Research And Implementation Of Dialogue Act Classification Based On Deep Learning

Dialogue act is a shallow category of user intention,which expresses the most direct behavior of users.As the basic task of natural language understanding,the classification of dialogue act plays an important role in the dialogue system.Its purpose is to identify the dialogue act of dialogic content and assist in making dialogic decisions.With the development of artificial intelligence,dialogue system is gradually integrated into people's life,providing intelligent and convenient services.In order to improve the user experience and fit the user's wishes,the system needs to provide more accurate semantic understanding,so the research of dialogue act classification has important significance and practical value.According to research findings,most existing work is based on the classification model for single-label dialogue act classification,however,multi-label situations may also exist,and the work of mining the dialogue act with multiple labels is not much,and rarely consider the relationship between clauses and the impact of syntactic information.Meanwhile,most work treats the dialogue act classification as a single task learning,without considering the role of other related tasks.Based on the above question,the investigation of work related to dialogue act classification and characteristics of dialogue act and task-oriented dialogue system are considered,the specific content of which is as follows:First,a classification model based on syntactic and clause information is proposed.According to the characteristics of the dialogue acts,BiLSTM is used to learn the context information of each word and GCN is adopted to learn the syntactic information between words to obtain the semantic representation of the sentence by word.Meanwhile,since the sentence is composed of multiple clauses and each clause represents a dialogue act label,the dependency relationship between clauses can also be obtained through BiLSTM.Then,based on the selfattention mechanism,important clause information can be identified to get the semantic representation of the sentence.The sentence representation based on words and clauses is integrated as the semantic representation of the whole sentence and the corresponding dialogue act labels are obtained.Finally,the validity of the approach is verified by a number of comparative experiments.In the method of joint learning,the loss of training process contains the loss of above two tasks.Finally,a better model for dialogue act classification and slot filling is obtained.In the experimental part,compared with several neural network classification models,the effectiveness of the method is verified,and the joint learning is proved to have a positive effect on improving the generalization ability of the model.Then,joint learning of related tasks is considered to improve the generalization ability of the target task.Dialogue act classification and slot filling are important components of natural language understanding.Dialogue acts represent semantic intents,and slots represent semantic constraints.Therefore,this thesis considers intents and slots have a certain relationship.The model learns word context information based on BiLSTM and word syntax information based on GCN.Then,through the Self-Attention mechanism,the semantic expression of the sentence can be obtained.By the semantic representation of words and sentences,the slot information and dialogue act labels of sentences are output respectively.Finally,due to the relatively abstract process of dialogue act,this thesis implements the process from input dialogue content to output dialogue behavior based on the open-source Web Framework Django of python,that is,the back end forecasts the results by loading the two neural network pre-trained models proposed in this thesis,so as to visualize the process of dialogue behavior classification.