Research On Entity Link And Relation Classification Method In University Welcome Question Answering System

A knowledge-based intelligent question answering system is an important approach for implementing knowledge and automated question answering.With the continuous development and growth of knowledge graphs,their accuracy and coverage have become increasingly high and widely attention in many application fields.This article focuses on studying the short-text entity linking and relationship classification tasks involved in a knowledge-based university welcome question answering system.The main work is as follows.Firstly,to address the problem of difficult entity linking due to the informal language,insufficient context information,and the use of homophonic characters and abbreviations in short-text questions,this article proposes a short-text entity linking model based on multifeature fusion.In addition to using the question itself,this model also selects information from the knowledge graph,extracts features from multiple perspectives,and fuses them to solve the problem of insufficient short-text information.Secondly,the relationship classification task refers to predicting the relationship category between two given entities in the context.To address the incomplete use of information in each layer of the BERT model and the fixed and inflexible training labels in most relationship classification methods based on BERT pre-training models,this article proposes a relationship classification model based on the weighted CLS layer and dynamic label assignment algorithm.This model first reassigns weights to the CLS position vectors of each layer of the BERT model using the weighted CLS layer algorithm to predict the relationship category,and then uses the designed dynamic label assignment algorithm to dynamically assign training labels based on the semantic information of the samples.The relationship classification model is trained by calculating the loss value with the predicted vector.Thirdly,the effectiveness and comparative experiments of the short-text entity linking model based on multi-feature fusion and the relationship classification model based on the weighted CLS layer and dynamic label assignment algorithm were conducted on the dataset constructed in this study,the public dataset KORE50,and the public dataset Sem Eval.Results were analyzed.Finally,a knowledge-based university welcome question answering system was designed and implemented using the proposed short-text entity linking model based on multi-feature fusion and the relationship classification model based on the weighted CLS layer and dynamic label assignment algorithm.