Multilingual Dependency Parsing for Low-Resource Languages: Case Studies on North Saami and Komi-Zyrian

Multilingual Dependency Parsing for Low-Resource Languages: Case Studies on North Saami and Komi-Zyrian KyungTae Lim, Niko Partanen, Thierry Poibeau LATTICE (CNRS & ´Ecole normale sup´erieure / PSL & Universit´e Sorbonne nouvelle Paris 3 / USPC) 1, rue Maurice Arnoux, 92120 Montrouge, France kyungtae.lim@ens.fr, nikotapiopartanen@gmail.com, thierry.poibeau@ens.fr Abstract The paper presents a method for parsing low-resource languages with very small training corpora using multilingual word embeddings and annotated corpora of larger languages. The study demonstrates that specific language combinations enable improved dependency parsing when compared to previous work, allowing for wider reuse of pre-existing resources when parsing low-resource languages. The study also explores the question of whether contemporary contact languages or genetically related languages would be the most fruitful starting point for multilingual parsing scenarios. Keywords: dependency parsing, word embeddings, Uralic languages 1. Introduction Developing systems for low-resource languages is a cru- cial issue for Natural Language Processing (NLP). Most NLP systems are built using supervised learning techniques (Weiss et al., 2015; Straka et al., 2016; Ballesteros et al., 2016). These systems require a large amount of annotated data and are thus targeted toward specific languages for which this kind of data exists. Unfortunately, producing enough annotated data is known to be time- and resource- consuming, which means that annotated data, especially of the type required for parsing, is lacking for most lan- guages. To take a recent example, the 2017 CoNLL Shared Task concerned around 50 languages, roughly all of the lan- guages for which enough syntactically annotated data is available in the Universal Dependency format. This was probably (by far) the most ambitious parsing challenge ever undertaken with regard to language diversity, but the fig- ure of 50 languages should be viewed relative to the 6,000 languages in the world: even if one includes only those lan- guages for which written data is available, the 50 languages targeted at CoNLL 2017 cover only a fraction of all the world’s languages. When it comes to parsing, the supervised, monolingual ap- proach based on syntactically annotated corpora has long been the most common one. However, thanks to re- cent developments involving feature-representation meth- ods (a.k.a. word embeddings) and neural network models, it is now possible to develop accurate multilingual models, too. The multilingual approach has yielded encouraging results for both low- (Guo et al., 2015) and high-resource languages (Ammar et al., 2016a). Generally speaking, the multilingual approach can be implemented in two ways. The first involves projecting annotations available for a high-resource language onto a low-resource language us- ing a parallel corpus, while the second aims at producing a cross-lingual transfer model that can work for several lan- guages. Guo et al. (2016) and Ammar et al. (2016a) have conducted multilingual parsing studies for Indo-European languages using the model transfer approach. They demonstrated that a multilingual model can yield better results than monolin- gual models for different European languages. However, their approach relied on the existence of a massive par- allel corpus, as their experiment was based on Europarl.1 Thus the problem …

Content truncated to 500 words