Towards a computational account of knowledge, action and inference in instruction

This is a very simple instruction:

Enter your name in the first line of the form.

Yet it places an almost paradoxical requirement on a speaker issuing it. The speaker means this description to identify an action to the hearer. Thus, to be confident in the instruction, the speaker must know that the hearer can select the right action using the description. Yet, normally the speaker will not know what the hearer will enter to fulfill the directive. In this sense, the speaker cannot know what the hearer will do.

Reconciling these requirements depends on the speaker's representing different states of knowledge explicitly and reasoning about them correctly. In my overview talk, I motivate one such reconciliation, based on three points:

• A modal first-order logic of knowledge can provide a language in which specifications and queries of agents' abilities have a natural form involving existential quantifiers and hypothetical implications.
• Such queries can be posed and evaluated with predictable search using a logic programming proof procedure for this language.
• To plan and validate abstract instructions in NLG, we can therefore create a modal specification of the private and shared information of the hearer and perform an incremental assessment of sentence interpretation using logic programming queries.