Keywords-aware dynamic graph neural network for multi-hop reading comprehension

Abstract

The multi-hop reading comprehension (RC) is challenging for machine reading comprehension. It is crucial for multi-hop RC to comprehend complex questions and contents between multiple paragraphs. In this paper, we propose a strategy of keywords-aware dynamic graph neural network (KA-DGN) to improve the performance of multi-hop reading comprehension. First of all, KA-DGN focuses on the salient information in the text, extracts keywords from the question and context. A window is specifically designed to frame the interrogative pronoun/adverb and its nearby words in the question, which encourages the model to focus on the answer. Next, the token-level answer span is predicted under the guidance of the keywords. And the boundary loss function is also invented to enhance the boundary awareness of the model on extracting the answer, which maximizes the probability of answer span bound while minimizing that of the noise. Finally, the model builds a dynamic reasoning graph combining explicit keywords and implicit semantic information among sentences. Graph neural network is applied to predict the sentence-level supporting facts. While evaluating on HotpotQA, the proposed KA-DGN achieves competitive performance in distractor setting.