Speech act modeling in a spoken dialog system using a fuzzy fragment-class Markov model

In a spoken dialog system, it is an important problem for the computer to identify the speech act (SA) from a user's utterance due to the variability of spoken language. In this paper, a corpus-based fuzzy fragment-class Markov model (FFCMM) is proposed to model the syntactic characteristics of a speech act and used to choose the speech act candidates. A speech act verification process, that estimates the conditional probability of a speech act given a sequence of fragments, is used to verify the speech act candidate. Most main design procedures are statistical- and corpus-based to reduce manual work. In order to evaluate the proposed method, a spoken dialog system for air travel information service (ATIS) is investigated. The experiments were carried out using a test database from 25 speakers (15 male and 10 female). There are 480 dialogs, containing 3038 sentences in the test database. The experimental results show that the speech act identification rate can be improved by 10.5% using the FFCMM and speech act verification with a rejection rate of 6% compared to a baseline system.