Research On Multi-document Machine Reading Comprehension Method Based On Heterogeneous Graph Neural Network

Machine reading comprehension is the most basic and essential research problem in the field of natural language processing,and it is the basic theory and foundation task to accomplish Artificial Intelligence,which has broad practical application demands in real life.Increasingly,the theoretical research has been deeply studied,besides its form of the appliance has changed,then the machine reading comprehension is now facing the new challenge that reading and reasoning in the multi documents within more complex requirements.For the sake of full usage of both the unstructured textual information and structured knowledge bases,to adapt the incremental of the reading documents and reason across multi documents,this dissertation based on the graph neural networks has studied the construction of the heterogeneous reading graphs and researched the method of fusion the external knowledge in multi-document machine reading comprehension models,as follows.(1)This dissertation studies and proposes a heterogeneous reading and reasoning graph model,dubbed ClueReader,which imitates the research findings in neuroscience i.e.,the concept of “the grandmother cells”.Firstly,it puts forward a clue-type reading model that constructs the reading links from questions to the reasoning entities contained in documents,then to the latent answer candidates covered in the contexts.It aggregates the global information and sequentially reasons across multi documents.Secondly,this model conducts the links and informative selections in multi entities and relations by improving the spatial graph neural networks.Conversely,it creatively visualizes the inner status in the graphs before the answer prediction phase and illustrates the heuristic explanations,which further improves the analyzability and reliability.The experiment elucidates that the model improves the accuracy of answer prediction in multi-document machine reading comprehension.(2)Considering the lack of the entities’ attributes,not merely reading from the closed-domain documents,this dissertation studies and proposes a graph reading model fusing the external knowledge into the graph neural networks,named Med KGQA,which can supplement the missing natures of the entities in the texts of molecular biology.In the first place,it analyzes the essence of the task,i.e.,predicting the drugdrug interactions based on the machine reading comprehension method from the closeddomain texts,collects and constructs a knowledge base containing the “drug-protein target” triplets from external structured knowledge.Afterward,it applies to the method of knowledge base embeddings to obtain the drugs and proteins’ representations align to the node-level representations in the model’s graph.Besides,it makes use of the recorded biology pathways contained in the human body to connect the directed edges among drugs and their targeted proteins entities in the reading graph.These two steps introduce the knowledge into the model.In conclusion,this model’s performance on the reading task surpasses the other models’,and illustrates its rationality and effectiveness,which further shows the feasibility of integrating external knowledge in reading comprehension tasks.In addition,it also could provide references to the other closed-domain reading comprehension in terms of architecture designing.In the final analysis,basing on graph neural networks and combining with the deep learning technologies,this dissertation proposes two models,i.e.,ClueReader and Med KGQA,respectively aim to the specific tasks in open-domain and closed-domain machine reading comprehension,and they effectively improve the accuracy of answer prediction,which have certain theoretical significance for the further study of graphbased reading comprehension methodologies.Moreover,it provides some demonstration effect for promoting the cross-domain application of artificial intelligence.