Second language acquisition through human computer dialogue

The paper describes our recent research in developing tools for second language acquisition based on spoken dialogue interaction with a computer. We argue that language proficiency can best be achieved through active communication, and that the computer is very patient and provides a non-threatening environment in which to practice. We have adapted our pre-existing multilingual dialogue systems for this application, focusing in our initial prototype on an English-speaking student learning Mandarin within the weather domain. Two significant new contributions are a Web-based interface for practice exercises to gain proficiency in carrying out a live conversation and a high-quality narrow-domain speech translation capability. In an evaluation on 695 spoken utterances drawn from a corpus of English weather data, our translation system produces an incorrect result less than 2% of the time, with a rejection rate of 8%.