Specification of communicative acts of utterances based on dialogue corpus analysis

The aim of this paper is to present a specification of speaker's communicative acts of utterances to be carried out gradually by using contextual and cotexual informations. Communicative act is defined as a communicative goal or aim which can be expressed in language L by a distinctive set of conventional cue patterns in specified discourse contexts. In our case, the communicative acts are classified into 28 sorts. The approach based on speech act theory to discourse analysis is considered generally as an efficacious method for topic tracking, ellipsis recovery, anaphora solution, etc. in machine translation of oral dialogues or analysis/generation of utterances on man-machine systems. However, the method requires that speaker's communicative act is interpreted and determined in the process of the analysis. Here is produced an ambiguity problem ; for example, the utterance A could be analysed as either an instruct, or an action-request in automatic analysis, while it is interpreted as an instruct according to human reading in a context given. A : Prenez le bus 3 et descendez à l'arr恡 Gustave Rivet,...(you wanna take the bus 3 and get out at bus stop Gustave Rivet) Aiming at a solution of the ambiguity, we observed task-oriented dialogue corpus, to clarify elements which are presumed to realise the speaker's communicative act of the utterance. There are two different types of elements involved in the communicative act ; contexual elements and cotextual elements. The contexual elements contain surface cue patterns, grammatical and linguistic aspects of utterances, precedent or/and next utterance, connectors, etc. The cotexual elements contain the speaker and hearer, turn taking, conversation stages (open conversation, starting of main topic, close conversation), the relation between speaker's knowledge in the domain and vocabularies used in the conversation. So, we formalised these elements to use as disambiguation informations of the specification of the communicative acts. Finally, we make experiments on the disambiguation using the contextual and cotexual informations. In this paper, we pick up the ambiguity between "Yes" and "Acknowledge", "Action-request" and "Instruction", and "Yn-question" and "wh-question", and show an evaluation concerning them.