Research On Question And Answer Retrieval Based On Generative Adversarial Network

With the popularity of the Internet and the rapid development of big data,the sources of information available for learning on the Internet are becoming more and more abundant,everyone can quickly and easily obtain the information they want through search engines.And because of the large amount of data available for selection,search engines need to have strong algorithmic support to match the information that users really need.However,the existing search engines still have many shortcomings,which are mainly divided into the following two aspects: Firstly,they return too many results,making it difficult for users to quickly and accurately find the information that best meets their needs;secondly,the technical basis of the search engine,that is,the key Word matching mainly focuses on the grammatical form of language,but pay less attention to semantics.At the same time,because the ability of users to express individual needs is uneven,it is difficult to accurately express information requirements by using simple query words,so that the retrieval effect is general.In addition to search engines,a question and answer retrieval system can be used to meet the information needs of users.Different from the traditional search engine,the question and answer retrieval system can not only use natural language sentences to ask questions,but also directly return the best answer to the user according to the query result returned by the model,instead of the related web page.The question-and-answer retrieval system finds the best answer by sorting the questions and answers by relevance matching.Therefore,selecting the appropriate training data is the first step in the training question-and-answer retrieval model.Based on this,in the semantic matching of this paper,the learning of the input text should be completed first,and then the semantic similarity between the question and the answer should be calculated.This paper first analyzes the research purpose and significance of the question-andanswer retrieval system,and elaborates the research status in the field of question-andanswer retrieval at home and abroad,including information retrieval,question-answering system,deep semantic matching,etc.,laying a theoretical foundation for the development of the model..Then introduces the techniques and methods used in the research,mainly the method of deep semantic matching and generating the anti-network(GAN)method.Then the QAGAN model is proposed,and the GAN model-based method is used to semantically match the problem and answer of the qualified domain.The purpose is not only to find the correct answer from the candidate answer set to present to the user,but also to improve the model to recognize the correct answer during the continuous training process.Ability.In the application of the model,for the limited domain question and answer retrieval task,select the insurance domain corpus,first sample the positive question and answer pair from the training set containing the correct question and answer pair,and select the correct answer for each group.In addition to the n interfering answers,the original sentence input of the three items is separately learned to obtain the corresponding sentence vector,and the similarity score is calculated and sorted by the method of finding the cosine value,and the generated model generates an answer pair that approximates the order of the real answer.The discriminant model needs to distinguish between the real answer pair and the generated answer pair.Finally,the results identified by the algorithm are compared with the data labels to verify the validity of the method.The innovations of this paper are mainly reflected in the following two aspects:1.Use the deep semantic matching model for the question and answer retrieval system to effectively reduce the data dimension.The deep semantic matching model has a certain degree of development in the field of information retrieval,but it is rarely used in short text tasks such as question answering systems.The best answer is selected from the answer candidate set by performing deep semantic matching on the candidate set of questions and answers one by one.2.Generate a confrontational search task in the field of natural language processing.Previously,some scholars have tried to combine GAN with NLP,but most of them are difficult to achieve satisfactory results.This time,GAN is applied to the question and answer system,combined with the generation model and the confrontation model in GAN,and the game theory method is used to iteratively optimize the two models.On the one hand,a discriminant model designed to mine valid signals from both labeled and unlabeled data provides guidance for training the generated model to accommodate the correlation distribution implicit in the documentation for a given query.On the other hand,the generation model generates a question-and-answer pair that is difficult to distinguish for the discriminant model by minimizing its discriminative target in a confrontational manner.