Research On Text Retrieval In Intelligent Question Answering Of Insurance Industry With Multi-feature

As a kind of risk protection,with the steady improvement of people’s living standards,risk awareness of people is increasing,and the demand for insurance is also increasing.Most users choose to get more information about insurance through the Internet.The rapidly increasing number of questions and a large number of repeated,similar and irregular questions bring great challenges to the traditional way of relying on manual customer service to answer users’ questions.Intelligent Question Answering enables computers to simulate people’s behavior habits of understanding corpus through machine learning,thus answering questions in professional fields and effectively improving business processing efficiency.How to get more accurate answers to personalized questions in professional fields is the core content of intelligent question answering research.As one of the key technologies of intelligent question answering,the accuracy of text retrieval is related to the development of the intelligent question-answering.Based on the investigation of the text retrieval technology in the existing question answering system,this thesis summarizes the development status of the existing technology,analyzes the shortcomings of the existing work,and puts forward corresponding improvement methods.The work of this thesis mainly includes the following three aspects.(1)Problem analysis and reconstruction of question-and-answer data set in the Chinese insurance field.By analyzing the existing insurance question-and-answer data set,it is found that there are some problems such as inaccurate English translation,which can not be directly applied to the construction of Chinese text retrieval methods.In order to promote the development of key technologies in the Chinese insurance question-and-answer system,this thesis has retranslated the original English insurance data set,cleaned the data,segmented the words and marked the data,and constructed a high-quality Chinese insurance question-andanswer data set.At the same time,the characteristics of the reconstructed data set are analyzed,which lays the foundation for this study.(2)Construction of BiLSTM-Text CNN text matching model with multi-features.The existing text matching methods have the problem of insufficient extraction of text information.Based on the analysis of the question-and-answer data set in the Chinese insurance field,this thesis proposes and establishes a text matching model which integrating BiLSTM and Text CNN fusing multi-feature by combining the characteristics of Chinese characters.In order to solve the problem of polysemy of the text,the ERNIE model is used to obtain the word vector representation of the text,which makes up for the lack of prior knowledge in the traditional word vector representation model.At the same time,there are also homophones and polyphones in Chinese,so the phonetic character sequence of the text is introduced to distinguish them.In addition,in view of the problem that there are many proper nouns in the insurance field and it is difficult to identify them,after the conventional part-of-speech tagging,proper nouns are distinguished by especially defining the part-of-speech of proper nouns.After extending the above three text-based semantic features,the BiLSTM and Text CNN models are used to extract the global features and local features of the text respectively,in order to obtain the feature representation of the text more comprehensively.The experimental results on the Chinese insurance data set show that the text matching model which integrating BiLSTM and Text CNN fusing multi-feature has higher evaluation index values than other models,including accuracy,recall and F1 value.(3)The proposal of a text retrieval method with multi-feature.In order to give consideration to both the efficiency and accuracy of text retrieval in question answering systems,a multi-feature text retrieval method is proposed,which divides text retrieval into two stages:text retrieval and text sorting.In the text retrieval stage,the recall method combining words and semantics is used,and in the text sorting stage,the text matching model which integrating BiLSTM and Text CNN fusing multi-feature proposed in this thesis is used for more complex and accurate sorting operations.Taking the insurance question answering data set as an example,this thesis compares and evaluates the multi-feature text retrieval method with other methods.The experimental results show that the performance of the multi-feature text retrieval method is better than other methods,which shows that the model trained by historical search data can better help users in the insurance question answering community get the answers they need and improve their satisfaction.