Token-based Chunking of Turn-internal Dialogue Act Sequences
Piroska Lendvai, Jeroen Geertzen · 2007
In this study we compare two sequence learning approaches to chunk dialogue acts within a speaker's turn.We assign a dialogue act label to each token in the transcribed speech stream of a dialogue participant, additionally classifying if the token is at the beginning of, inside, or outside that specific dialogue act.Experimental findings show that both our approaches -conditional random fields and memory-based tagging -largely improve over local classification methods, obtaining comparable scores on distinct datasets.We discuss the interplay between transcription granularity of turns and dialogue act chunking.