Extraction and Clustering of Two-Dimensional Dialogue Patterns

Abstract

This article proposes a two-step methodology to ease the identification of dialogue patterns in a corpus of annotated dialogues. The annotations of a given dialogue are represented within a two-dimensional array whose lines correspond to the utterances of the dialogue ordered chronologically. The first step of our methodology consists in extracting recurrent patterns. To that end, we adapt a dynamic programming algorithm used to align two-dimensional arrays by reducing its complexity and improving its trace-back procedure. During the second step, the obtained patterns are clustered using various heuristics from the literature. As evaluation process, our method is applied onto a corpus of annotated dialogues between a parent and her child in a storytelling context. The obtained partitions of dialogue patterns are evaluated by an expert in child development of language to assess how the methodology helps the expert into explaining the child behaviors. The influence of the method parameters (clustering heuristics, minimum extraction score, number of clusters and substitution score array) are studied. Dialogue patterns that manual extractions have failed to detect are highlighted by the method and the most efficient values of the parameters are therefore determined.