Research On Multi-hop Reasoning Machine Reading Comprehension Based On Graph Convolutional Neural Network

In recent years,the research on Machine Reading Comprehension(MRC)has continued to be hot.The task of machine reading comprehension can be regarded as a text-based question and answer system.Given a context and question,the machine can get the answer.The MRC task measures the machine.The ability to understand human natural language and has high application value in real life.Multi-hop reasoning machine reading comprehension In the traditional reading comprehension method that focuses on a single article,more emphasis is placed on the reasoning ability of the machine,aiming to require the system to combine multiple reasoning facts of multiple documents to make inferences,and then to obtain the final answer.We propose an effective multi-hop reasoning machine reading comprehension method based on graph convolutional neural network,which is based on the context entity index extracted by text feature extraction and Self-Attention.The entity index is obtained by string matching.In this way we get the entity representation of the answer in the text,and each entity can fully integrates the context and the information of the internal characters of the entity.When using graph convolutional neural networks for multi-hop reasoning,we designed two graph construction rules for different types of data sets,constructed a single entity graph in wikihop data,and constructed two types of entities and sentences in Hotpot QA.Heterogeneous graphs of node types to strengthen the interaction between entities and sentences,which reflects the strong supervision effect of supporting sentences on answer prediction.In order to combine the query information,we co-attention the graph node representation and the query representation after the information fusion of the graph convolutional neural network to generate query-aware entity nodes,thus reflecting the guiding role of the query in model prediction.In addition,we hope to convert the segment extraction task of Hotpot QA data into a text classification task,so we designed a multi-answer candidate generator based on a single-answer reading comprehension model,and also used the Bert classification task to generate supporting sentence candidates.Experiments show that the method in this paper has advantages in the support sentence tasks of the unmasked version of the wikihop dataset and the Hotpot QA dataset.In the ablation analysis of this article,the mutual attention between the question and the entity,and the interaction between the supporting sentence and the answer have a significant impact on the model.