Nucleus Composition in Transition-based Dependency Parsing

Abstract Dependency-based approaches to syntactic analysis assume that structure can be analyzed in terms of binary asymmetric dependency relations holding between elementary units. Computational models for parsing almost universally an unit is a word, while the influential theory Lucien Tesnière instead posits more abstract notion nucleus, which may realized as one or words. In this article, we investigate effect enriching computational with concept nucleus inspired by Tesnière. We begin reviewing how defined framework Universal Dependencies, has become de facto standard training and evaluating supervised parsers, explaining composition functions used make neural transition-based parsers aware nuclei thus defined. then perform extensive experimental study, using data from 20 languages assess impact across different typological characteristics, utilizing variety analytical tools including ablation, linear mixed-effects models, diagnostic classifiers, dimensionality reduction. The reveals gives small but consistent improvements accuracy most languages, improvement mainly concerns main predicates, nominal dependents, clausal coordination structures. Significant factors rate include entropy structures frequency certain function words, particular determiners. Analysis reduction classifiers suggests increases similarity vectors representing same type.