Cross-lingual and Supervised Models for Morphosyntactic Annotation: a Comparison on Romanian

Cross-lingual and supervised models for morphosyntactic annotation: a comparison on Romanian Lauriane Aufrant1,2, Guillaume Wisniewski1, Franc¸ois Yvon1 1LIMSI, CNRS, Univ. Paris-Sud Universit´e Paris-Saclay, F-91405 Orsay 2DGA, 60 boulevard du G´en´eral Martial Valin, F-75509 Paris {lauriane.aufrant, guillaume.wisniewski,francois.yvon}@limsi.fr Abstract Because of the small size of Romanian corpora, the performance of a PoS tagger or a dependency parser trained with the standard supervised methods fall far short from the performance achieved in most languages. That is why, we apply state-of-the-art methods for cross-lingual transfer on Romanian tagging and parsing, from English and several Romance languages. We compare the performance with monolingual systems trained with sets of different sizes and establish that training on a few sentences in target language yields better results than transferring from large datasets in other languages. Keywords: Cross-Lingual transfer, Part-of-Speech tagging, Dependency parsing, Romanian 1. Introduction While most Romance languages are well studied in the Natural Language Processing field and have large sets of annotated data, Romanian still stays behind. As a result, most of the tools generally used in the pre- processing steps of NLP pipelines such as lemmatizer, PoS-tagger or dependency parser are not available for Romanian and, when they exist, their performance of- ten fall far short of the performance achieved, for in- stance, on French or English (Straka et al., 2015). Romanian is therefore a prime candidate for applying transfer methods (Pan and Yang, 2010). Many works have considered the task of transferring PoS-tagger or dependency parser from a resource-rich to a resource- poor language (McDonald et al., 2011; T¨ackstr¨om et al., 2013; Lacroix et al., 2016). However the proposed approaches have generally been evaluated only on resource-rich languages for which the annotated data required to evaluate the transfer models were readily available. In this work, we intend to compare state- of-the-art cross-lingual methods for PoS tagging and dependency parsing on an actual resource-poor lan- guage, Romanian, that, in addition, present several lin- guistic challenges. Indeed, even if Romanian is a Ro- mance language, it has rare properties among this fam- ily: it is a morphologically rich language which kept a case system inherited from Latin and uses more clitics than languages from Romance languages. Moreover, the Romanian went through Slavic influences, notice- able on 15 to 20% of its vocabulary (Haspelmath and Tadmor, 2009; Roegiest, 2006). Our contribution is twofold: • we implement two state-of-the-art transfer meth- ods for PoS-tagging (T¨ackstr¨om et al., 2013) and dependency (McDonald et al., 2011) parsing for Romanian; • we evaluate thoroughly the interest of cross- lingual methods in a realistic use-case of transfer- ring syntactic models from a resource-rich into a resource-poor language. Our work contrasts with previous studies of the two considered transfer methods, as they are evaluated on an actual low-resourced language instead of the usual well-resourced ones. While the data amount in Ro- manian is gradually increasing, and should make it possible to train reasonably competitive systems with monolingual annotated data, we restrict our study to linguistic information available with simple means and without language-specific …

Content truncated to 500 words