Measuring Synchrony in Dialog Transcripts

Abstract

A finite register method of processing dialog transcripts is used to measure interlocutor synchrony. Successive contributions by participants are measured for word n-gram repetitions and temporal overlaps. The Zipfian distribution of words in language use leads to a natural expectation that random re-orderings of dialog contributions will unavoidably exhibit repetition -- one might reasonably expect that the frequency of repetition in actual dialog is in fact best explained as a random effect. Accordingly, significance is assessed with respect to randomized contrast values. The contrasts are obtained from averages over randomized reorderings of dialog contributions with temporal spans of the revised dialogs guided by the original durations. Benchmark distributions for allo-repetition and self-repetition are established from existing dialog transcripts covering a pair of pragmatically different circumstances: ATR English language ``lingua franca'' discussions, Air-Traffic communications (Flight 1549 over the Hudson River). Repetition in actual dialog exceeds the frequency one might expect from a random process. Perhaps surprisingly from the perspective of using repetition as an index of synchrony, self-repetition significantly exceeds allo-repetition.