The paper presents an algorithm for locating parts of an utterance that is canceled by the speaker using repairs, or self-correct ions. Especially, we concentrate on identification of the on-set of a canceled part. The underlying idea is that the repair and reparendum intervals are as similar to each other as are two coordinated phrases. We adopted the similarity-based algorithm for analyzing Japanese coordination structures proposed by Kurohashi and Nagao. The success rate of the algorithm is 92% for tagged utterances in ATR dialogue database.