Derstanding. a Computational Approach to Human Living with Clas- Sic: When and How to Use a Kl-one-like Language

provide role functions on event tokens that yield the parameters of the event, the parameters lling the roles are basic terms, and therefore no structural matching is performed on them. In general, events are terms rather than structures, and terminological inferences are not of concern. There are only few systems known to me that exploit subsumption to perform reasoning about plans: SUDO-PLANNER Wellman, 1988], CLASP Devanbu and Lit-man, 1991] and T-REX Weida and Litman, 1992]. Of these systems, CLASP is the closest to my work. The CLASP representation language provides description-forming operators that specify temporal and conditional relationships between actions represented in CLASSIC; plans are built out of actions, whose deenition follows the classic STRIPS planning operator deenition. Subsumption in CLASP builds on and expands term subsumption in CLASSIC, in order to deal with plan and scenario classiication | scenarios are plan instances. T-REX Weida and Litman, 1992] is similar in spirit to CLASP, but is able to capture ner temporal relations, and integrates terminological reasoning with constraint network reasoning to classify plans. CLASP and T-REX therefore take the rst step toward building and managing a hierarchy of plans | however, these system don't provide a linguistically sound representation of actions and plans. 6 Conclusions I've suggested that two desiderata for Plan Recognition as applied to NLU are providing more exibility in the matching process, and including in the representation not only planning, but also linguistic components. I've proposed that DLBS's provide a solution to both representing linguistic and planning knowledge about actions , and computing relations between such descriptions. Note that a similar approach to action representation could also beneet planning as applied to NL, which is required both in interpretation (ie to interleave the step Open the basement door (if it's closed) in the plans developed from Exs. (4a) and (4b)), and in generation. Finally, notice that the classiication algorithm is also useful in relaxing step (2a), using as an index into the KB to nd recipes that achieve. In my current implementation such match is only allowed when is an instance of the header of a recipe. However, this step can be relaxed to also retrieve recipes when is more abstract than the header: could be used to select the best matching recipe, if it has enough discriminatory characteristics .