Research On Key Technology Of Slot Filling In Human-computer Dialogue System

With the development of artificial intelligence,the human-computer dialogue system has continuously penetrated into people's lives.In the pipeline method of the human-computer dialogue system,the slot filling task plays a vital role,which directly determines whether the machine can accurately understand the user's intention.In the slot filling task,there are two challenging problems,one is the error propagation problem caused by automatic speech recognition(ASR)errors,and the other is the difficulty of model migration due to the lack of training data.Both of these problems will directly affect the accuracy of the slot filling task and the effectiveness of the human-computer dialogue system.Therefore,this paper focuses on the above two problems.The content of this paper includes how to improve the model structure of the slot filling task,how to alleviate the error propagation problem caused by automatic speech recognition errors without changing the automatic speech recognition system,and how to use the information in domains with sufficient training data to improve the accuracy of domains lacking training data.First of all,this paper proposes a network that combines word information and char information based on the hierarchical decoding model,and uses pre-trained word embeddings in this model.In the CATSLU dataset,the F-score and joint accuracy in all domains have been generally improved.The F-score has increased by an average of 7.34%,and the joint accuracy has increased by an average of 8.75%.Secondly,based on the hierarchical decoding model which combines word information and char information,this paper builds a network that uses transfer learning methods to achieve ASR-error adaptation.Through the probability distribution adaptation method and feature augmentation method,the information in the manual data is integrated into the model,decreasing the maximum mean discrepancy(MMD)of feature distributions between the manual data and ASR data,reducing the slot errors caused by ASR errors,and improving the robustness of slot filling task.In the map domain and the music domain of the CATSLU dataset,the F-score can be increased by 1.67% and 2.05%,respectively,and the joint accuracy can be improved by 1.74% and 4.72%,respectively.Finally,in the hierarchical decoding model which combines word information and char information,this paper implement domain adaptation by sharing models and parameters in multiple domains,enhancing the features of the encoder,and fine-tuning some parameters in the model with target domain data,and successfully integrated the effective information into the model and improved the effect of the domain which lacks training data.In the video domain of the CATSLU dataset,the F-score is increased by 1.98%,and the joint accuracy is increased by 2.56%.