Dialogue Acts Recognition - Man-Machine Dialogue: Design and Challenges - page 117

On the one hand, it has to determine a precise classification of the types of act
so that it becomes the system's main reasoning criteria: as the number of the
types of acts increases, the automatic identification can become problematic,
but the interpretation also becomes all the more detailed and the system can
react with all the more relevance. In general, a classification with a reasonable
amount of types is more practical for the system to manage. It has to consider a
corpus annotated in terms of dialogue acts to use these corpora to improve the
system identification performances, for example, thanks to a machine learning
phase. Yet building a reference corpus on these aspects is still done by hand
for now, and it is difficult to ask annotators to choose, for each utterance, one
type of act among a hundred. The annotation task becomes too onerous and
could lead to a number of errors. Moreover, it is far from easy for a human to
conceptualize a hundred types of act, and it does not help the system designer.
Various technical act classification suggestions have been made, from
Foundation for Intelligent Physical Agents (FIPA) to Dialogue Act Markup in
Several Layers (DAMSL), from an XML-based language such as KQML to
the recent ISO 24617-2 standard specified by a work group bringing together
the main international MMD research specialists. The point, especially for the
latter, is to suggest a hierarchical classification that allows the system to
approach the types of act at different levels of granularity; see also the
summary table in [HAR 04, p. 104]. This might be the solution to the
previous section's dilemma: in the case where an MMD system can exploit all
the levels of granularity, a corpus annotator will make do, at first, with the
first level.
The identification of the speech act or dialogue act is a process requiring
the following parameters:
- the words of the utterance themselves and their semantic properties,
especially for the verb [ROS 07];
- the syntactic analysis of the sentence, especially the mood;
- the prosodic analysis of the utterance, especially with the intonation
outline [WRI 02];
- thetypeofthepreviousactinthedialogue, andingeneralanyinformation
arising from the dialogue history that can help link the current utterance to the
previous utterances;