Research And Implementation Of Cross-domain Spoken Langage Understanding

With the development of science,the people's life become more convenient.People can use a variety of portable devices for natural interaction to get the information and services that previously had to be obtained through human-human interaction,such as Customer Service System,Hotel Reservation System,Q&A System and etc.At the heart of these services is the human-machine dialogue system.Recent years,natural language understanding module obtained a lot of important research results.With the complex of the requirement for information,the traditional task-oriented or single-domain human-machine dialogue system can't meet people's needs well.A cross-domain language understanding module is an integral part of implementing a cross-domain human-machine dialogue system.This thesis researched a technique of cross-domain spoken language understating to support and realize a cross-domain human-machine dialogue system.Based on using the Bi-LSTM for language understanding,this article mainly studies on maximum mean discrepancy and Adversarial Neural Network,which used to model the discrepancy between domains and learn the domain invariant features.And use the learned feature to do the sequence labeling task to complete the slot filling task.The method proposed in this thesis improves the F1 score by 0.0304 on the semi-supervised task in the best case compared with baseline;on the unsupervised task,the F1 score is increased by 0.003.The main work of this thesis are as follows:1.A Bi-LSTM-based dialogue understanding technique is proposed to convert the slot filling task into a sequence labeling task,which forms the basis for follow-up research.2.A cross-domain dialogue understanding model based on the maximum mean discrepancy is proposed.The model minimizes the-loss of the maximum mean discrepancy value,extracts the invariant features of the domain,and realizes the transfer from the source domain to the target domain.3.A cross-domain dialogue understanding model based on adversarial neural network is proposed.The model uses a domain discriminator to extract domain invariant features and realize the transfer from source domain to target domain.4.Based on the above technologies,a cross-domain human-machine dialogue understanding module was designed and implemented to support the research and development of cross-domain human-machine dialogue system.