Research On Joint Intention Classification And Slot Filling Method Of Task-oriented Chinese Question Answering System

With the continuous development of deep learning and other technologies,intelligent question answering systems have gradually entered people’s lives.Natural language understanding is an important part of the task-based question answering system,which mainly uses two subtasks of intent classification and slot filling to obtain information to accurately identify the domain/intent category and semantic slot of the text.In recent years,the joint intention recognition method has gradually begun to develop.Researchers have jointly trained the two subtasks of intent classification and slot filling,which simplifies the processing flow of the intelligent question answering system,and at the same time enhances each other task through information interaction,hoping to explore more Highperformance intent recognition accuracy,thereby improving the performance of the dialogue system.This paper analyzes the text data and system process characteristics of the task-based Chinese question answering system,and focuses on the problems of intent recognition and slot filling joint optimization and system processing.The following research is carried out:1)This paper proposes a joint intention classification and slot filling model based on part-of-speech attention mechanism.Considering that the actual Chinese dialogue text information is concentrated in the characteristics of nouns and verbs,and the existing methods do not explicitly provide different text weights according to different parts of speech,the model introduces a part-of-speech attention mechanism to increase the model’s attention to different parts of speech.The results of part-of-speech attention extraction are added to the task information interaction using the gate mechanism,which improves the performance of the intention recognition model,and the experimental performance is improved compared with the baseline model.2)This paper proposes a joint intention classification and slot filling model based on an autoencoder.Taking into account that some homophonic errors will be caused during the speech conversion process in the dialogue system processing flow,in order to ensure the accuracy of the intent recognition task and the part-of-speech attention module information,this paper introduces the homophonic error correction module based on the autoencoder into the above model The word embedding of the input text is corrected by the autoencoder,and the autoencoder finds the optimal parameters through pre-training and fine-tuning,which improves the robustness and experimental effect of the intention recognition model.3)Since the public Chinese data set for joint training of the task-based question answering system is relatively rare,this paper collects and annotates the Chinese data set of the task-based question answering system,which contains 3567 pieces of training data.The actual labeling process uses the method of multi-person annotation voting.The Chinese data set of the task-oriented question answering system can be used for subsequent research.