Universal Dependencies v1: A Multilingual Treebank Collection

Universal Dependencies v1: A Multilingual Treebank Collection Joakim Nivre∗ Marie-Catherine de Marneffe◦ Filip Ginter• Yoav Goldberg† Jan Hajiˇc‡ Christopher D. Manning⋄ Ryan McDonald◁ Slav Petrov◁ Sampo Pyysalo▷ Natalia Silveira⋄ Reut Tsarfaty⋆ Daniel Zeman‡ ∗Uppsala University ◦The Ohio State University •University of Turku joakim.nivre@lingfil.uu.se mcdm@ling.ohio-state.edu ginter@cs.utu.fi †Bar-Ilan University ‡Charles University in Prague ⋄Stanford University yoav.goldberg@gmail.com {hajic,zeman}@ufal.mff.cuni.cz {manning,natalias}@stanford.edu ◁Google Inc. ▷University of Cambridge ⋆The Open University of Israel {ryanmcd,slav}@google.com sampo@pyysalo.net reutts@openu.ac.il Abstract Cross-linguistically consistent annotation is necessary for sound comparative evaluation and cross-lingual learning experiments. It is also useful for multilingual system development and comparative linguistic studies. Universal Dependencies is an open community effort to create cross-linguistically consistent treebank annotation for many languages within a dependency-based lexicalist framework. In this paper, we describe v1 of the universal guidelines, the underlying design principles, and the currently available treebanks for 33 languages. Keywords: treebanks, dependency, annotation, multilingual, cross-linguistic, universal. 1. Introduction Multilingual research on syntax and parsing has for a long time been hampered by the fact that annotation schemes vary enormously across languages, which has made it virtu- ally impossible to perform sound comparative evaluations and cross-lingual learning experiments. A striking illus- tration of this problem can be found in Figure 1, which shows three parallel sentences in Swedish, Danish and En- glish, annotated according to the guidelines of the Swedish Treebank (Nivre and Megyesi, 2007), the Danish Depen- dency Treebank (Kromann, 2003), and Stanford Typed De- pendencies (de Marneffe et al., 2006), respectively. The syntactic structure is identical in the three languages, but the percentage of shared dependency relations across pairs of languages is at most 40% (and 0% across all three lan- guages). As a consequence, a parser trained on one type of annotation and evaluated on another type will be found to have at least a 60% error rate when it functions perfectly. The Universal Dependencies (UD) project seeks to tackle this problem by developing cross-linguistically consistent treebank annotation for many languages, aiming to cap- ture similarities as well as idiosyncracies among typolog- ically different languages (e.g., morphologically rich lan- guages, pro-drop languages, and languages featuring clitic doubling). In this way, we hope to be able not only to support comparative evaluation and cross-lingual learning but also to facilitate multilingual natural language process- ing and enable comparative linguistic studies. To serve all these purposes, the framework needs to have a solid lin- guistic foundation and at the same time be transparent and accessible to non-specialists. Several separate initiatives exist to build consistent re- sources for many languages, and the UD project is a merger of some of the initiatives. It combines the (universal) Stan- ford dependencies (de Marneffe et al., 2006; de Marneffe and Manning, 2008; de Marneffe et al., 2014), the universal sv: en katt jagar r˚attor och m¨oss det nsubj conj dobj conj da: en kat jager rotter og mus nsubj ? dobj cc conj en: a cat chases rats and mice det nsubj dobj cc conj Figure 1: Divergent annotation of parallel structures Google dependency scheme (Universal …

Content truncated to 500 words