Research On Intent Recognition In Task-based Dialogue System Based On Deep Learning

As one of the core components of the human-machine dialogue system,natural language understanding has very significant scientific research value.Intention recognition is a sub-task of the natural language understanding system,its accuracy directly affects the performance of natural language understanding,and then affects people's experience of using a humancomputer dialogue system.With the continuous development and improvement of the humanmachine dialogue systems,more and more task-based human-machine dialogue systems are constantly deployed to people's real life,such as smartphone assistant,vehicle voice assistant,and intelligent customer service system in APPs,etc.However,due to the randomness and simplicity of human oral language in the real scene,there are a large number of nonstandard sentences in the existing human-computer dialogue system,which usually show the characteristics of short language and wide content.It is a challenging task to determine the user's intention by analyzing such dialogue text.Considering that the supervised learning method is currently the mainstream method of intent recognition,it requires a large amount of publicly labeled training corpus.However,the work of corpus labeling is not only heavy and time-consuming but also has very high labor cost.Therefore,the first content of this dissertation is to collect a large-scale Chinese multi-intention dialogue corpus—CMID-Transportation in the domain of transportation customer service.The corpus contains 251,094 dialogues,about 1.04 million sentences,involving eight intentions,including not only human to human interaction forms but also human to machine interaction forms.With this corpus,a series of experiments are carried out based on the existing mainstream classification algorithms,and an intention recognition method based on the context modeling of BERT is proposed.This method first digitizes the dialog text into a machine understandable embedded form.The embedding is obtained by simply summing the token embedding,segment embedding,and position embedding.Then it is input into the BERT model to obtain the semantic information of different levels in the text.Then,a classification model is constructed to train the weight of the text semantic vector output by the BERT model and finally obtain the intent classification result.In the second part of this dissertation,a simple and effective method of intention recognition based on data enhancement is presented for word loss or transcription errors in speech recognition.The method firstly imitates the possible problems in the speech recognition process and performs random homophone replacement and random deletion operations on the existing training corpus to obtain training corpus with different semantic information.Then,the training corpus is digitally expressed as a computer understandable embedding and input into the BERT model to obtain different levels of text semantic vectors.Finally,a classification model is constructed to train the text semantic vector to obtain the intention classification results.Finally,in this dissertation,the CMID-Transportation dataset of the two proposed intent recognition methods were used to conduct experiments and compare the experimental results with other methods.The experimental results show that these two methods have certain advantages on Hamming loss,zero-one loss,Micro-F1,and Macro-F1.