PICARD: The Next Generator

This paper 1 introduces a new line of research which ensures soundness and completeness in Natural Language text planners on top of an efficient control strategy. The work builds on the HUNTER-GATItERER analysis system (Beale. 96; Beale & Nirenburg, 96). That system employs constraint satisfaction, branch-and-bound and solution synthesis techniques to produce near linear-time processing for knowledge-based semantic analysis. PICARD enables similar results for the field of text planning by recasting localized means-end planning instances into abstractions connected by usage constraints that allow HUNTER-GATHERER to process the global problem as a simple constraint satisfaction problem. PICARD is currently being used to plan En-ghsh and Spanish text in the Mikrokosmos Machine Translation Project. ambiguity; these are just a few of the immense challenges a computationM-semantic system faces. Nevertheless, humans process language in real time every day with very little apparent misunderstanding. How can we make a computer do the same? By constraining the problem. With search spaces that can reach 10 TM and more just to deal with basic semantic dependencies in a sentence (disambiguating word senses and determining the semantic connections between them-or lexical choice and implementing semantic connections for generation), exhaustive search techniques are untenable. Constraint satisfaction problem (CSP) techniques allow early dis-ambiguation and drastic search space reduction , while maintaining the integrity (sound-ness) of the solutions found. Beale (1996) introduced a new control strategy for computational-semantic processing. The HUNTER-GATHERER methodology uses knowledge of constraint dependencies to identify small sub-problems which can be processed independently. Solution synthesis methods are then utilized to combine (gather) solutions to these sub-problems, or circles, into larger and larger solutions until the entire sentence is analyzed. As solutions for each circle are created, branch-and-bound and constraint satisfaction techniques are used to prune away (hunt down) non-optimal solutions. HUNTER-GATHERER is a general control strategy that works particularly well for NL problems. Central to our application of this methodology to computational semantics, both in analysis and generation, is the hypothesis that such problems can almost always be viewed as bundles of tightly constrained sub-problems, each of which combine at higher, relatively constraint-free levels to produce a complete solution. Constraint dependency information , retrieved from the semantic co-occurrence information stored in the lexicon, which in turn exploits syntactic information, can be used to partition the complex disam-biguation problem into simpler sub-problems, or "circles of dependency." The concept of relatively independent "cir-cles of dependency" can be exploited to …