Enhance Machine Reading Comprehension on Multiple Sentence Questions with Gated and Dense Coreference Information

Abstract

Machine reading comprehension (MC) is one of the most important problems in natural language processing. Most of the previous works rely heavily on features engineering and handcrafting techniques. Since the release of SQuAD, a large-scale MC dataset, many deep learning models have been proposed. However, these models are limited by the soft attention mechanism only relied on keywords that appears in a question. Therefore, the performance is always poor in a question that needs to infer an answer from multiple sentences, which cannot depend on keywords in a question. In this paper, we propose a deep learning model that incorporates coreference information to improve the prediction performance especially on multiple sentence question. We also propose the bi-directional answering technique that can help the model avoid a local maxima of the single directional answering method in a traditional model. The results have shown that our approach outperforms the baseline in terms of F1 and Exact Match (EM).