Research And Implement Of Machine Reading Comprehension Based On Semantic Reasoning And Induction

Machine reading comprehension is one of the most challenging tasks in the field of natural language processing which aims to answer related questions by understanding unstructured text.Thanks to the development of deep learning technology and the emergence of large-scale machine reading comprehension datasets,the field of machine reading comprehension has developed rapidly in recent years,reaching or even surpassing human levels on many datasets.These traditional machine reading comprehension tasks are often "single-hop",which means you can get the answer to a question based on an article or a sentence in most cases.However,in reality,we need to reason between multiple documents to get the answer to the question from time to time.The "multi-hop" reading comprehension task poses a greater challenge to semantic reasoning and induction capabilities of the model.In this paper,we focus on the task of the multi-hop reading comprehension,different models are studied in order to better summarize and reason on the original text.We use the WIKI HOP dataset.The main work of the paper is as follows:(1)Study the construction method of multi-hop reading comprehension dataset.Inspired by the WIKIHOP dataset,we construct a data collection and processing pipeline,analyze the collected dataset,and finally perform benchmark experiments on the dataset;(2)Study the multi-hop reading comprehension model based on graph convolutional network.First,a basic model GCN_Base which is based on R-GCN is constructed,and on this basis,we make further improvements which include optimizing the entity graph,adding a double attention layer and adding a double scoring mechanism to get the improved model GCN_Enhance.Experiments show that GCN_Base has a 15%accuracy improvement compared to the benchmark model,and GCN_Enhance has a 3.3%accuracy improvement compared with GCN_Base;(3)Study the context-enhanced multi-hop reading comprehension model.We propose a context-enhanced model CEG to overcome the shortcomings of the graph models.CEG uses soft extraction mechanism and hard extraction mechanism to make context enhancement.Soft extraction embeds contextual representations in entity coding,mainly by introducing the pre-trained language model BERT.Hard extraction directly extracts the context content around the entity and uses the memory network to utilize the content in the graph inference stage.The accuracy of CEG relative to GCN_Enhance is increased by 5.7%,and further ablation experiments have verified the effectiveness of each module of the model.The research in this paper focuses on multi-hop reading comprehension,explores the use of graph models in multi-hop reading comprehension tasks,and further proposes context enhancement methods to overcome the shortcomings of graph models.Experiments show that our proposed method can effectively improve the model's induction and reasoning capabilities.