Empirical Verification of Adjacency Pairs Using Dialogue Segmentation

Abstract

A problem in dialogue research is that of finding and managing expectations. Adjacency pair theory has widespread acceptance, but traditional classification features (in particular, 'previous-tag' type features) do not exploit this information optimally. We suggest a method of dialogue segmentation that verifies adjacency pairs and allows us to use dialogue-level information within the entire segment and not just the previous utterance. We also use the X2 test for statistical significance as 'noise reduction' to refine a list of pairs. Together, these methods can be used to extend expectation beyond the traditional classification features.