In order for a dialogue to proceed successfully, it is important that the speakers maintain a consistent view of the interaction. Nevertheless, vagueness and ambiguities can intervene in a speaker's utterance and interfere with the process of interpretation of that utterance. Many researchers have recognized the role of various ambiguity types in generating misunderstandings, from different perspectives Zaefferer:77,Dascal:85,Blum-Kulka-Weizman:88,Vendler:94.
In the linguistic and philosophical research, misunderstandings have been analyzed under different points of view: for example, [Zaefferer1977] and Dascal support the idea that the study of misunderstandings is related with comprehension in an indirect, external way, ``much in the same way as pathological behavior is often said to be able to illuminate the nature of `normal' behavior'' Dascal:85. More recently, misunderstanding has been recognized by [Blum-Kulka & Weizman1988] and [Weigand1997] as a normal phenomenon in communication, to introduce into an harmonical model of dialogue.
In the research on conversational analysis, [Schegloff1987], [Schegloff1992] has analyzed misunderstandings with respect to the sequences of turns in a dialogue, in order to identify which specific mechanisms are used by the interactants to defend the common set of beliefs necessary for the interaction to go on successfully: he calls ``intersubjectivity'' this common ground in dialogue and points out that, during an interaction, the speakers monitor their partners' reactions and interpret them as displays of understanding / misunderstanding of the previous turns. ``Having registered the observation that, through their talk, speakers can display aspects of their understanding of prior talk, it remains to be noted that, in doing so, they can reveal understandings that the speakers of that prior talk find problematic - in other words - what they take to be misunderstanding'' Schegloff:92. Schegloff considers various cases where there are problems in maintaining the intersubjectivity among the interactants; he studies different types of repairs, which can be started by the misunderstood speaker, or by the agent who has misunderstood a previous turn, when they realize that something is wrong with the interpretation. He explains that repairs are an important instrument for reconstructing the mutuality in the interactants' beliefs.
In this paper, we describe a plan-based agent model that takes into account the problem of misunderstandings in (cooperative) agent interaction. This model supports NL communication and interprets dialogue as a rational, cooperative form of interaction among agents. The coherence of a dialogue is assessed by identifying the relation among each turn and the previous interaction context: we model a goal-based notion of coherence, which builds on the idea that the explicit and implicit goals identifyable by interpreting a conversational turn can be related with the previous explicit / implicit goals of the interactants Allen:83,Litman-Allen:87,Carberry:90a. Under the assumption that the interactants cooperate and that every turn is performed to carry on some of their goals jointly, a misunderstanding is hypothesized when the coherence of the dialogue is lost, i.e. an utterance comes which is not related with the contextual goals. In fact, the presence of an utterance that does not contribute to the goals pursued in the receiver's interpretation of the dialogue is taken as a sign that his interpretation context is different from the speaker's one (who certainly considered his own turn coherent).
The processes of analysis and resolution of a misunderstanding are rational behaviors caused by the acquisition of a supplementary goal when an incoherent turn comes: when an agent A interprets a turn and finds it incoherent with respect to the previous context, he can adopt the intention of looking for an alternative interpretation of the interaction. This choice is the consequence of another goal which he decides to pursue: A wants that his view of dialogue converges with his interlocutor's one. If he considers it unlikely that the speaker B has performed a topic shift, or that a breakdown in cooperation has occurred, he looks for an action to modify the wrong interpretation which caused the misunderstanding. Such an action must lead one of the interactants, possibly himself, to change the interpretation context: so, A restructures his own interpretation of the previous dialogue, or he induces B to restructure his dialogue context, depending on who A believes to have made the mistake. This commitment leads A to produce a repair turn, which initiates a subdialogue aimed at restoring the common interpretation ground.
Many researchers (e.g. consider
[Pollack1990], [Perrault1990], [Hobbs et al. 1993],
[Nagao1993] and [Hirst et al. 1994]) have used abductive frameworks
for carrying on the interpretation of a dialogue and possibly
restructuring the dialogue context, when a failure occurs in the integration of
new utterances.
In our plan-based model, the whole interpretation process is represented as
the execution of interpretation (and reinterpretation) actions, as well as
restructuring actions which modify the agent's dialogue
interpretation; moreover, the maintenance of a correct
interpretation context is a mutual goal of the interactants
Cohen-Levesque:91. In this sense, the actions
undertaken by agents to recover from a misunderstanding are not a separate
activity with respect to their normal rational behavior: they are adopted as
commitments by them, when necessary.
We believe that this is a major difference
with respect to the other models of dialogue processing, where the treatment of
misunderstandings (and general communication problems) is embedded in
the normal interpretation process, or is managed by external recovery strategies
(like the metarules used in [Eller1993] to restructure the dialogue
context).
Moreover, some of these models are mainly focused on misunderstandings in the
interpretation of the illocutionary force of utterances and are constrained by
a rigid notion of coherence. Instead, we claim that the misunderstanding
phenomenon goes beyond dialogue and regards wider kinds of interactions;
in our model, we analyze misunderstandings due to ambiguities both
at the level of communicative and domain activity.
Incoherent turns are not always due to misinterpretations: also
topic shifts and intentional breakdowns in cooperation should be considered.
Currently, we don't model topic shifts due to the initiation of new dialogues;
however, as pointed out in other works,[2]For example, consider
[Cohen1984], [Cohen1987], [Litman1986], [Litman & Allen1987],
[Grosz & Sidner1986] and [Grosz & Sidner1990].
focus and topic shifts are usually marked by the presence of ``cue" words.
So, they can be distinguished from misunderstandings.
In our model, this means that when an unexpected change in a
dialogue occurs, the presence of a cue word should
trigger the hypothesis that a topic shift has occurred, before hypothesizing
that a misunderstanding has occurred. So, no misunderstanding analysis would
start. On the contrary, since we model cooperative dialogues, we exclude
the hypothesis that a breakdown in cooperation can occur.
The paper is organized as follows: after a brief description of the phenomenon of misunderstandings (section 1.2), we will present our notion of coherence in dialogue and our approach to the repair of a misunderstanding (section 2). Then, we will describe our computational model of dialogue (section 3): the agent model (section 3.1), the interpretation process of utterances (section 3.2) and the recognition and recovery from a misalignment in dialogue (section 3.3 and 3.4). We will then provide a detailed example to show how our model works (section 4). Section 5 describes the evidence about the occurrence of misunderstandings gathered by studying a number of dialogues. Finally, section 6 compares our model to other related works and section 7 concludes the paper.