Phylogenic Multi-Lingual Dependency Parsing

Proceedings of NAACL-HLT 2019, pages 192–203 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 192 Phylogenetic Multi-Lingual Dependency Parsing Mathieu Dehouck Univ. Lille, CNRS, UMR 9189 - CRIStAL Magnet Team, Inria Lille 59650 Villeneuve d’Ascq, France mathieu.dehouck@inria.fr Pascal Denis Magnet Team, Inria Lille 59650 Villeneuve d’Ascq, France pascal.denis@inria.fr Abstract Languages evolve and diverge over time. Their evolutionary history is often depicted in the shape of a phylogenetic tree. Assuming pars- ing models are representations of their lan- guages grammars, their evolution should fol- low a structure similar to that of the phylo- genetic tree. In this paper, drawing inspira- tion from multi-task learning, we make use of the phylogenetic tree to guide the learning of multi-lingual dependency parsers leverag- ing languages structural similarities. Experi- ments on data from the Universal Dependency project show that phylogenetic training is ben- eficial to low resourced languages and to well furnished languages families. As a side prod- uct of phylogenetic training, our model is able to perform zero-shot parsing of previously un- seen languages. 1 Introduction Languages change and evolve over time. A com- munity that spoke once a single language can be split geographically or politically, and if the separation is long enough their language will di- verge in direction different enough so that at some point they might not be intelligible to each other. The most striking differences between related lan- guages are often of lexical and phonological order but grammars also change over time. Those divergent histories are often depicted in the shape of a tree in which related languages whose common history stopped earlier branch off higher than languages that have shared a longer common trajectory (Jäger, 2015). We hypothe- size that building on this shared history is bene- ficial when learning dependency parsing models. We thus propose to use the phylogenetic structure to guide the training of multi-lingual graph-based neural dependency parsers that will tie parameters between languages according to their common his- tory. As our phylogenetic learning induces parsing models for every inner node in the phylogenetic tree, it can also perform zero-shot dependency parsing of unseen languages. Indeed, one can use the model of the lowest ancestor (in the tree) of a new language as an approximation of that language grammar. We assess the potential of phylogenetic training with experiments on data from the Universal De- pendencies project version 2.2. Our results show that parsers indeed benefit from this multi-lingual training regime as models trained with the phy- logenetic tree outperform independently learned models. The results on zero-shot parsing show that a number of factors such as the genre of the data and the writing system have a significant impact on the quality of the analysis of an unseen language, with morphological analysis being of great help. The remaining of this paper is organized as fol- lows. Section 2 presents both the neural pars- ing model as well as the phylogenetic training procedure. Section 3 presents some experiments over data from UD …

Content truncated to 500 words