Research on the combined classification model of question in insurance Q&A system

• 1、School of Software, Beijing University of Posts and Telecommunications,Beijing 100876

Abstract：In the question answering system, the commonly used method is to map the user\'s question to the existing question in the database, and then retrieve the answer to give the user feedback. This method often leads to excessive time overhead for question mapping, which affects system performance. If you can pre-classify the question, you can narrow the scope of the question, reduce the workload of the question mapper, and improve the performance of the system. An excellent classification model can improve the response speed of the system while ensuring the accuracy of the response. In order to get the question classifier with better classification effect, this paper selects the question answering corpus in the insurance field, trains the word2vec word vector model, converts the original Chinese sentence into a computer-readable vector form, and investigates and implements three common neural network classifiers: MLP, CNN and LSTM. Linear regression method is used to combine these three models to obtain a combined classification model. In the experimental stage, in addition to using the test set that comes with the corpus, this paper also uses a test set of variant questions which is manually generated to test the performance of the classifier when it encounters unfamiliar questions. Experiments show that compared with a single classification model, the combined classification model has higher classification performance and better adaptability in the face of unfamiliar sentences with insufficient training.

Keywords： Classification of questions Neural network Combinatorial classifier