Research On Reasoning Machine Of Reading Comprehension Model With Multiple Choice

Machine Reading Comprehension is a long-term task in the field of Natural Language Processing,which aims to help the machine automatically read and fully understand human natural language.This task requires the machine to answer context-specific questions to evaluate its understanding of the semantics and logic of the text.At present,the research work on Machine Reading Comprehension mainly focuses on the coded representation and logical interaction of different texts,of which the difficult problem is the deep understanding and logical reasoning of text semantics.At the same time,logical reasoning ability,as an important part of human intelligence,is often used to evaluate whether a machine can think like a human being.The fact that machines are capable of logical reasoning can help humans truly achieve the goal of Artificial Intelligence,which has great research value.However,the current mainstream machine reading comprehension technologies are difficult to deal with complex logical reasoning problems,and the latest machine reading comprehension models are seriously declining on machine reading comprehension datasets that require logical reasoning ability.Therefore,this paper focuses on the general characteristics of multiple-choice reasoning machine reading comprehension tasks.The multiple-choice reasoning machine reading comprehension models are studied from the perspective of improving the accuracy of mainstream machine reading comprehension models in dealing with logical reasoning problems.The research contents and innovations of this paper are as follows:(1)Research on Reasoning Machine Reading Comprehension Model Based on Multigranularity Attention Matching Mechanism.Considering the effects of attention mechanism and the difference information between answer candidates on the logical reasoning ability of the machine reading comprehension model,a reasoning machine reading comprehension model with multigranularity attention matching mechanism is proposed.The distributed word vector representation method is used to process the text content into a mixed form of vectors with various granularities,which is convenient for machine recognition and processing.Second,the difference information between answer candidates is encoded into valid reasoning evidences.Finally,the attention mechanism is used to highlight the key information in the text in order to improve the logical reasoning and anti-interference ability of the model.The experimental results of the model on the target dataset demonstrate the effectiveness and feasibility of the method,and the ablation experimental results demonstrate that the option comparison network module in the model has the greatest effect on improving the overall performance of the model.The main innovation of the model is the proposal of the answer candidate comparison mechanism,which encodes the differences between answer candidates into feature vectors and splices them into the vector representation of the original answer candidates.This mechanism not only enriches the reasoning evidences that the model can receive,but also improves the overall performance of the model on multiple-choice reasoning machine reading comprehension tasks.(2)Research on Reasoning Machine Reading Comprehension Model Based on Pretraining Technology.Considering the effects of contextualized text representation model and external reasoning knowledge on the logical reasoning ability of the machine reading comprehension model,a reasoning machine reading comprehension model based on pre-training technology is proposed.First,the encoder module in the pre-trained language model on a largescale English text corpus is used to encode the input text,which helps the model deeply understand text semantics and logical dependencies.Then,the fully fine-tuned model and its training parameters on the source datasets are transferred to the target dataset for adaptive learning in a specific way,which helps the model absorb external reasoning knowledge and improves the robustness and generalization ability of the model.The experimental results of the model on the target dataset demonstrate the effectiveness and feasibility of the method,and the ablation experimental results demonstrate that the text encoding module in the model has the greatest effect on the overall performance of the model.The main innovation of the model is that the combination of the coding module of the pre-trained language model and the attention mechanism can help the model deeply understand natural language text which contains multiple logical dependencies.Also,this mechanism can improve the overall performance of the model on multiple-choice reasoning machine reading comprehension tasks.