In the linguistic research, the term repair has been used in a wide sense; in fact, phenomena due to problems in the understanding process (e.g. consider questions like ``Are you asking me or do you want to know?'', ``What do you mean?" or problems in the identification of a referent as described in [Heeman & Hirst1995]) have been classified as repairs (e.g. see [Schegloff1992]). In our model these questions are related with subgoals which derive from the execution of a ``Build-interpretation'' action: they are not associated with the resolution of misunderstandings because the speaker has not yet committed to one interpretation of his partner's turn, but he is trying to build one.
If, during the interpretation of an input utterance, the process searching for the coherence of the last turn fails but a local interpretation of the turn is still possible, then the interpretation action produces a context in which this local interpretation is unrelated with respect to the previous context. Since our model does not currently manage topic shifts and breakdowns in partners' collaboration, the agent (A) can only hypothesize a misunderstanding. Hence, in the second phase of the main agent loop (the one where the agent decides which goal to adopt next) A assumes that this interpretation hypothesis is not admissible and does not reflect what his partner has in mind. Under assumption that the partner (B) is actually collaborating, this implies that the intersubjectivity in the interaction has been lost.
In this case, A adopts the goal of realigning the subjective views of
the dialogue; in our model, this is done by the execution of a ``Satisfy"
action on the goal that the A's (private) interpretation of the dialogue is
the same as that of B, up to the misinterpreted turn t
:
Satisfy(A, B,
inter(A, [
, . . .,
], 
) 
inter(B, [
, . . .,
], 
) 
equal(
, 
))
Given this goal, there is a planning phase, in which the agent looks for the
actions having the goal among their effects. The two alternative ways to
achieve the goal are restructuring his own context, or inducing the partner to
change his context. These alternatives correspond to two different
instantiations of the same ``Restructure'' action.
The ``Restructure'' action must be executed by the agent who is misunderstanding
his
partner and consists of modifying his own interpretation of the turns of the
dialogue to obtain the correct interpretation (i.e. the one intended by
the partner).
Restructure 
The execution of the ``Restructure'' action is ruled by the same Agent Modeling actions that describe the performance of the domain-level and linguistic actions. As we said in section 3.1, before committing to an action (and before executing it) an agent checks whether its constraints are true; this checking process corresponds to the diagnosis of the misunderstanding, since the constraints of ``Restructure" are the following ones:
- the agent x (who has to perform the action in order to correct his
wrong interpretation) has an interpretation ctx of the whole dialogue,
up to the last turn 
;
- an alternative interpretation 
of the dialogue can be
attributed to x's partner y;
- 
(determined by comparing ctx and 
) is the coherent
interpretation of the misunderstood speaker, up to the misinterpreted turn
(j < n);
- the agent y corresponds to the speaker of the
first turn that has been interpreted differently in ctx and
(i.e. the last turn of 
, denoted as 
[j]);
To check whether the constraints of an action hold, the agent must perform a
further ``Satisfy" action on the goal of knowing if they are true. The truth
value of a partially instantiated condition depends on whether a
value for its unbound parameters is found that satisfies the condition;
in this specific case, the value is the interpretation of the dialogue
(i.e. the goal is ``Knowref(A, 
, inter(B, [...], 
))'').
The ``Knowref(...)'' goal is obtained by executing an action corresponding to the ``Reinterpret" algorithm described in Figure 2. This action, that the agent undertakes to identify the alternative interpretations of the turns of the dialogue, is very similar to the one previously performed in the standard interpretation task. The main difference is that while in the basic interpretation process the agent considers the interpretation of his own turns as fixed, in this phase, he must consider the alternative interpretations of such turns as well: by doing this, he tries to understand why the partner believes that the interaction is coherent.
When a hearer A chooses an alternative interpretation of a turn 
,
he has not only to check its coherence with the last problematic turn, but he
also has to reconstruct the interpretation of the whole sequence of turns.
In fact, if the speaker B has uttered the last turn, he must have a wrong,
but coherent view of the whole interaction, otherwise he would have started a
repair.
Figure 2: The algorithm for finding alternative interpretations of a dialogue.
The reinterpretation process traces back the dialogue, turn after turn
and looks for a relation between the problematic turn and some previous
(probably misinterpreted) turn.
In principle, it should be possible to look for a global alternative
interpretation of the dialogue, starting from the last
contribution, like in a standard
backtracking procedure. But time would be wasted in producing
many candidate alternatives which fail to relate with the last turn.
Instead, we exploit the newest information for pruning the inadequate
hypotheses: first, we look for an alternative interpretation of a
single turn 
, which explains the last contribution; then, we propagate
the change, to see if the rest of dialogue becomes coherent:
if the interpretations of the other turns are not adequate any more, an alternative for them is looked for.
``Reinterpret'' has two arguments: the agent (agt) and a sequence of
turns ([
, ..., 
]).
When it is called on a list of uninterpreted turns [
, ..., 
], it
goes backward from the problematic last turn 
, towards the beginning of
the dialogue, and stops when it finds the most recent turn 
for which
``Build-interpretation'' has found a new interpretation, coherent with that
of 
.
Then, ``Reinterpret'' propagates the new interpretation ip
( interpretation pair) to the whole context, using the interpretation of
in ip (the first component of this context, denoted by
ip[1]) as a ``pivot", backward by calling itself recursively, and
forward by means of a loop of ``Build-interpretation" on the remaining turns.
When the loop stops, the new interpretation context is returned; this
context is composed of the (sub)context for turns 
, ...
(restC in the figure) with that of turn 
(denoted
as 
).
The algorithm stops changing the previous interpretation when an alternative
interpretation of a turn 
is found that is coherent with the
interpretation of the turns 
..., 
.
The speaker of the last changed turn (
) is the agent who has been
misunderstood.