Research On Multi-hop Question Answering Based On Graph Neural Network

Machine reading comprehension is an important research direction in the field of natural language processing,which requires machines to answer questions by reading text.Previous machine reading comprehension models mostly focused on single text retrieval.These models can only answer questions based on a single paragraph,but in actual question answering scenarios,it is often necessary to reason across multiple paragraphs to obtain the answer of a question.In recent years,the multi-hop reasoning question answering task has been proposed to solve the above problems.Multi-hop reasoning question answering is a subtask of machine reading comprehension,which aims to find the answer of a given question across multiple paragraphs and has the ability of knowledge reasoning.At present,many models including DRN,QFE and DFGN have achieved certain results on multi-hop reasoning question answering tasks,but there are still some shortcomings.First,most existing models usually only obtain the answer by visiting the question once,so the model may not be able to obtain enough text information;Secondly,many models do not select the question-related paragraphs before performing knowledge reasoning,resulting in the amount of data that the reasoning model needs to process is too large and the time cost is too high.This paper conducted the following work for the above two problems:Firstly,a multi-hop reasoning question answering model based on paragraph filtering-reasoning pipeline mechanism is proposed.A dynamic reasoning network composed of dynamic graph attention and question reshaping mechanisms is established to reason across entity graphs,taking into account the mechanism that humans cannot pay attention to too much content at the same time,so as to solve the problem of model's insufficient understanding of query information.Experiments are conducted on public data sets,and the experimental results prove the effectiveness of the multi-hop reasoning answering model proposed in this paper;Secondly,an information retrieval model based on learning to rank is proposed.This paper uses the multi-head self-attention mechanism and the pairwise algorithm of learning to rank to consider the relative order among documents to improve the result of mean reciprocal rank.At the same time,this paper uses the model in the passage selecting stage of the multi-hop reasoning question answering task,removing paragraphs that are not related to the query as much as possible to reduce the scale of the data to be processed of the reasoning model.Experiments are carried out on public data sets,and the experimental results prove the effectiveness of the information retrieval model and multi-hop reasoning answering model combined with the latter proposed in this paper.