Tagging Speech Repairs

Tagging Speech Repairs Peter A. Heeman and James Allen Department of Computer Science University of Rochester Rochester, New York, 14627 {heeman, james}@cs, rochester, edu ABSTRACT This paper describes a method of detecting speechrepairs that uses a part-of-speech tagger. The tagger is given knowledge about category transitions for speechrepairs, and so is able to mark a transition either as a likely repair or as fluent speech. Other contextual clues, such as editing terms, word fragments, and word matchings, are also factored in by modifying the transition probabilities. 1. Introduction Interactive spoken dialog provides many new challenges for spoken language systems. One of the most critical is the prevalence of speech repairs. Speech repairs are dysfluencies where some of the words that the speaker utters need to be removed in order to correctly understand the speaker's meaning. These repairs can be divided into three types: flesh starts, modifications, and abridged. A fresh start is where the speaker abandons what she was saying and starts again. the current plan is we take - okay let's say we start with the bananas (d91-2.2 uttl05) A modification repair is where the speech repair modifies what was said before. after the orange juice is at - the oranges are at the OJ factory (d93-193 utt59) An abridged repair is where the repair consists solely of a fragment and/or editing terms. we need to -um manage to get the bananas to Dansville more quickly (d93-14.3 utt50) In these examples, the "-" marks the interruption point, the point that marks the end of the removed text (including word fragments). and precedes the editing terms, if present. In our corpus of problem solving dialogs, 25% of turns contain at least one repair, 67% of repairs occur with at least one other repair in the ram, and repairs in the same turn occur on average within 6 words of each other. As a result, no spoken language system will perform well without an effective way to detect and correct speech repairs. We propose that speechrepairs can be detected and corrected within the local context of the repair. So, clues are needed for detecting repairs that do not depend on such global properties as the syntac- tic or semantic well-formedness of the entire utterance. But this does not mean that syntactic clues cannot be used. One power- fig predictor of modification repairs is the presence of a syntactic anomaly (c.f. Bear, Dowding and Shriberg, 1992) at the interruption point. The anomaly occurs because the text after the interruption point is not intended to follow the text before the interruption, but to replace it, so there is no reason why the text before and the text after need to be syntactically well-furmed. In this paper, we describe how the syntactic anomalies of modification repairs can be detected by a part-of-speech tagger, augmented with category transition probe- bilities for modification repairs. Because we use a statistical model, other clues, such as the presence of editing terms, word fragments, and word conespondence can be …

Content truncated to 500 words