Research And Implementation Of Vertical Human-Machine Dialogue System For Automobile Field

With the prosperity and development of online services,there are more and more vertical human-machine dialogue systems,all of which require domain-related knowledge as support.However,there is a lack of mature professional knowledge base and dialogue system in the automotive field.This thesis studies and improves the models used in the two stages of semantic understanding and dialogue text generation in the vertical humanmachine dialogue system,and on this basis,constructs a vertical humanmachine dialogue system for the automotive field.The main research contents are as follows:1)In view of the problems of poor standardization of dialogue texts and few proprietary domain datasets,this thesis uses BERT as word embeddings in the semantic understanding stage to build training models.The semantic understanding model includes two tasks:intent recognition and slot filling.This thesis proposes the BERT-CNN model for intent recognition,and conducts a full experimental comparison of different pretrained models.The experimental results show that Mac-BERT-CNN performs the best,reaching an accuracy of 88.36%.In the slot filling task,this thesis proposes the BERT-Bigram-BiLSTM-CRF model.The model learns and filters textual context information,and strengthens the limitation of sequence annotation when outputting,making it better on entity-f1 and more suitable for dialogue systems.2)To balance the accuracy and diversity of dialogue text generation,this thesis proposes the Attention-PHVM model.The PHVM model can generate diverse text,but the accuracy of this model is poor.Considering that the human-machine dialogue system in the automotive field requires higher accuracy,this thesis introduces an attention mechanism to enhance the semantic information of the input.After experimental comparison,the model has improved in multiple indicators,and it is suitable for the dialogue system in general.3)Combining the research of the above two stages,this thesis constructs a vertical human-machine dialogue system in the automotive field,using the knowledge graph stored in Apache Jena as the database.Users can use text to consult car-related information,and the system will identify user semantics and retrieve graph information,and generate diverse and accurate responses back to users based on this information.