Resource-Adaptive Model Generation as a Performance Model

Abstract

ABSTRACT: Model generation calculi, close relatives of tableau calculi for theorem prov-ing, can be used as competence models for semantic natural language understanding. Un-fortunately, existing model generation calculi are not yet plausible as performance models ofactual human processing, since they fail to capture computational aspects of human languageprocessing.We outline an extended model generation calculus that solves the most unpleasant com-putational inadequacy; In the extended calculus, tableau expansion rules are equipped withcosts, and model construction is a process that optimizes model quality under resource con-straints with respect to these costs. We embed the new calculus into an abstract inferencemachine and illustrate the possibilities of this approach by presenting a partial theory ofdefinite descriptions in this setting.In this case study, the constants in the universe are given saliences, that are maintainedacross the model generation process. This additional data serves as one important source ofinformation for model quality and resource cost estimation.KEYWORDS: model generation, natural language understanding, bounded optimality, per-formance models