Automatic speech recognition (ASR) systems often normalize transcripts before they are used downstream. This typically means removing filled pauses like "um" and "uh," stripping punctuation, and expanding colloquial forms into standard spellings. The goal is to produce clean, readable text, but this preprocessing may also discard information.
The new arXiv paper focuses on dialogue act classification, the task of labeling each utterance with its communicative function, such as a question, statement, or backchannel. These functions are often signaled by pragmatic cues carried in disfluencies and punctuation. The authors evaluate how normalization affects classification performance.
The paper's title and abstract indicate that this normalization removes useful pragmatic signal. The authors argue that preserving these cues could improve dialogue act classification, and they evaluate the effect of normalization on the task. Their findings suggest that aggressive transcript cleaning may be counterproductive for downstream dialogue understanding.