Parser engineering and performance profiling

We describe and argue for a strategy of performance profiling and comparison in the engineering of parsing systems for wide-coverage linguistic grammars. A performance profile is a precise, rich and structured snapshot of system (and grammar) behaviour at a given development point. The aim is to characterize system performance at a very detailed technical level, but at the same time to abstract away from idiosyncracies of particular processors. Profiles are obtained with minimal effort by applying a specialized profiling tool to a set of structured reference data (taken from both existing test suites and corpora), in conjunction with a uniform format for test data and processing results. The resulting profiles can be analyzed and visualized at various levels of granularity in order to highlight different aspects of system performance, thus providing a solid empirical basis for system refinement and optimization. Since profiles are stored in a database, comparison with earlier versions, different parameter settings, or other processing systems is straightforward. We apply several salient performance metrics in a contrastive discussion of various (one-pass, bottom-up, chart-based) parsing strategies (viz. passive vs. active and uni- vs. bidirectional approaches). Based on insights gained from detailed performance profiles, we outline and evaluate a novel ‘hyper-active’ parsing strategy. We also present preliminary profiles for techniques for ‘packing’ of local ambiguities with respect to (partial) subsumption of feature structures.