An efficient language understanding approach for mixed-initiative spoken dialogue systems

摘要

A practical spoken dialogue system in a mixed-initiative scheme must be able to handle a variety of concepts supplied by users. This is mostly responsible by a language understanding module, which converts an input word string to a semantic concept understood by the system. This article proposes a novel language understanding approach, which consists of two modules, a subframe extraction module that utilizes weighted finite state automata, and a neural network based concept interpretation module. Given an input sentence, the automaton acts as a robust semantic parser that produces a semantic frame called subframe, and a parsing score. The extracted subframes and their scores are used to interpret a final concept of the sentence using a neural network. Various techniques based on the proposed model are empirically and comparatively evaluated. With more than 40' target concepts founded in our dialogue corpus of hotel reservation, the proposed model achieves considerable results on either typed-in test set or spoken test set.