Reconstructing False Start Errors in Spontaneous Speech Text

This paper presents a conditional random field-based approach for identifying speaker-produced disfluencies (i.e. if and where they occur) in spontaneous speech transcripts. We emphasize false start regions, which are often missed in current disfluency identification approaches as they lack lexical or structural similarity to the speech immediately following. We find that combining lexical, syntactic, and language model-related features with the output of a state-of-the-art disfluency identification system improves overall word-level identification of these and other errors. Improvements are reinforced under a stricter evaluation metric requiring exact matches between cleaned sentences annotator-produced reconstructions, and altogether show promise for general reconstruction efforts.