Modelling Expectation in the Self-Repair Processing of Annotat-, um, Listeners

This paper describes a statistical corpus study of self-repairs in the disfluencyannotated Switchboard corpus which examines the time-linear nature of self-repair processing for annotators and listeners in dialogue. The study suggests a strictly local detection and processing mechanism for self-repairs is sufficient, an advantage currently not used effectively under the bonnet of state-of-the-art automatic disfluency processing. We then show how simple local fluency measures using modified language models can be strongly indicative of repair onset detection, and how simple information theoretic measures could characterize different classes of repairs.