Research On Multi-jump Reading Comprehension Based On Multi-stage Reasonin

Machine Reading Comprehension(MRC)is a subfield of natural language processing aimed at training machines to answer questions based on given context.Multi-hop Machine Reading Comprehension(MHMRC)requires multiple jumps of inference across several related paragraphs to understand and answer complex questions,making it closer to human language reasoning abilities.In the process of solving multi-hop reading comprehension problems,capturing key paragraphs,sentences,and named entities in each document is crucial for forming reasoning chains in the human brain to ultimately find the answer,posing great difficulty and challenge for machines.In existing research,graph neural networks have achieved good performance in multi-hop reading comprehension tasks on relevant datasets.This is due to the excellent performance of graphstructured neural network methods in information propagation,feature information interaction,and relationship induction biases.However,existing models still suffer from insufficient information utilization,ambiguous reasoning paths,and redundant interaction information.This paper proposes a machine reading comprehension model based on multi-stage reasoning to address these issues.The main contributions of this paper are as follows:A two-stage document extraction model based on Prompt is proposed.This model addresses the overlook of the strength of the relationship between the combined information of various documents and the question when selecting documents in existing models.In order to enhance the ability of the model to extract relevant documents,this paper proposes a two-stage document extraction model based on Prompt.Through comparative experiments,it is verified that the twostage document extraction model based on Prompt can effectively improve the prediction results.A two-stage multi-hop reasoning model based on the relational graph convolutional network(R-GCN)is proposed.This model addresses the problem of fuzzy reasoning paths and redundant interaction information in existing models,and proposes a support sentence selection and answer prediction model based on relational graph convolution.By comparing with other advanced models in experiments,the effectiveness of the two-stage multi-hop reasoning model is verified.