Study On Semantic Matching And Sentence Reasoning For Machine Reading Comprehension

Machine reading comprehension is a fundamental and crucial topic in the field of natural language processing.Given a passage,the task requires a system to automatically answer questions based on the semantic information of the questions and options.With the rapid development of deep learning and the release of many large-scale datasets,deep learning based machine reading models are becoming more and more popular.Basically,most deep machine reading models follow the“encode-interact-predict" framework.Under this framework,the main contents of this research could be summarized as follows.First of all,this research proposes a multi-granularity co-reasoning model to tackle the problems of semantic matching and sentence reasoning in machine reading comprehension.A multiple granularity text matching module is introduced to model the interactions between passage,question and options.We make use of information extracted from diverse semantic spaces to conduct more extensive matching between text sequences,which could help to match the passage against the question and options more efficiently.Furthermore,in the information aggregating layer,a multi-sentence co-reasoning module is employed to perform sentence inference across multiple sentences.Specifically,1D Convolution Neural Network and self-attentive Recurrent Neural Network are utilized to model the relationships of relevant sentences.Experimental results demonstrate that our proposed model achieves state-of-the-art performance for single models on the RACE dataset.Then,focusing on how to effectively aggregate the information from multiple evidence clues,this research proposes an option attentive capsule network.Instead of traditional pooling operation,we regard the information aggregating as a routing problem and incorporate a capsule network to iteratively aggregate the evidence clues,and dynamically refine the matching representation vectors.Since traditional dynamic routing algorithm does not consider the semantic information of each option,we design an option attention-based routing policy to focus more on each option when clustering the features of low-level capsules.Experimental results demonstrate that our proposed capsule aggregating layer could bring relatively large improvement and it could better handle the complex questions which require sentence reasoning.