Conversational Machine Comprehension Based On Knowledge Reasoning

Machine reading comprehension is an important part of natural language processing.Its goal is to make the machine understand the content of articles and answer relevant questions by using algorithms,which is the embodiment of machine intelligence.In real life,humans communicate through conversation.Humans acquire additional information based on what they have learned through multiple rounds of conversational interaction.In order to simulate the ability of humans to answer current questions using historical information in conversation,conversational machine comprehension is proposed by researchers.Conversational machine comprehension adds conversational question and answer on the basis of traditional machine reading comprehension.There are often language phenomena such as topic transfer and referential substitution in the process of dialogue.Compared with traditional machine reading comprehension,conversational machine comprehension requires stronger logical reasoning ability of the model.In recent years,due to the wide application of deep learning model in conversational question answering task,conversational machine comprehension task has also been rapidly developed.However,there are still some problems in these models,such as ignoring conversational reasoning process,neglecting textual structure relations and semantic relations.In order to solve these problems,this thesis proposes a new conversational machine comprehension model based on knowledge reasoning.The model establishes connections between conversation through reasoning mechanism based on hierarchical information flow.At the same time,the reasoning mechanism based on dynamic conversation graph is used to construct the structural relationship between texts.The research contents are as follows:1.This thesis proposes a hierarchical conversation flow transition and reasoning model.In order to solve the problem that current machine reading comprehension models ignore the conversational reasoning process,this model uses the inference information of previous rounds to assist the reasoning process of current rounds,and makes full use of the potential semantic information in the process of sessions.2.This thesis proposes a dynamic conversation graph reasoning model.In view of the current machine reading comprehension model cannot build global semantic relations and to solve the problem of long distance reasoning,the model dynamically build problems in each conversation and the conversation history figure of reasoning,utilization of guiding the transmission process of information,can effectively capture the semantic structure information in the session and the session history information.3.This thesis develops online conversational machine comprehension question answering system prototype.In view of the lack of online reading comprehension question answering system in the market,this thesis integrates the above model algorithms to design and build an online reading comprehension question answering system.The system can carry out several rounds of word question and answer according to the user input articles and questions,and get a good conversation experience.