Realism

One of the desirable properties for the set of negotiation acts is realism. One way to demonstrate realism is to compare the repertoire with that of a standard agent communication language, such as FIPA [22]. FIPA has a communicative act library, whose repertoire of negotiation acts is similar to the acts given here. Each is given a pragmatic specification using preconditions and effects. The tell act has the effect that the hearer believes the told proposition, and the effect of a propose is only that the hearer believes that the speaker intends the proposed act. These are quite similar to the schemas developed by Allen and Perrault [51] [1] and those of Appelt [5], which were discussed in Section 2.4.

A second way to demonstrate realism is to look to a corpus of human negotiation dialogue to see whether the acts that appear in the corpus are well represented. The TRAINS corpus [33] is a good candidate for this, where a human planner converses with a planning assistant in a meta-level plan to decide how to route some trains in a transportation network. The problem differs from that considered here in that only the user acts in the domain-level plan. However, the expert may still have beliefs and preferences with respect to that plan. Without making too much effort to develop annotation rules or to process large amounts of the corpus, a small sample of ten dialogues was taken so that an rough idea of the distribution of dialogue acts could be established.

The TRAINS dialogues are populated mainly with ask as the naive user consults the system, who is an expert about the domain state. Since the user is the decision-maker, the initiative tends to stay with him, and so it is rare that the system takes initiative in providing unasked-for tells. propose makes up about half of his dialogue. pass occurs only occasionally. The system's acts were mainly tells, as responses to asks. tell was occasionally used as a response to a propose, but never used otherwise. ask was rarely used by the system, since the system is the domain expert. In the cases that it was used, it was for the purpose of clarifying the problem description. In all, there were no instances in these ten dialogues of acts that could not be readily classified using the repertoire developed here, apart from greetings.