Research On Automatic Question Answering System In Insurance Field Based On Attention Mechanism

With the enhancement of computer computing power,AI has also developed rapidly,meanwhile information and data have shown an exponential explosion.How to accurately obtain the required information from massive data has become a hot topic and objective for people to study.One of the important ways to solve such problems is question answering system.As a application,intelligent customer service can not only reduce labor costs,but also provide standard answers to questions with high repetition rates,which will not be affected by emotional and other factors,and is more efficient and efficient.People’s insurance awareness has gradually increased,because of the living standards have been improved greatly,more and more people are transferring family risks through insurance.So many people want to understand relevant insurance mechanisms.However,some existing question answering systems are often similar to chat robots,once they are involved in relevant professional fields,they may appear to be somewhat mentally retarded.Therefore,designing a question answering system in the insurance field can meet people’s needs in this regard.Based on the above purpose,this article uses the open source insurance dataset,Insurance QA,to complete the research work on the insurance field question answering system.Because the Insurance QA dataset is based on people’s questions on the Insurance Library and answered by senior professionals in the insurance field,it constitutes the dataset.The main contents of this study are as follows:(1)Data set construction and pre-processing.This paper completes the extraction of positive tags from Insurance QA corpus information and the labeling of category tags for each corpus.Combining Wikipedia corpus with 60% corpus information of Insurance QA categories,word vectors of Word2 vec are generated after deactivation,word segmentation and so on.(2)A fast text matching method is designed.Firstly,the text is completed to identify the insurance category corresponding to the user’s question,then the Top-k most similar candidate answers are found through the improved cosine similarity algorithm in the corresponding category question and answer library.Finally,the K candidate answers are calculated with the improved deep learning similarity calculation model to find the most similar questions and answers.(3)A text BLCNN classification model is proposed.Bi-LSTM is used to replace the fixed-window convolution layer in the text CNN model,and the convolution network with an additional convolution core is fused with the original text CNN.By maximizing the pooling,the optimal potential information is extracted and the classification effect is improved.(4)An improved ESIM model AG-ESIM based on attention mechanism is proposed.By introducing attention mechanism to Bi-GRU and replacing the encoding and inference synthesis layer of ESIM model,the model can learn the importance of each word that constitutes a sentence,thus improving the sentence representation ability and accuracy of similarity calculation.The question answering system model designed in this thesis shows that it improves the accuracy,recall rate and F1 value compared with the original model,and the test results of the model on the test set can meet the expectations..