Grammatical gender is assigned to nouns differently in different languages. Are all factors that influence gender assignment idiosyncratic to languages or are there any that are universal? Using cross-lingual aligned word embeddings, we perform two experiments to address these questions about language typology and human cognition. In both experiments, we predict the gender of nouns in language X using a classifier trained on the nouns of language Y, and take the classifier’s accuracy as a measure of transferability of gender systems. First, we show that for 22 Indo-European languages the transferability decreases as the phylogenetic distance increases. This correlation supports the claim that some gender assignment factors are idiosyncratic, and as the languages diverge, the proportion of shared inherited idiosyncrasies diminishes. Second, we show that when the classifier is trained on two Afro-Asiatic languages and tested on the same 22 Indo-European languages (or vice versa), its performance is still significantly above the chance baseline, thus showing that universal factors exist and, moreover, can be captured by word embeddings. When the classifier is tested across families and on inanimate nouns only, the performance is still above baseline, indicating that the universal factors are not limited to biological sex. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 265 Cross-lingual Embeddings Reveal Universal and Lineage-Specific Patterns in Grammatical Gender Assignment Hartger Veeman Uppsala University Department of linguistics and philology Box 635, 75126 Uppsala hartger.veeman.7544@student.uu.se Aleksandrs Berdicevskis University of Gothenburg Spr˚akbanken Box 200, 40530 Gothenburg aleksandrs.berdicevskis@gu.se Marc Allassonni`ere-Tang University Lyon 2 Lab Dynamics of Language 14 avenue Berthelot 69363 Lyon marc.tang@univ-lyon2.fr Ali Basirat Uppsala University Department of linguistics and philology Box 635, 75126 Uppsala ali.basirat@lingfil.uu.se Abstract Grammatical gender is assigned to nouns dif- ferently in different languages. Are all fac- tors that influence gender assignment idiosyn- cratic to languages or are there any that are uni- versal? Using cross-lingual aligned word em- beddings, we perform two experiments to ad- dress these questions about language typology and human cognition. In both experiments, we predict the gender of nouns in language X using a classifier trained on the nouns of lan- guage Y , and take the classifier’s accuracy as a measure of transferability of gender systems. First, we show that for 22 Indo-European lan- guages the transferability decreases as the phy- logenetic distance increases. The classifier’s ability to classify the test nouns is interpreted as an indication of the transferability of grammatical gen- der system from the source language to the target language (i.e., the higher the accuracy is, the more transferable the gender systems are). The entire setting is founded on the cross-lingual represen- tation of words, providing for knowledge transfer between the gender classification models across languages. The embeddings are used to represent nouns in both source and target languages. When the classifier is applied to related languages, does its success depend on how related they are (if yes, this in an indication that some factors are not uni- versal)? Does gender transfer work in the same way for all nouns or are there differences between certain noun classes? 2 Experimental materials and settings In this section, we present the languages involved in this study along with the source of our data. Then, we provide an overview of the cross-lingual word embedding method and the settings of the classifier used for gender transfer. 2.1 Materials Two sources of data are selected for each language. Second, word embeddings are se- lected from pre-trained cross-lingual embeddings published on the fastText website (Joulin et al., 2018).2 Further details about the embeddings are provided in the following subsection. We selected all languages that have grammatical gender and are present in both data sources, with the exception of Albanian due to its small treebank size and Norwegian because in pilot experiments, our classifier showed unexpectedly poor perfor- mance for reasons we were not able to establish. This results in the selection of 24 languages that are shown in Table 1. 273 5 Conclusion This study investigates how grammatical gender is transferable across languages from a transfer learn- ing point of view. The cross-lingual word embed- dings are considered as the source of knowledge shared between languages from which the gram- matical gender of nouns are predicted using a multi- layer perceptron. The empirical results reveals that there exist some universal and lineage-specific pat- terns in the grammatical gender assignment. First, our analysis of gender transfer between Afro-Asiatic and Indo-European languages indi- cated that partly successful gender transfer is pos- sible between non-related languages. While we address the universality of gender assignment cross-linguistically, our data is restricted to languages from two families and our word embeddings are trained on data from specific domains. Additional data from a more diverse sam- ple is needed to further confirm our observations. Furthermore, we Linguistically, grammatical gender is strongly tied to the semantic and for- mal properties of nouns. Since the cross-lingual word embeddings used in this study encode both the formal and semantic information, we cannot disentangle the relative contributions of form and semantics to gender transfer. Finally, it should be mentioned that an important line of research in modern NLP focuses on gender bias present in naturally occurring texts (Caliskan et al., 2017; Gonen et al., 2019). The combina- tion of these questions and approaches with our perspective might become an interesting research direction.