The Vocabulary Problem in Spoken Dialogue Systems

Designers of spoken dialogue systems need to be able to predict and constrain the words people use in speech directed at these systems. Larger vocabularies lead to longer processing times, as well as a substantial increase in perplexity errors; according to one estimate (Makhoul 1993), error rates increase with the square root of the number of words in the vocabulary (assuming all words are equally likely). The problem is that a speaker of English may know more than 100,000 words.[1] With such abundance in the mental lexicon, the potential for variability in a speakerÕs word choices is enormous. Consider, for example, the abstract geometric object in Figure 1 and the referring expressions for it that were spontaneously produced in thirteen different conversations. This object is not lexicalizedÐit has no conventional labelÐand in virtually every conversation, speakers took quite a different perspective on it. The mapping of terms to referents is many-to-one, especially when a domain is unfamiliar or when alternative conceptualizations are available. This potential for variability in word choice has been dubbed the vocabulary problem by Furnas, Landauer, Gomez, and Dumais (1983 1987).