On Language Spaces, Scales and Cross-Lingual Transfer of UD Parsers

Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL), pages 266 - 281 December 7-8, 2022 ©2022 Association for Computational Linguistics On Language Spaces, Scales and Cross-Lingual Transfer of UD Parsers Tanja Samardži´c1 Ximena Gutierrez-Vasques1 Rob van der Goot2 Max Müller-Eberstein2 Olga Pelloni1 1 Text Group, URPP Language and Space, University of Zurich, Switzerland 2 Department of Computer Science, IT University of Copenhagen, Denmark 3 Center for Information and Language Processing, LMU Munich, Germany {tanja.samardzic, ximena.gutierrezvasques, olga.pelloni}@uzh.ch {robv, mamy}@itu.dk b.plank@lmu.de Barbara Plank2,3 Abstract Cross-lingual transfer of parsing models has been shown to work well for several closely- related languages, but predicting the success in other cases remains hard. Our study is a com- prehensive analysis of the impact of linguistic distance on the transfer of Universal Dependen- cies (UD) parsers. As an alternative to syntactic typological distances extracted from URIEL, we propose three text-based feature spaces and show that they can be more precise predictors, especially on a more local scale, when only shorter distances are taken into account. Our analysis also reveals that the good coverage in typological databases is not among the factors that explain good transfer.1 1 Introduction The goal of cross-lingual parsing is to process a target language as well as possible by exploiting training data available from (an)other language(s). While we know that parsing models can be trans- ferred well across some well-known closely related languages (de Lhoneux et al., 2018), the success of cross-lingual transfer in all other cases remains hard to predict. Surprising cases of syntactic trans- fer between unrelated languages such as Irish and Indonesian (Lynn et al., 2014) illustrate well this unpredictability. A possible explanation for such cases is that genealogically unrelated languages can still be sim- ilar enough to allow transfer. But what is the rele- vant measure of language similarity in such cases? One possible solution is to rely on language fea- tures stored in typological databases such as WALS (Dryer and Haspelmath, 2013; Comrie et al., 2013) or Glottolog (Hammarström et al., 2018). Taking these features as vector representations, languages 1The analysis notebooks are available at https:// github.com/MorphDiv/transfer-lang. can be embedded and compared regardless of their genealogical relations. A popular library URIEL (Littell et al., 2017) has facilitated the use of ty- pological features to measure similarity between languages at different levels (phonology, syntax, ge- ographical distribution). The problem with this so- lution is that the information in linguistic databases is often incomplete and unevenly distributed. Some languages are fully described, while only a few fea- ture values are known for others (Ponti et al., 2019). Nevertheless, a study by Lauscher et al. (2020) on transferring models from English to several other languages suggests that the URIEL language sim- ilarity score is a good predictor of cross-lingual transfer for parsing Universal Dependencies (UD). Our study brings a comprehensive analysis of the relationship between language similarity and the cross-lingual transfer in UD parsing. It extends pre- vious work in two directions: first, we cover many more languages …

Content truncated to 500 words