Research And Design On Spoken Language Understanding Of Task-Oriented Dialogue System

With the rapid development of deep learning and neural network,task-based dialogue system based on deep learning model is gradually becoming a trend.As the core of the task-oriented dialogue system,spoken language understanding is divided into two parts:intention recognition and slot filling,which provides support for subsequent dialogue state management and dialogue generation.Previous spoken language understanding studies used the independent modeling method to carry out intention recognition and slot filling tasks in turn.However,this method has limitations.Independent modeling may produce cascading errors,resulting in the amplification of errors.Second,the information shared between intention and slot categories is not utilized.On the one hand,the slot filling task will blur the word boundary,resulting in recognition errors.On the other hand,the integration of word information is also beneficial to intention recognition.Finally,due to the lack of data and the difficulty of annotation,few-shot spoken language understanding is of great significance.To sum up,the main work of this paper is as follows:(1)Aiming at the limitations of existing spoken language understanding algorithms,this paper proposes a joint modeling algorithm for spoken language understanding based on multi-level word adapters.The algorithm innovatively establishes word channel and character channel,capturing word-based sentence vector and sequence embedding,wordbased sentence vector and sequence embedding respectively,and deeply integrates them through multi-level word adapter to improve the performance of intention detection task and slot filling task.This paper introduces inter-layer attention to get global information of different layers of BERT model.In addition,this paper designs the gradient balance method to prevent the imbalance in the process of multi-task learning and training,and to improve the model training rate and training effect.The test is completed on the SMP2019 Chinese man-machine dialogue evaluation dataset and CAIS dataset.The results show that the joint modeling algorithm of spoken language understanding based on multilevel word adapters proposed in this paper has advantages over stack propagation,joint BERT,and other algorithms.(2)Aiming at the problem of few-shot spoken language understanding,this paper proposes a few-shot spoken language understanding algorithm based on comparative learning.The algorithm creatively fuses the intention metric space and the slot metric space and uses the method of contrastive learning in the target space to enlarge the distance between irrelevant labels and reduce the distance between relevant labels based on the interval ternary comparison loss function.This paper also designs an abstract conditional random field to mark the sequence in alignment,so as to further improve the effect of slot filling.Experiments are carried out on the SMP2020 Chinese man-machine dialogue evaluation dataset.The results show that the few-shot spoken language understanding algorithm based on contrastive learning proposed in this paper exceeds the baseline algorithm in all indicators.(3)This paper designs the intelligent service layer in the intelligent outbound call system,and realizes its core function,the spoken language understanding module.The system realizes online inference through Tensorflow framework,and verifies the practical value of the algorithm proposed in this paper in the actual scene.