COHERENCE AS A RATIONAL PHENOMENON

We analyze dialogue from the intention recognition point of view Cohen-etal:81,Allen:83,Allen-Perrault:80,Cohen-Levesque:91: when an agent acts, a relation of his action with the interaction context is looked for, to see whether the action represents an attempt to satisfy any intention expressed explicitly by the partner, or an implicit goal which can only be inferred by reasoning on the partner's plans, or if it is a further step in a plan that the agent has already started.

A different approach to the analysis of coherence was adopted by the conversational analysts, who introduced adjacency pairs Sacks-etal:74 to model the expected continuations of an interaction: adjacency pairs are sequences of speech acts (e.g. question-answer pairs) such that, after the first element occurs, the second one is expected. However, as pointed out in [Levinson1981], agent behavior can not directly be explained by means of such type of strict interactional rules. On the contrary, the intentional approach to dialogue interpretation makes it possible to adopt a flexible notion of coherence: a new contribution is considered coherent as long as a relation can be identified among the intentions underlying an action and the previous pending intentions of the interactants.

Following the ideas of [Castelfranchi & Parisi1980], we consider an utterance coherent with the previous context if and only if its receiver can interpret it as a means of the speaker to achieve an unsatisfied goal g which realizes one of the following coherence relations:

1. Goal adherence: g is one of the goals addressed explicitly by the partner in a speech act (not necessarily in the last turn of the interaction).
   Goal adherence covers the phenomena treated by adjacency pairs, but it is a more general approach than that. We represent speech acts as actions Austin:62 having the effect that the interactants share the belief that the speaker intends the partner to execute a certain (domain-level or linguistic) action. An illocutionary act and the action occurring in its effect clearly correspond to the first and second components of an adjacency pair (e.g. a question is performed to induce the hearer to answer it). In the simplest case, dialogues are composed of sequences of adjacency pairs because the participants recognize each other's intentions and react to them as expected. However, other behaviors are modeled in our intentional approach: e.g. insertion sequences Schegloff:72, like ``question-question-answer-answer'' sequences (where further questions are asked before the answer to the initial question is provided) can be modeled as investigations on the actions addressed in the initial question. Asking questions about the preconditions of an action is explained as a coherent continuation, where the speaker tries to execute the action and, to do that, moves to the satisfaction of the subgoal of checking whether the action is executable or not. E.g. consider the following excerpt, taken from [Merritt1976]:

   T1: A: ``Sell me a bottle of whisky, please.''
   T2: B: ``Are you 18?''

   Also notifications of (un)successful performance of actions (e.g. acknowledgements) follow from the satisfaction of a joint goal (see [Cohen & Levesque1991]):

   T1: A: Could you register me for the Artificial Intelligence exam ?
   T2: B: Ok, you are registered.

2. Goal adoption: g is one of the goals that the speaker B has inferred the hearer A is aiming at; this goal has not been expressed explicitly by A, but B has inferred it by reasoning on A's plans. For example, in:

   T1: A: ``I need to borrow a book. Where is the library?''
   T2: B: ``The library is over there, but it is closed for the whole week.''

   B replies to A's utterances by providing him with the information that his plan can not be executed, because one of its constraints is false. B's second utterance follows from the fact the he has adopted A's goal to know whether he may enter the library Allen:83. A didn't express such a goal, but it can be identified from the fact that A is checking another condition (a knowledge precondition Morgenstern:87) of the plan of using the library.

3. Plan continuation: g is a subgoal deriving from the course of actions already undertaken by the speaker; so, the speaker is carrying on the execution of his plan.

   For example, in:

   T1: B: ``Where is the Computer Science library?''
   T2: A: ``It is in the underground floor.''
   T3: B: ``Do you know if it is open today?''

   T3 is not related with A's previous turn T2, but it can be related with T1, as another step in B's preparation for going to the library. Also turns aimed at checking whether actions have succeeded represent continuations of an agent's acting; for example, consider: ``Can you give me your book? .... Did you hear me?''.

In summary, the coherence of the turns of a dialogue is assessed by identifying a structured set of goals which relate the linguistic and domain actions of the interactants. During the interpretation of a turn, a relation among the goals pursued by the turn and the contextual pending goals is searched for in a ``bottom-up'' way (as we will explain in Section 3.2): the turn is matched on the ``nearest'' pending goals and, possibly, on the other, higher-level ones. In this way, the existence of a goal adherence relation is identified before a goal adoption relation, which in turn precedes a possible plan continuation relation. This corresponds to hypothesizing that, before carrying on one's own plan, an agent tries to satisfy his partner's open goals.

Note that this notion of coherence is not necessarily associated with cooperation among the interactants. In fact, it is possible to respect the first two conditions above, while thwarting the satisfaction of a goal g; as pointed out in [Castelfranchi1992], when this happens, an exchange between two debating agents can be perceived as fluent.

In our model, the ``display of understanding'' described by Schegloff is mirrored on the ability of producing goal-coherent turns.

For example, acknowledgements, as well as the production of a coherent turn, display understanding (e.g. consider a proper answer to a question). When an agent tries to interpret a new turn from his interlocutor, he might not be able to relate it with the previous context. In that case, he should reconsider such context to see whether there is an alternative interpretation of some turns which restores the coherence of the whole context.