CoNLL-UL: Universal Morphological Lattices for Universal Dependency Parsing

CoNLL-UL: Universal Morphological Lattices for Universal Dependency Parsing Amir More,1,∗ ¨Ozlem C¸ etino˘glu,2 C¸ a˘grı C¸ ¨oltekin,3 Nizar Habash,4 Benoˆıt Sagot,5 Djam´e Seddah,5,6 Dima Taji,4 Reut Tsarfaty1 1Open University of Israel, 2Institut fur Maschinelle Sprachverarbeitung, 3University of T¨ubingen, 4New York University Abu Dhabi, 5INRIA, 6Universit´e Paris Sorbonne habeanf@gmail.com, ozlem@ims.uni-stuttgart.de, ccoltekin@sfs.uni-tuebingen.de, nizar.habash@nyu.edu benoit.sagot@inria.fr, djame.seddah@inria.fr, dima.taji@nyu.edu, reutts@openu.ac.il Abstract Following the development of the universal dependencies (UD) framework and the CoNLL 2017 Shared Task on end-to-end UD parsing, we address the need for a universal representation of morphological analysis which on the one hand can capture a range of different alternative morphological analyses of surface tokens, and on the other hand is compatible with the segmentation and morphological annotation guidelines prescribed for UD treebanks. We propose the CoNLL universal lattices (CoNLL-UL) format, a new annotation format for word lattices that represent morphological analyses, and provide resources that obey this format for a range of typologically different languages. The resources we provide are harmonized with the two-level representation and morphological annotation in their respective UD v2 treebanks, thus enabling research on universal models for morphological and syntactic pars- ing, in both pipeline and joint settings, and presenting new opportunities in the development of UD resources for low-resource languages. Keywords: Morphology, Universal Dependencies, Morphological Analysis, Morphological Ambiguity 1. Introduction The development of the universal dependencies (UD) framework and its treebank collection (Nivre et al., 2016; Nivre et al., 2017) follows many shared tasks and multi- lingual evaluation campaigns in which the linguistic repre- sentation schemes across different languages vary (Buch- holz and Marsi, 2006; Nivre et al., 2007; Seddah et al., 2013; Butt et al., 2002; Zeman et al., 2012). The UD tree- banks collection, in contrast, obeys a single set of annota- tion guidelines, and respects the discrepancies between sur- face input tokens and the output nodes in the syntax trees (a.k.a., the two-level representation principle.)1 The UD initiative has paved the way to the development of cross-lingual models for word segmentation, part-of-speech tagging and dependency parsing (Straka and Strakov´a, 2017), as well as cross-linguistic typological investiga- tions (Futrell et al., 2015). Recently, the CoNLL 2017 Shared Task on Multilingual UD Parsing (Zeman et al., 2017), which used a variant of the UD datasets from the UD v2.0 release, introduced a truly end-to-end parsing set- ting: participants had to parse raw texts into dependency trees, implying the initial phases of sentence tokenisation, word segmentation, part-of-speech tagging and morpholog- ical annotation (if their parsing models required morpho- logical information in their input). The UD annotation scheme provides guidelines for unam- biguously annotating the morphological and syntactic lev- els of representation of natural language sentences, but it does not provide means to formally capture a range of po- tential analyses, and in particular the ambiguous morpho- logical analyses, of raw surface tokens. Lexical resources that capture this ambiguity, in the form of morphological analyzers or existing lexicons, are available for many of the ∗ Corresponding author. 1http://universaldependencies.org/ participating treebanks,2 but they are far from useful for …

Content truncated to 500 words