Research On Semantic Analysis Technology For Question And Answer System In Tourism Field

As the demand for information accuracy grows rapidly,the importance of vertical field question and answer systems is becoming increasingly prominent.The semantic parsing module of the vertical domain question answering system can understand the real needs of users,and then quickly provide accurate information to users through the knowledge base behind.The paper focuses on the construction of the semantic parsing module of the specific question and answer system in the tourism field,and provides a reference for the construction of the semantic parsing module of different vertical domain question answering systems.The semantic parsing module in the thesis contains three sub-modules: main classification intent recognition sub-module,fine classification intent recognition sub-module and slot identification sub-module.In order to meet different application scenarios,two different construction methods are given for each sub-module.The main classification intent recognition sub-module uses the design method based on LightGBM algorithm and the design method based on BERT model.The design method based on LightGBM algorithm uses template information for automatic feature engineering.The fine classification intent recognition sub-module uses the design method based on Text-CNN model and the design method based on BERT model.The Text-CNN-based design method is used to vectorize the domain knowledge to improve the performance of the model.The slot identification sub-module uses a design method based on the LSTM model and a design method based on the BERT model.Test comparison and result analysis were carried out for different sub-module design methods,and the application scenarios of different sub-module construction methods were determined.Due to the high memory requirements of the BERT model,if the different sub-modules use the BERT-based construction method at the same time,there may be a problem that the model memory usage is high.The paper attempts to design a different sub-module to share the construction method of a BERT model,so as to alleviate the problem of high memory usage of the model when different sub-modules are simultaneously constructed using the BERT model.Since the data acquisition in the vertical domain has been difficult,it is an ideal way to use the data accumulated in other fields when constructing the semantic parsing module in the vertical domain.In the paper,the BERT model is used to carry out the migration experiment of the main classification intent recognition sub-module,and it is found that the BERT model has weak dependence on the data field.The experiment found that the problem of insufficient data in the existing field can be compensated by enhancing the amount of data in the tourism field.That is,when constructing a new vertical domain main classification intent recognition sub-module,the data of different vertical fields accumulated before can be used to enhance the performance of the module.