Cross-Domain Classification Using Generalized Dialogue Acts

Much of the development time of spoken dialogue systems is spent on hand-crafting understanding grammars for every new application domain. Our goal is to speed up this process by building models that are portable across domains. The understanding system is broken up into two parts: dialogue acts (domain independent) and application actions (domain dependent). We propose a two-level classifier that consists of a set of dialogue acts in the first tier and a set of application actions in the second tier. The first tier classifier can be trained from data from various application domains; domain specific data can also be used for adaptation if available. The second tier classifier is simpler and can be trained from a limited amount of domain dependent data. In this paper, we investigate the first tier classifier by building a cross-domain classifier for dialogue acts that uses training data from previously studied domains to categorize utterances in a new domain by the set of the abstracted dialogue acts. We used training data from three different domains: Movie (phone requests for movie information), Carmen Sandiego (computer game) and Travel (flight, car, and hotel phone reservations), to build models. There are a total of 14 dialogue acts used for classification. Examples include: Thank, Suggest, Request_Action. Word and concept bigram models were trained for each of the dialogue acts and used in a maximum likelihood classification framework as described in Potamianos et al., EuroSpeech 1999. Training and testing were performed within each domain (matched) as well as across domains (mismatched).