also conduct experiments with the large language models (LLMs) ChatGPT and LLaMA to illustrate the impact of QD on QA tasks in the LLM era.

Paper Type: long

Research Area: Question Answering

Contribution Types: NLP engineering experiment

Languages Studied: English

Consent To Share Submission Details: On behalf of all authors, we agree to the terms above to share our submission details.