Research On Spoken Language Understanding In Task-Based Dialogue Systems

Task-based human-machine dialogue refers to a dialogue system that can guide users to complete specific tasks.The dialogue process usually has a clear purpose,and the quality of the dialogue system is mainly measured by the completion of the tasks.Spoken Language Understanding(SLU)is one of the core technologies of the task-oriented human-machine dialogue system.It is mainly responsible for understanding the question of what users are saying.The goal of spoken language understanding is to map the natural language input to the user's intention and the corresponding slot value,mainly including the two tasks of intent detection and slot filling.The performance improvement of spoken language understanding module plays an important role in the completion of user tasks and the improvement of user experience in the task-oriented human-machine dialogue system.However,there are many challenges in spoken language understanding task.Due to the characteristics of the dialogue text,many challenges must be addressed.For instance,speech variability exists among different speakers,and user dialogues are generally ultrashort text sequences with abbreviations and stronger contextual dependence than ordinary text sequences.At present,the rapid development of deep learning provides new ideas for the realization of the spoken language understanding technology.This paper studies the characteristics and difficulties of spoken language understanding in the task-oriented human-machine dialogue system,and proposes two deep learning models to complete the spoken language understanding task of single-turn and multi-turn dialogues respectively.The specific work is as follows:Firstly,in the spoken language understanding of single-turn dialogue,how to effectively obtain the deep semantic information of user dialogue is the first step.We propose an attention-based encoder-decoder neural network model for joint intent detection and slot filling.The model encodes dialogue text with a combination of Convolutional Neural Networks and Bidirectional Long Short-Term Memory Networks(CNN-BLSTM)and decoder uses a Long Short-Term Memory Networks(LSTM)based on attention mechanism and alignment inputs.The loss of intention detection and slot filling is added as the loss of the joint model to better learn the connection between intention detection task and slot filling task.Experiments were carried out on the public dataset and good results have been achieved.Secondly,in the spoken language understanding of multi-turn and multi-task dialogue,how to better encode the history context,how to screen the effective history context and how to integrate the history context information into spoken language understanding model are crucial to correctly understand the user dialogue.This paper proposes a joint model of intent detection and slot filling tasks that can effectively encode history context.The model is divided into two parts:a history context encoder and a joint model of intent detection and slot filling.In the process of encoding history context,on the one hand,the history context is combined with the current utterance to reduce the negative impact of irrelevant history context;on the other hand,the attention mechanism is used to filter the history context,so that the history context encoder pays more attention to the history context related to the current utterance to obtain the key information.In addition,a History Gate Unit is introduced in the joint model proposed in this paper to learn the relationship between history context and slot.Experiments were carried out on the public dataset and good results have been achieved.