Evolvable And Robust Spoken Language Understanding

Due to the rapid development of mobile internet and the application of such emerging technologies as deep learning,speech recognition has almost been in a practical stage.As the interface between speech recognition and back-end application,spoken language understanding(SLU)is needed urgently and more and more important.Decades ago,many universities and research institutes have carried out much exploration and investigation,from rule-based methods to statistical approaches with significantly improving on the performance of SLU.However there are still some lacks on the robustness to speech recognition errors and the portability for dialogue domain extension.Following these two problems,this paper aims to exploring an evolvable and robust spoken language understanding method.It focuses on dealing with speech recognition errors and the extensibility for dialogue domain extension,to seek novel resolutions which could provide the users a better interactive experience in practice.In this paper,the method of language model selection based on domain classification is adopted.It significantly improved the accuracy of speech recognition,and indirectly improved the robustness of understanding spoken language.In term of dialogue domain extension,methods of data expanding and semantic rescoring domain independently are exploited to improve the performance of spoken language understanding as well as dialogue state tracking significantly.