Dialogue Acts and Dialogue Structure

Dialogue Acts and Dialogue Structure

T. Daniel Midgley (University of Western Australia, Australia)
DOI: 10.4018/978-1-60960-741-8.ch006


This chapter discusses historical and recent work in dialogue act tagging and dialogue structure inference. Dialogue act tagging is a classification task in which utterances in dialogue are marked with the intentions of the speaker. It is possible to classify utterances using only features relating to the utterance itself (for example, words and prosodic features), but much work has also utilized dialogue-level features such as previous speaker and previous dialogue acts. The structure of dialogue can be represented by dialogue grammar, segmentation, or with a hierarchical structure.
Chapter Preview


Dialogue act tagging (or DA tagging) is a classification task in which utterances in a dialogue are labeled automatically according to the intentions of the speaker. More formally, for each utterance, given the available evidence e, the tagger will try to select the dialogue act d that has the highest posterior probability P(d|e). Then, by Bayes’ well-known equation:

DA tagging involves using data from a dialogue corpus, which has been marked with DA tags by human annotators. Features and machine learning techniques are selected which will maximize the likelihood of obtaining correct classifications.

Table 1 shows an example of text from VERBMOBIL-2 (Alexandersson 1997), a corpus of appointment scheduling dialogues. Note that each utterance has information about the speaker, the words in the utterance, and annotators’ opinion as to the dialogue act. Other information is available, including phonetic transcriptions and part-of-speech tags.

Table 1.
An example of dialogue from the VERBMOBIL-2 corpus
SpeakerWordsDA tag
DNCI have nothing in NovemberINFORM
DNCpretty much unless you can do that Wednesday Thursday and FridaySUGGEST
RGMno that is bad for me unfortunatelyREJECT
RGMI am <uh> I have got to set up chairs at the county fair <uh>GIVE_REASON
DNChow about <uh> nothing in OctoberREQUEST_SUGGEST
RGMin October let us seeDELIBERATE
RGMI have got <uh> the second through the sixthSUGGEST

Complete Chapter List

Search this Book: