Research On Spoken Language Understanding In Task-based Multi-round Dialogue Systems

With the popularization of artificial intelligence technology,human-machine dialogue systems have gradually appeared in people's daily lives.This task-based human-machine dialogue system can assist users to complete tasks in specific fields through information provided by users.Spoken language understanding is a key module of the task-based human-machine dialogue system.The module can extract the structured data that the computer recognized the user's natural language.Spoken language understanding mainly includes two tasks:intent recognition and semantic slot filling.The deep neural network model can effectively complete the spoken language understanding task in an end-to-end manner,and improve the overall effect of the task-based dialogue system.In the multiple rounds of conversation,the historical dialogue information between the user and the machine can provide additional information support for the spoken language understanding model leading to a better understanding of the user's current statement.However,not all historical information is related to the current sentence.In the human-machine interaction,there might be a transformation of the user's intent,then irrelevant historical information will have a negative impact on the understanding of the current sentence.In addition,in the interaction between people,the common sense and knowledge that people are familiar with are usually omitted from the speech,but the machine does not understand these common sense and knowledge,and cannot reason in the same way as humans.In this paper,historical information and external knowledge are introduced into the joint model of intent recognition and semantic slot filling,so that the module can understand the dialog more accurately.The main research of this article includes:(1)This paper proposes a joint model of deep learning based on historical information for end-to-end spoken language understanding.The model can use historical dialogue information between users and machines to complete the tasks of intent recognition and semantic slot filling.In this model,we use the attention mechanism to assign the corresponding weight to the historical dialogue information to obtain a more relevant and effective historical vector representation,and integrate it into the joint model.In addition,this paper also makes use of the strong relationship between the intent recognition task and the semantic slot filling task in spoken language understanding to design an auxiliary door structure that can use the result of intent recognition to limit slot generation.Experimental results show that the spoken language understanding model proposed in this paper can more accurately identify user intention and semantic slot information.(2)This article proposes a joint model of spoken language understanding based on an external knowledge base.While adding historical information as an auxiliary,it also introduces an external knowledge base as a supplement to knowledge and common sense.How to find the most-required candidate knowledge set from the knowledge base and how to filter the knowledge of the set are the main problems currently facing.For the screening of candidate knowledge,this paper proposes an attention mechanism based on "words",using "words" as query information,and using the attention mechanism to weight the average to obtain candidate knowledge vectors for downstream association model.The experimental results show that the spoken language understanding model added with external knowledge can achieve better results in the tasks of user intent recognition and semantic slot filling.