Four-way Bidirectional Attention for Multiple-choice Reading Comprehension

As one of the crucial tasks of natural language processing, machine reading comprehension has gained increased attention in recent years. In this paper, we propose a four-way bidirectional attention network for a multiple-choice reading comprehension task, where every question comes with a set of candidate options and only one correct answer. Current methods on such tasks usually judge options independently and ignore their relations. Thus, this work designs a four-way bidirectional attention strategy to formulate the interactions among the passage, questions and candidate options. In particular, the relations among options are well represented. This enables the model to leverage the option correlation information for inferring the final answer accurately. The experimental evaluations on the CosmosQA dataset demonstrate the competitive performance of our model, and confirm the effectiveness of the option comparison strategy.