Research On Answer Selection Method Based On BERT Model

Faced with the rapid growth of information in the process of Internet development,people can access all kinds of information and knowledge more conveniently than before,but the massive data also brings challenges to the current question answering system.How to quickly and accurately obtain valuable information has become the focus of the current research.We focus on the answer selection task in question and answer systems in this thesis.The key of the task is to calculate the semantic similarity between question sentences and candidate answer sentences,so as to filter out the answer that best matches the question.In order to solve the problems of weak semantic interactions and insufficient background information of question and answer sentences in previous models,we improve on a deep learning model based on the BERT model and proposed an answer selection model that incorporates question classification information and additional knowledge representations.The model first uses the question classification information to process the candidate answer sentences to enhance the semantic interaction between the question and answer sentences;then expands the context vector by introducing the exogenous knowledge base to enable the model to extract richer semantic features.Our main research work is as follows:(1)In the current model,there is less interaction of semantic information between question and answer sentences,and the classification information of question sentences cannot be directly applied.An answer selection model that combines question classification and pre-training model BERT is proposed.First,the expected answer type of the question sentence is obtained through the question classification,and then the irrelevant words in the candidate answer sentence are masked according to the expected answer type of the question sentence.Finally,the BERT model is used to further integrate the syntactic and semantic features in the question sentence and the answer sentence,and then calculate the semantic similarity of question-answer pairs.The experimental results show that the answer selection model combined with question classification information can effectively improve the accuracy of answer selection task,In the Trec QA Clean dataset,the MAP and MRR indicators are improved by 1.6% and2.5%,while in the Wiki QA dataset,the MAP and MRR indicators are improved by 1.4%and 0.8%.(2)Due to the limited length of question-answer pairs,it is difficult for the model to obtain complete semantic information in the existing input sequence.To solve this problem,we proposed a knowledge-based answer selection model.By combining the knowledge representation of the exogenous knowledge base and the vector representation based on the BERT model,the semantic features of the question-answer pairs are represented using the fused vector representation,which enhances the application of background knowledge in the question-answer sentences and enriches the vector representation of the question-answer pairs.The experimental results show that in the Trec QA Clean dataset,the MAP and MRR indicators are improved by 1.5%and 1.1%,and in the InsuranceQA dataset,the accuracy is improved by 1%.