Research And Application On Joint Learning Technology In Spoken Language Understanding

With the development of computing technology and artificial intelligence technology,man-machine dialogue system gets more and more extensive research, especially the specific task oriented man-machine dialogue system, which can be used in the customer service such as ticket booking, and effectively reduces enterprises' operating costs, has important application value.Spoken language understanding is an important part of man-machine dialogue system, it contains three key tasks: domain identification, slot filling and intention detection. Classic spoken language understanding technology will model three subtasks respectively and execute serially.This method of pipeline defects such as cumulative error. In recent years,there have been some joint model methods that can model some tasks,which can effectively alleviate the error accumulation. On the basis of previous work, this paper research on the joint modeling technology, the main work includes:Proposed a joint model method based on bi-direction Long Short-Term Memory that can joint model the slot filling and intention detection subtasks. The effecting of intention identifier, historical dialogue information, corpus size and other factors to the performance of the joint model are analyzed, And compared with the joint model based on Conditional Random Fields (CRF), experimental results show the effectiveness of the method.Implemented a conference room reservation task oriented man-machine dialogue system integrated using the above models, the system can interacts with users on meeting room reservation and the system has good interactive ability. At the same time, collected a certain size of Chinese man-machine conversation corpus based on the system and tagged the corpus with intention and slot information.