Research On Dialogue State Tracking Based On Multi-granularity Information And Multi-round State Prediction

Task-oriented dialogue systems are used to solve specific tasks and support multiple rounds of dialogue on a certain demand.Task-oriented dialogue system implementation can be divided into two approaches:end-to-end and pipeline.Dialogue state tracking is a crucial part of the pipeline approach.The dialogue state tracking task generates slot operations in the state operation prediction module,and the slot operations are sent to the slot value generation module to generate slot values that are dialogue states.This thesisi focuses on the dialogue state tracking technology,and implements the design improvement of the state operation prediction module based on multi-level information granularity and multi-round state prediction.The specific work is as follows:Aiming at the problem that the redundancy of most information in a long dialogue will cause great trouble to the model,this thesisi uses the Triple Copy strategy dialogue state tracking model(TripPy)to feed three types of discourses in the complete dialogue history into different levels of Transformer encoding layers,and adopts different pooling methods to obtain information of different levels of granularity.Then combined with the attention mechanism,the information is divided into multiple rounds and sent to the slot operation prediction module according to the slot operation type,and the output is sent to the slot value generation module to generate the complete dialogue state.Finally,the effectiveness of the proposed method was verified by comparing the performance of the model with the TripPy model.Aiming at the difficulty of obtaining a large number of correctly labeled samples,this thesis further proposes a data augmentation method based on keyword extraction.A small number of samples in the dataset are randomly selected,the score of each word in the utterance is calculated by combining the clustering algorithm and the attention mechanism,the frequency of the word with the highest score is recorded,and the word with the highest frequency is set as the keyword to be retained.At the same time,the words with the highest score in the remaining words are retained and the words with the lowest score are deleted.The data augmentation method achieves the purpose of enhancing the few-shot learning ability of the model by reducing the data noise.The experimental results show that the proposed state-action prediction module improvement and data augmentation method achieve significant overall performance improvement on the mainstream question answering datasets without significantly increasing the computational resource consumption.Finally,combined with the algorithm proposed in this thesis,a web-based dialogue state tracking system is designed and developed.The system has comprehensive functions,friendly interface,stable performance,and can complete the dialogue state tracking task with high quality.