Research And Implementation Of Machine Reading Comprehension For Logical Reasoning

In recent years,technologies such as deep learning and natural language processing have been developing rapidly.As an important research direction in the field of natural language processing,machine reading comprehension plays an important role in search engines and question-answering systems by understanding articles and questions to make corresponding answers.For the model to understand the text,the text’s logical reasoning ability plays a crucial role.The model needs to explore the logical relationship between different sentences and entities.In order to improve the robustness of the reading comprehension system,most of the existing technologies use the pre-trained model as the encoder to understand the text and design the corresponding decoder to infer the text to get the corresponding answer.For logical reasoning questions with non-factual answers,the generative model is not limited to extracting answers directly from paragraphs but can generate more natural and complete statements as answers.However,the slow reasoning speed and long response time of the generative model hinder the industrial application of the model.At present,researchers also widely use model compression technologies such as pruning,quantification,and knowledge distillation to improve the reasoning speed of the model.To solve the problem of logical reasoning reading comprehension,this paper proposes a verb-driven dual-graph neural network.Verbs have the ability of context reasoning,and our network uses the reasoning ability of important verbs in sentences to construct the logical relationship between different sentences in the text.Compared with the previous methods,this paper makes up for the shortcoming that previous models based on graph neural networks do not explore the relationship between sentence level and word level at the same time.For the graph neural network algorithm proposed in this paper,two strategies are designed to construct the graph structure.First,use verbs to connect different sentences to explore the relationship between sentences,and then use dependency parsing to capture the clue chain in the sentence through the shortest dependency path.This paper constructs an article graph through the article and a question graph through the problem and carries out information dissemination respectively.In order to correctly predict the answer,the framework proposed in this paper integrates the information in the article graph and the question graph and applies a bi-direction attention mechanism to the graph data to predict the answer.This paper evaluates the method proposed on two public logical reasoning reading comprehension data sets ReClor and LogiQA.The experiment shows that the method proposed in this paper achieves excellent performance.At the same time,this paper implements a multi-task reading comprehension prototype to demonstrate the innovative algorithm proposed.This experiment uses a prototype system to infer and analyze real examples to demonstrate the effectiveness of reading comprehension models.Aiming at the problem of accelerating the reasoning of a generative model,although there are various model compression methods to compress BERT and its related variants,few people try to compress the generative model.In this paper,the difficulties of distilling the knowledge of the BERT model and compressing the generative model are studied from two questions.Because of the unique encoder-decoder architecture of the generative model and the characteristics of word-by-word answer generation,this paper analyzes the potential difficulties in the process of knowledge distillation and proposes a multi-task knowledge distillation framework.To solve the problem of word homogenization after knowledge distillation through contrastive learning distillation,a method of decoupling positive and negative label distribution is designed to make the student model better fit the output distribution of the teacher model,which can effectively transfer the knowledge in the teacher model to the student model,and greatly improve the reasoning speed of the student model.This paper evaluated the knowledge distillation method proposed in this paper on three public Chinese and English data sets,CNN DailyMail,LCSTS,and Quora.The experiment results show that the method proposed in this paper achieves better performance.