Research Of Text Modeling Based Recommendation Method In Community Question Answering System

Community question answering(CQA)system provides users with a knowledge platform,where users can directly acquire the required knowledge and experts with domain knowledge can efficiently share their knowledge.Community question answering system has attracted a large number of users due to its advantages of convenience,interaction,openness,etc.while a large amount of knowledge data also entered this knowledge platform with the rapid growth of users.How to accurately extract the information required by users from the massive amount of knowledge data is the core question of community question answering system to improve user satisfaction.This paper uses two currently popular text modeling methods to improve answer recommendation and expert recommendation ability in the field of community question answering system recommendation research.This thesis designs an answer recommendation model based on convolutional neural network and attention mechanism and incorporates fine-grained relevance matching into the process of modeling coarse-grained sentences.This model also relieves the problem that modeling a long sequence into a single vector cannot capture all the important information from the sequence and the underused global information which is not available for compare-aggregate framework approaches.This part of experiment uses 3 public datasets for performance verification.The results show that our method can recommend answers more accurately than the current high-level text modeling methods.Pre-trained model has achieved excellent results in many natural language inference tasks in recent years.This paper compares recommended method based on the pre-trained model BERT.The results show that our method has the same level of recommendation ability as the pre-training model based recommendation method,and can learn domain knowledge faster and stronger.For expert recommendation tasks,this thesis designs a heterogeneous graph neural network for modeling expert features and problem features.After mapping questions,answers,and experts as nodes in the graph,we use long short term memory neural networks to perform information interaction according to the corresponding relationships and use these features for final relevance matching.This part of the experiment use pre-processed training set of Zhiyuan-Zhihukanshan Cup 2019 Expert Discovery Contest.The results show that our model can capture more grammatical and semantic information than the non-graph structure neural network model,thereby alleviating the problem of poor modeling effect for long text in the field of non-graph structure neural network based expert recommendation.