Error-corrective discriminative joint decoding of automatic spoken language transcription and understanding

Following recent trends in the development of spoken dialogue systems, this paper proposes to improve the performance of the user’s intent extraction by means of joint decoding of automatic spoken language transcription and understanding. Gains are expected not only from a better connectivity and mutual awareness of both tasks but also through the use of discriminant models and integration of an error-corrective intermediate mechanism. This latter is based on a statistical post-editing of the speech recognizer word lattice and conditional random fields instantiate the former in our system. An overall absolute reduction of 1.1% is observed by direct application of the proposed techniques on the MEDIA task.