Generation-based question decomposition for multi-hop question answering

Abstract

We target to improve multi-hop question answering (QA) by decomposing hard questions into simpler ones that can be answered by existing single-hop QA models. Instead of manually labeling question with decompositions or building decompositions from an external question dataset, we leverage an existing question generation (QG) model to produce high-quality dataset for training a question decomposition model, in which the decompositions are semantically relevant to the questions. Then, we adapt the previous multi-hop QA architecture utilizing decomposed sub-questions, by answering sub-questions with an existing QA model and using corresponding supporting sentences to reformulate the multi-hop questions. Experiments on HOTPOTQA show that, with equivalent amount of training data, our newly-generated decompositions help to boost the performance by 3.1 points on Ans EM score. By combining with previous training data, our model outperforms all existing question-decomposition-based models, and shows competitive results with state-of-the-art graph-based approaches.