Sam Steel

Extracting Semistructured Data - Lessons Learnt

The Yellow Pages Assistant (YPA) is a natural language dialogue system which guides a user through a dialogue in order to retrieve addresses from the Yellow Pages. Part of the work in this project is concerned with the construction of a Backend, i.e. the database extracted from the raw input text that is needed for the online access of the addresses. Here we discuss some aspects involved in this task as well as report on experiences which might be interesting for other projects as well.

The bread and butter of planning

This is an introduction to the main areas of planning. It treats the most standard areas in detail, and only sketches less standard ones, since this seems the most sensible thing to do in limited space. It attempts to give the reader a general grasp that will make reading the original and current literature easier. An annotated bibliography is provided at the end of the paper.

The YPA - An Assistant for Classified Directory Enquiries

The Ypa is a directory enquiry system which allows a user to access advertiser information in classified directories [1]. It converts semi-structured data in the Yellow Pages machine readable classified directories into a set of indices appropriate to the domain and task, and converts natural language queries into filled slot and filler structures appropriate for queries in the domain. The generation of answers requires a domain-dependent query construction step, connecting the indices and the slot and fillers. The Ypa illustrates an unusual but useful intermediate point between information retrieval and logical knowledge representation.

Knowledge Subgoals in Plans

Plans naturally contain subgoals about how the world must be. They arise as preconditions of actions in the plan — having keys in order to unlock doors and so on. Plans also contain subgoals about what the agent must know — combinations of safes, Mary’s phone number, which way the outlaw went. Merely stipulating that actions and operands must be known is a hack. This paper offers an account of actions done under uncertainty. If the account is accepted, knowledge subgoals arise automatically, since plans done under uncertainty must still be adequate descriptions of what to do. Plans with knowledge subgoals turn out to be very like ordinary plans.

Quantification in Generative Refinement Planning

This paper brings together a collection of new ideas from generative refinement planning with some more well established results from theorem proving. We add full quantification to a generative refinement planning framework, not by expanding to a universal base [9], but by Skolemizing. We apply our results to causal link planning which leads to a new conflict resolution strategy, a notion called weakening the label.

Sound substitution into modal contexts

I review the problems involved in substituting into modal contexts, and state a rule which permits this to be done soundly. Such a rule justifies a expanded substitution process for modal logic. The process is syntactic but incurs proof obligations. I show constructs that permit the naming of objects outside the current modal contexts. These devices were motivated by problems found while using modal logic in planning.

Actions on belief

ABSTRACT This paper shows how to represent actions on belief (such as deducing and evaluating) in a language in which it is also possible to represent actions on the world (such as putting one block on top of another). It is done by combining modal logics of belief and of action in such a way that actions on belief can be represented as perfectly respectable modals with standard and well-motivated semantics, obeying sensible rules. One basic actions is proposed, representing finding that a fact S is true; it has a variant, representing finding that S is true given a fact (or theory) T. These can be used to define other, more usable, actions: finding whether S, finding an object with property o, and evaluating a term.