A Mixed Language Model for a Dialogue System over the Telephone

This paper describes the work, under way in ITC-irst, on spoken dialogue technology over the telephone. Several activities have been carried out in order to improve either the flexibility and usage facility of ITC-irst voice recogniser and to efficiently afford difficult tasks arising when the system has to cope with large domains. At present, the dialogue technology is sufficiently robust for handling information access in restricted domains (e.g. train timetable inquiry). Furthermore, the portability of the technology towards different domains is guaranteed by an application independent architecture and by easy to use development tools and interfaces. To port the technology on larger domains (e.g. requests of tourism information) several improvements have still to be done, especially in the language models. These latter ones must be capable of handling spoken interactions aimed at retrieving information contained in large and variable databases. Hence techniques and tools for efficiently adapting language models need to be developed. Finally, language generation methods for giving the retrieved information to the callers in a useful way need to be studied