R. De Mori Semantic interpretations can be complete if they cause the dialogue component to have gathered all the required information to produce an answer to the user. They can also be incomplete. There may be more than one complete or incomplete interpretations in competition, due to inherent ambiguities of the spoken message or to imprecision of the recognizer.
All the pertinent conceptual structures C of a given task can be seen as phrases generated by a semantic grammar. The terminal symbols of this grammars are elementary concepts and the non-terminal symbols are conceptual categories. For example, a conceptual category ``trip'' may be a structure containing the elementary concepts ``origin'' and ``destination''. A conceptual structure can be seen as a parse tree representing the generation with the semantic network of a set of elementary concepts triggered by the detection of word patterns. Semantic grammars can be stochastic context-free grammars. Competing interpretations due, among other things. to the detection of competing words and elementary concepts, can be ranked using a combination of acoustic, linguistic and semantic scores, the latter being obtained by parsing with a stochastic context-free semantic grammar (scfsg). Recent results on scoring partial parses of a scfg can be used to score partial semantic interpretation and to find the most probable completion of it. The dialogue strategy may accept the semantic completion as an information intended but not expressed by the speaker or expressed but not detected by the recognizer and ask, for example, for a simple confirmation. The conceptual probabilities P(C/W ) and the language model probabilities P(W) can be dynamically adapted taking the dialogue history into account. History h of previous interpretations, coming, for example, from previously exchanged sentences in a person-machine dialogue can be taken into account in a dynamic model as follows: Pr(CW/h) = Pr(C/Wh)Pr(W/h) Assuming Pr(C/Wh) = Pr(C/h), semantic expectations can be adapted to the evolution of the dialogue. P(W/h) is obtained by a dynamic language model. In a dialogue framework, both types of adaptation can be made dependent on a dialogue state. A dialogue can be described by logical formulae and in theory a dialogue state is defined by the set of values to which all the variables are bound. The number of states defined in this way can be prohibitively high. A good clustering of dialogue states into ``dialogue situations'' can be performed by considering the predicates used in the formula for generating an output message and selecting the one which is more likely to condition the answer.