Research On Community Question-answer Matching Method Based On Reinforcement Learnin

Community Q&A matching is one of the important research directions of natural language processing.With the continuous development of Internet technology and the continuous popularity of mobile intelligent devices,people’s way of acquiring knowledge has gradually shifted to the community Q&A platform represented by Zhihu.Community Q&A is a public knowledge platform.It not only provides users with the services of question search,answer acquisition and information sharing,but also solves the problem that search engines cannot directly return answers.However,due to the different quality of the answers,it takes a lot of time for users to find the possible useful answers through a group of candidate answers,and question and answer matching can automatically select the correct answers in the candidate answers according to the given questions,maximize the user’s participation in the community question and answer,and reduce the time for users to find the correct answers.Sort the candidate answers based on their quality,and then perform question and answer matching based on the sorting results.The main content is as follows:(1)To address the problem of a single method for evaluating answer quality and introducing a large amount of noise from user information,a ranking method combining historical answer features and multi granularity semantic interaction is proposed to determine answer quality.In this paper,the pointer network method in reading comprehension task is used to extract the characteristics of users’ historical answers,and then the multi granularity semantic interaction method is used to model the semantic relationship between question and answer pairs and users’ historical answers from the local and global dimensions.Through local semantic interaction,the semantic association information of question and answer pairs and historical answers is extracted from the word level,and the weak correlation parts are eliminated with the dynamic attention mechanism,And the method of comparative aggregation pooling is used to extract local semantic interaction feature vectors of question answering pairs and historical answers.Extracting semantic association information between question answering pairs and historical answers at the sentence level through global semantic interaction.Finally,the local and global semantic interaction vectors of question and answer pairs and historical answers are fused,input into the classifier for scoring,and candidate answers are ranked based on their scores.The experimental results show that the proposed method effectively improves the semantic representation ability and sorting performance of the model.(2)Traditional question and answer matching methods usually judge the correctness by independently matching questions and individual candidate answers.However,the matching information between a question and a single candidate answer is usually limited,and only using the question as the only evidence to determine the correctness of the candidate answer can lead to the answer being misjudged as incorrect if the evidence is not intuitive enough.To address this issue,we propose a question and answer matching method based on reinforcement learning.It can continuously accumulate evidence of candidate answers,helping the model judge the correctness of the answers.Through the reward mechanism of reinforcement learning,agents are encouraged to learn a good evidence extraction mechanism,continuously accumulating evidence from candidate answers that are judged to be correct as the basis for subsequent Q&A matching.Match the candidate answer with the current question and evidence separately,obtain a matching representation,and use it as the state of the agent to perform the action of judging the correctness of the candidate answer.If the agent determines that the answer is correct,potential useful information is extracted to update the evidence,and feedback rewards are given based on evaluation indicators to improve the evidence extraction mechanism.In addition,extracting consensus related to the question from external answers to determine the credibility of candidate answers and using it as another basis for question and answer matching.The experimental results indicate that extracting evidence from candidate answers can effectively improve the performance of question and answer matching.