Research On Question Response Time Prediction Of Community Question Answering Based On Quality-Analysis

The Community Question Answering(CQA)has emerged as an extremely popular interaction platform for knowledge acquisition and sharing.CQA provides different services to improve the efficiency of knowledge acquisition,one of which is the service of question response time prediction.It aims to analyze the question to further predict the time when the question is answered so as to help users manage time scientifically and promote user experience in the CQA.The existing methods for predicting the response time of a question tend to estimate the time when the question is first answered by constructing the question and the answerer model.However,the answers provided by different users are often with different quality and usability.Therefore,the first answer to a question may not be able to effectively solve users' doubts,resulting in the response time prediction service aimed to estimate the arrival time of first answer may not truly meet needs of users to some extent.To address this problem,we propose a novel method named Question-Answerer Matching Model with High Quality for Question Response Time Prediction for predicting the time when high-level users give high-quality answers,which improves the validity of prediction.The main research contents and innovations of this article are as follows:(1)High-Quality Question Retrieval based on Coupling of LC-LDA and BERT(HQLB-QR)is proposed for question retrieval.This method couples the LCLDA(Label Cluster Latent Dirichlet Allocation)with the BERT(Bidirectional Encoder Representation from Transformers)and extracts the topic-level and word-grained semantics of the question for retrieving questions with similar semantics,which overcomes the impact of polysemy through the introduction of BERT.More importantly,the coupling of LC-LDA and BERT can comprehensively analyze the question semantics and improve the accuracy of semantic similar question retrieval.HQLB-QR further evaluates the quality of the question from multiple dimensions,including life cycle-based question popularity profile model,question evaluation quality model,answer evaluation quality model and user evaluation quality model.These quality models are used to optimize the results of semantic retrieval,so that the retrieved result is not only a set of questions with similar semantics but also with highquality content.Finally,a comparative experiment on the CQADup Stack dataset proves that HQLB-QR can more accurately retrieve high-quality similar question sets compared with LC-LDA,Word2 vec and BERT.(2)High-Quality Domain Expert Finding Incorporating Interest Shift(HQIS-DEF)is proposed for expert finding.It first couples LC-LDA and BERT to extract the semantic representation of the user's domain information from multiple granularities to improve the accuracy of domain expert matching.Then,according to user's activity,the user's potential activity distribution is obtained and used to update the expert's domain semantic representation,which can effectively characterize the expert's domain under current interest.Furthermore,the expert's domain representation modified by interest shift is used for matching with question representation,which renders that the retrieved results are both the experts with domain knowledge and potential interests to the proposed question.Additionally,HQIS-DEF models the quality of the experts and integrates it into the TSWPR(Topic Sensitive Weighted Page Rank)and optimize the result of domain expert finding,making the found expert set with potential interests,similar fields,high evaluation quality,and high professionalism.Finally,a comparative experiment on the CQADup Stack dataset proves that HQIS-DEF proposed in this paper can find high-quality experts more accurately compared with LDA-EF(LDA based Expert Finding)and LPR-EF(LDA based Expert Finding Incorporating with Page Rank).(3)Question-Answerer Matching Model with High Quality for Question Response Time Prediction(QAMHQ-QRTR)is proposed for estimating the arrival time of high-quality answers given by high-level expert.First,HQLB-QR is employed to retrieve the similar question sets with high quality and construct the question feature model that can describe the questioning and answering rules of retrieved questions.Second,HQIS-DEF is utilized to find high-level expert sets and construct answerer feature models for describing the questioning and answering rules of found experts.Then,question feature model is extracted to evaluate the quality of the proposed question according to the question statistic data.Based on the three constructed quality feature models,the Random Forest Classifier is used to build the question response time prediction model,which predicts the time when high-level experts give high-quality answers.Finally,a comparative experiment CQADup Stack dataset proves that proposed QAMHQ-QRTR can more accurately predict the response time of high-level experts giving high-quality answers compared to TSM(Text Statistic based Method)and TITP(Tag Information based Response Time Prediction Method).