Research On The Sorting Method Of Cross-modal Candidate Answers In Community Question And Answer

With the development of the Chinese Q&A(question and answer)community,like ZhiHu and Baidu knows that there are a huge number of Q&A pairs in the Chinese Q&A community.When a user asks a question,first of all,search for a question similar to the question in the Q&A community.If the question does not appear in community,then the new question is posted.If a similar question has been retrieved,then return the Q&A pair of similar questions to the user.With the development of the information age,when the people answer a questions,they are not just limited to using text to describe them,And use pictures,audio and video to further explain or prove,in a more intuitive way,thus forming a form of multi-modal data complementing each other,so the answers in the current Chinese Q&A community mostly include text,pictures and audio and video.There are still a large number of other multimodal data problems and answers in the current Q&A community that have not been rationally utilized.Among them,the association between multi-modal data such as text,picture,video,audio,etc.,does not consider the characteristics of other multi-modal information when sorting answers,so that some high-quality answers are not effectively arranged.How to match the questions raised by the questioners to more reasonable answers and improve the efficiency of the Q&A community users to obtain high-quality answers is one of the key issues that the Chinese Q&A community needs to solve.This paper mainly studies the expressions of multimodal combination in the current Chinese Q&A community.In addition to text representation,it also contains other modal information expressions such as pictures,videos and audio.By extracting the textual features of the question,the textual features of the answer,and the characteristics of other multimodal information(pictures as examples)in the answer,Reanalyze the correlation between text features and image features.Through the correlation analysis between different modal data,this paper studies how to use the multimodal information in the answer to adjust the score of the answer,and proposes a cross-modal answer ranking method for the question and answer community,from the perspective of crossmodality.The sorting problem of the answer returns to the user a more reasonable sort of answers,returning the high quality answer to the user more efficiently.The main research contents of this paper are as follows:(1)Relevant question retrieval.When the user asks a new question,it is necessary to calculate the similarity between the problem and the problem,and return the Q&A pair to the user with most similar questions.In this paper,a question retrieval method based on Word2 vec word vector and GRU neural network for question coding is used.And compared with the LDA method,verify the validity of the relevance query method proposed in this paper;(2)The construction of a cross-modal retrieval model of Q&A.First,calculate the degree of association between the multi-modal information(pictures as examples)in the question and answer,and then calculate the degree of association between the text in the answer and other multimodal information(pictures as examples).The LDA method is used to extract the text features in the Q&A,the SIFT method is used to extract the image features in the answers,and the correlation analysis is performed by the Canonical Correlation Analysis(CCA)method to calculate the other multimodalities in the problem texts and answers.The degree of relevance of the information and the relevance of the answer text to other multimodal information in the answer.Mapping the text feature vector set and the image feature vector set to the same maximum subspace,and when the user gives a question query,find the distance between the feature projection of the question text and the image feature projection in the answer.The image with the smallest distance value is used as the search image that best matches the feature text feature.Similarly,the image that best matches the answer text can be obtained,and the cross-modal retrieval of text and image can be realized,thereby constructing a cross-modal retrieval model for text and image.(3)An answer ranking method based on cross-modal retrieval.By calculating the similarity of the problem,this paper solves a series of Q&A that are similar to the problem in the Q&A library of the Q&A community when the user asks a new question.Then from the Q&A pairs of these similar questions,find the multimodal information with high degree of relevance to the problem.Then look at the relevance of the text in the answer to the multimodal information in the answer,and give these associations a reasonable score on the importance of answering the question.Finally,use the problem similarity score,the multi-modal information in the answer and the relevance score of the text in the answer and the multi-modal information in the answer and the relevance score of the question are used to evaluate the quality of the answer and obtain the sort result.In addition,in order to further improve the effect of the sorting method,this paper combines the answer sorting method based on cross-modal search and the Ranking SVM answer sorting method,and proposes the answer sorting method combining the two methods,and the data set in the ZHIHU community.In summary,in the research of this paper,the multi-modal data in the Q&A community is considered,the correlation between different modal data is explored,and the cross-modal retrieval model of the text and image in the question and answer is been constructed.On this basis,the crossmodal sorting method of the question answer community candidate answer and the Ranking SVM sorting method are compared to prove the effectiveness of the sorting method proposed in this paper.