Evaluation for WFST-based dialog management

Abstract

To construct an expandable and adaptable dialog system which handles multiple tasks, we proposes a dialog system using a weighted finite-state transducer (WFST) in which users concept and system action tags are input and output of the transducer, respectively. To test the potential of the WFST-based dialog management (DM) platform using statistical DM models, we construct a dialog system using a human-to-human spoken dialog corpus for hotel reservation, which is annotated with Interchange Format (IF). A scenario, a Spoken Language Understanding (SLU) and a Sentence Generation (SG) WFSTs are obtained from the corpus and then composed together and optimized to generate a Dialog Management (DM) WFST. We evaluate the detection accuracy of the system next actions using Mean Reciprocal Ranking (MRR). We evaluated how WFST optimization operations contribute to dialog systems and confirmed the optimization enhance the performance of accuracy of the next action detection.