Example 1: Risking misinterpretation

This Section will present the first example dialogue planning problem. Often in dialogues, communicative choices come with different levels of risk. In a paper by Carletta [12], it was explained that by making a low risk communicative choice, the speaker must make more effort, but with the benefit that the hearer is less likely to make a misinterpretation that would lead to failure of the plan. On the other hand, the speaker can choose a risky alternative, which requires less effort, but in turn has the risk of costly recovery. Carletta explains a number of such choices such as specificity of referring expressions and indication of focus shifts. There are also corresponding recovery strategies, such as replanning, clarification subdialogue, and repetition of the plan. Sometimes both agents have the opportunity to initiate a recovery strategy. There are many other risky devices that can be used in dialogue. At the semantic level, examples are anaphora and ellipsis, syntactic ambiguity, or use of words and phrases with ambiguous senses. At the pragmatic level, there can be incomplete evidence given by the speaker to unambiguously identify his dialogue plan, or insufficient amount of clarification used by the hearer in eliciting it. Such choices are commonly made when agents try to establish mutual beliefs (known as 'grounding') [15]. Hearers must continually decide whether to open clarifications when it is not certain that the speaker's belief has been established as mutual. As an example of establishing mutual belief, consider giving someone a phone number over a noisy telephone line. The noise creates risk in the outcomes of the agents' communicative choices. It may take several exchanges of confirmations before the hearer is happy that it has the right number, that the other agent believes it has the right number and so on.

In this section, a quantitative approach is taken in planning to take risks in dialogue, complementing Carletta's paper which gives details of particular types of risk-taking, but does not explain how risk-taking might relate to dialogue planning. Only one example is used here, that of the planning of a referring expression to describe an object car-spanner, but other types of risk taking would be planned in the same way. There is a choice between a low risk referring expression, which cannot fail to correctly identify the object, and a high-risk referring expression, which requires less effort, but risks misinterpretation by the hearer as bike-spanner instead of car-spanner The low-risk expression is:

"I need the two-inch hexagonal double-jointed wrench so I
can fix the wheel nut"

The high-risk expression is:

"I need the spanner"

In choosing the high-risk expression, the speaker is declining the initiative in the objective of grounding his intention. Instead, the hearer has a choice of using a clarification dialogue with which it takes the grounding initiative instead. The risk varies according to the state of the belief model. If the belief model causes the agent's candidate intentions to be recognised as equally likely, the risk is high. However, if one of those intentions were to be more likely than the other, the risk is low and so the risky alternative would be a better choice.

The effect of stereotype acquisition will be demonstrated in this example, showing a performance improvement with each dialogue run, as more and more data accumulates. To this end, the sampler as described in Section 3.4.8 is used.