Term Similarity-aware Extensive and Intensive Reading For Multiple Choice Question Answering

Abstract

Multiple Choice Question Answering(MCQA) aims to automatically choose a correct answer from candidate options when given a passage and question. Existing approaches generally model attention mechanisms based on whole-passage information or manually tag key sentences for weakly supervised learning, which leads to the models focusing extensively on redundant information and costly manual annotation. In this paper, we consider evidence sentence extraction work in an unsupervised way to precisely pinpoint evidence sentences and minimize the impact of redundant information while avoiding costly manual annotations. Specifically, we propose a novel model called Term Similarity-aware Extensive and Intensive Reading(TS-EIR), which dynamically and automatically refines critical information by term similarity. In detail, it intelligently selects sentences more relevant to the question from the passage and deeply extracts features by enhanced graph convolutional neural network. We apply the proposed TS-EIR to a typical pre-trained language model, BERT, for encoding and evaluate it on the RACE and Dream benchmarks, which verify our model achieves substantial performance improvements over the current baseline.