Co-Parsing with Competitive Models

International Conference RANLP 2009 - Borovets, Bulgaria, pages 173–179 Co-Parsing with Competitive Models Lidia Khmylko Natural Language Systems Group University of Hamburg, Germany khmylko@informatik.uni-hamburg.de Kilian A. Foth smartSpeed GmbH & Co. KG Hamburg, Germany kilian.foth@smartspeed.com Wolfgang Menzel Natural Language Systems Group University of Hamburg, Germany menzel@informatik.uni-hamburg.de Abstract We present an asymmetric approach to a run- time combination of two parsers where one component serves as a predictor to the other one. Predictions are integrated by means of weighted constraints and therefore are subject to preferential decisions. Previously, the same architecture has been successfully used with predictors providing partial or inferior infor- mation about the parsing problem. It has now been applied to a situation where the predictor produces exactly the same type of information at a fully competitive quality level. Results show that the combined system outperforms its individual components, even though their performance in isolation is already fairly high. Keywords Dependency Parsing, Hybrid Parsing 1 Introduction Machine learning techniques for automatically acquir- ing processing models from a data collection and tra- ditional methods of eliciting linguistic knowledge from human experts are usually considered as two alterna- tive roadmaps towards natural language processing so- lutions. Since the resulting components exhibit quite different performance characteristics with respect to coverage, robustness and output quality, they might be able to provide some kind of complementary infor- mation, which could even lead to a notable degree of synergy between them when combined within a single system solution. For the task of dependency parsing the high poten- tial for such a synergy has indeed been demonstrated already. A popular approach for combining alternative de- cision procedures is voting [18]. It makes use of a sym- metric architecture, where a meta component chooses from among the available candidate hypotheses by means of a (weighted) voting scheme. Such an ap- proach not only requires the target structures of all components to be of the same kind, but in case of com- plex structures like parse trees also requires sophisti- cated decision procedures which are able to select the optimal hypotheses with respect to additional global constraints (e.g. the tree property). Since this opti- mization problem has to be solved by the individual parser anyhow, an asymmetric architecture suggests itself as an alternative. i.e.In asymmetric architectures, a master component, a full fledged parser, is solely in charge of de- ciding on the target structure, whilst the others (so called helper or predictor components) provide addi- tional evidence which is integrated into the global de- cision by suitable means. Such a scheme has been extensively investigated for the Weighted Constraint Dependency Grammar, WCDG [3]. External evidence from the predictor components is integrated by means of constraints, which check for compatibility between a local structure and a prediction, and penalize this hypothesis in case of a conflict. So far, however, all the additional information sources which have been considered in this research differed considerably from the master component: They either focused on partic- ular aspects of the parsing problem (e.g. …

Content truncated to 500 words