ON ADAPTIVE ACQUISITION OF LANGUAGE.

Language plays an important role in large-vocabulary speech- understanding systems. It is the code by which we convey meaning, providing a mapping from spoken input to semantics. Furthermore, for the intermediate goal of recognition, language improves performance by constraining the recognizer to consider only grammatical and meaningful word sequences. A deficiency in current large-vocabulary systems is that they are not habitable. That is, a casual user finds it difficult or impossible to speak naturally while remaining within the prescribed vocabulary and syntax of a given recognition system. This paper reports on progress towards understanding how to build habitable large- vocabulary speech-understanding systems.