During an interaction, the agent performs an incremental interpretation of the dialogue, by maintaining as a context the model of his and his partner's activities. Such dialogue context (ctx) is represented by the sequence of local interpretations of the (interleaved) turns of both speakers; in the context, the interpretation of each turn which has been considered coherent is explicitly related with some previous turn by reporting the coherence relation (adherence, adoption and plan continuation) identified with a contextual goal. Instead, incoherent turns are unrelated with respect to the previous context.
The local interpretation of a turn is a complex structure which represents its Agent Modeling, linguistic, internal and domain level actions (Figure 3 shows one of these structures; there, the three action types are contained in boxes labeled suitably, with the ``AM", ``SAM'', ``IM'' and ``DM'' labels). In this interpretation structure, the AM plans relate the linguistic and domain activity, since they not only specify the steps necessary for (linguistic and domain) actions execution, but they also explain how a given speech act contributes to the problem solving activity of an agent who is working to satisfy a domain goal.
The process of local interpretation of a turn takes in input an interpretation structure which only contains the observation that the agent has performed an action (e.g. ``Exec(A, Utterance-act(A, B, ``Do you have a calendar?''))''); at this stage, the input utterance has not yet been interpreted syntactically and semantically. The interpretation process expands this structure by traveling along the plan libraries in order to find which high level goals explain the observation. Since the interpretation structures are composed of two levels, the AM actions and the object-level ones are expanded in parallel. During the expansion of a SAM (Speech Act Model), this process builds the syntactic Lesmo-Torasso:85 and semantic Dieugenio-Lesmo:87 representation of the input sentence (the ``Utterance-act'' action generates a ``Locutionary-act'' if a semantic interpretation can be found for its object text string), and identifies the performed illocutionary act.
In such a complex interpretation structure, the possible ambiguities that lead
to alternative interpretations (the main cause of the phenomenon of
misunderstanding) may concern several aspects: the syntactic and semantic
interpretation (including references) of the sentence can cause the
activation of more than one ``Locutionary-act" hypothesis.
The ambiguity in the illocutionary force leads to identify different actions
representing speech-acts. A domain action can be interpreted as a means
for more than one goal (e.g. going to the library can be a step of both
borrowing book and studying in the library). Even the AM plans allow to
construct different hypotheses on what the agent is doing (e.g.
the applicability conditions of a domain action are checked in two
places: when an agent is exploring the action and when he is executing it).
Finally, the same local interpretation of a turn
can be related to the previous context in different ways.
The interpretation and coherence-seeking processes are represented
by the ``Build-interpretation'' action, which is at the object level, and
takes as arguments the interpreting agent x, the agent y to which x
attributes the interpretation, the dialogue context for turns
, ..., 
, and the last turn 
. The effect of
``Build-Interpretation'' is to build agent x's view of y's interpretation
of the dialogue, for the context 
: 
is obtained from
by incrementing it with the interpretation of turn 
and,
possibly, specifying the relation which links 
with some previous
pending goal of 
:[2]The meaning of predicate
``inter(y, [
, ..., 
], 
)'' is that agent y
interprets turns 
, ... ,
as a context 
.
Build-interpretation 
The Agent Modeling plans in a local interpretation provide the basic information to decide whether the new utterance is coherent with the previous context or not. In fact, the top AM action of a turn interpretation represents a speaker's goal G which led him to produce the turn: if the turn is coherent, G must appear also in the context.
As described in section 2, a new utterance from a speaker A can be related to the dialogue context in different ways: it can aim at the satisfaction of some goal of the partner B (goal adherence / adoption), or at the continuation of the speaker's plan (plan continuation). In the first case, the goal which explains A's new turn appears in some previous turn interpretation of his partner B, having been expressed explicitly by him (by means of a speech act), or inferred by A from B's overt actions. Instead, in the plan continuation case, the top Agent Modeling action of A's turn interpretation can be linked to A's previous metalevel activity (the continuation of an AM plan likely corresponds to the continuation of an object level plan, either a domain or a communicative one, or to the satisfaction of some subsidiary goal due to the cooperation with the partner).
It must be noted that the process performing the local interpretation of a turn operates in a goal directed fashion, in cooperation with the coherence-seeking process: in principle, when a new action of uttering a sentence is recognized, one could try to interpret it locally, and take the context into account only after having recognized the speaker's goal (by looking for it in the context itself). On the contrary, we use some heuristics which start from the (speaker's and hearer's) pending contextual goals to guide the interpretation of the new contribution. This is done by looking for the possible paths which relate an object level action with the plans occurring in the interpretation of the previous context. The search for those paths makes it possible to avoid trying every expansion of the AM local interpretations towards any higher-level goals. Anyway, if the top goal of a local interpretation can not be related with some goal pending from a previous turn interpretation, there is still the chance to expand these interpretations further, in order to find some higher-level goals that explain them. So, our heuristics also search for some higher-level goal, for which there are both a path leading to the new turn interpretation, and a path from some previous turn interpretation to the high-level goal.
Although the last turn of a dialogue may be incoherent with respect to the previous context, it is possible that the interpretation function produces a local interpretation, which remains unrelated. A real failure in the interpretation of a turn happens only when also the local interpretation fails, maybe because the sentence is syntactically / semantically ill formed, or it is impossible to identify the performed speech act, or the underlying domain goal.