Bilexical Dependencies as an Intermedium for Data-Driven and HPSG-Based Parsing

Bilexical dependencies capturing asymmetrical lexical relations between heads and dependents are viewed as a practical representation of syntax that is well-suited for computation and intelligible for human readers. In the present work we use dependency representations as a bridge between data-driven and grammar-based parsing, both for cross-framework parser comparison and for parser integration. We observe that the state of the art in dependency parsing for English is characterized by broad diversity of dependency representations and seek to systematize properties of various dependency formats pointing out their similarities and differences by carrying out qualitative and quantitative structural analysis and furthermore exploring learnability of four of these representations in automatic syntactic analysis. In addition to comparing syntactic dependencies along several evaluation measures for parsing, we also evaluate the representations in application to the negation resolution task. Using a dependency representation extracted from HPSG structures we contrast three different approaches to parsing—data-driven dependency, phrase structure and a hybrid grammarbased—observe what trade-offs apply along accuracy, efficiency, coverage, and resilience to domain variation and show that explicit, hand-engineered grammatical knowledge helps in both accuracy and cross-domain parsing performance. We complement intrinsic parser evaluation with extrinsic comparison on the negation resolution and semantic dependency parsing tasks discovering that accuracy gains sometimes but not always translate into improved end-to-end performance. A combination of complementary approaches is often a good strategy for achieving improvement. We explore parser integration as a method for advancing the efficiency of a grammarbased parser. Bilexical dependencies serve as an interface for enforcing constraints drawn from the output of the statistical, data-driven systems on the unification-based processing of the grammar-based parser. We experiment with confidence thresholding, filtering and parser ensembles for tackling the problem of selecting high-quality dependencies and propose a technique of static analysis as preliminary evaluation in navigating a large space of various combination setups. We choose configurations optimizing for speed, coverage and balancing the two metrics and carefully evaluate the trade-offs along efficiency, coverage, accuracy and domainresilience.