Research For Question Intent Recognition Based On Question Answering System

With the development and popularization of artificial intelligence,the emergence of intelligent question answering system has been widely used by people.It can quickly and concisely and accurately return the answers to questions raised by people in natural language.The demand for is increasing,so improving the response time of the question answering system has become the main research goal in this field.The analysis of questions in the question answering system directly affects the answering performance of the question answering system.The question intention recognition task in question analysis is mainly studied in this thesis,which aims to determine the intention category of the question text.The machine learning method is one of the current mainstream methods,but when a large number of labeled Chinese corpora are used to train the classification model There is also the problem of poor classification performance and long time-consuming.Therefore,while improving the classification accuracy,shortening the classification time is of great research value.Moreover,the coarse-grained intent classification task lacks the ability to classify text in text classification.In response to the above problems,the main research work of this article is as follows:(1)In order to solve the problem of low classification accuracy and time-consuming when using machine learning classification for a large number of labeled text corpora.On the basis of Intent classification based on deep learning,introduces a self-attention mechanism and combines the self-attention based on gated loop unit Force mechanism,a user intention recognition method based on gated self-attention(Recurrent Neural Network and Attention inner GRU based on query,RAGRU)is proposed in this thesis.The gating self-attention mechanism is applied before feature learning to focus on the global features of the front to back of the sentence,to enhance the accuracy of sentence feature extraction,and to improve the performance of the classification model.(2)In response to the problems of lack of coarse-grained intent classification task classification ability and the introduction of more invalid characters to affect the original text information,introduces the CD?SFT model on the basis of the RAGRU method and the question intent based on the CD?SFT text representation model.Identification method(RAGRU based CD?SFT,RAGRUT)is proposed in this thesis.The CD?SFT method combines wor2 vec and improved TF-IDF.Through the CD?SFT model to obtain enhanced text information with more class distinguishing ability,the RAGRU method is used to perform deep learning on the enhanced text information,so as to obtain a better classification effect.(3)Insurance question data set in the vertical field is presented in this thesis to verify the effectiveness of the RAGRU and RAGRUT methods.The RAGRU method is compared with other commonly used intent recognition methods based on deep learning.The experiments show that the RAGRU method guarantees the accuracy of classification,Greatly reducing the classification time,the accuracy rate is 1.14% better than the best,and the time is only the fastest 0.89%.Comparing the CD?SFT method with the Word2 vec model,the accuracy is increased by a maximum of 1.68%,and RAGRUT achieves the highest accuracy in the comparison method.