Multilingual Dependency Parsing from Raw Text to Universal Dependencies : The CLCL entry

Even for the most advanced NLP tasks, data goes through basic preprocessing steps, and syntactic parsing is one of them. However, such a poor analysis can completely annihilate the final performance of the downstream applications; it is therefore essential to bring the low-level operations' competitivity as high as possible. The aim of this work is to prepare the DINN system (Discriminative Incremental Neural Network Parser), grand-child of the first transition-based neural network dependency parser, for the University of Geneva's contribution at the CoNLL-2017 shared task, devoted to multilingual dependency parsing. This is the first competition with a strong multilingual vocation (45 languages) over many typologically different languages, taking place in real-world setting without any gold-standard annotation on input (i.e. starting from raw text). This task has been made possible by the Universal Dependencies project, which provides annotated treebanks for a large number of languages using a cross-linguistically consistent annotation scheme. The submitted model performs with respect to the state-of-the-art, and the corresponding results can serve as a baseline for future work evaluating to what extent recently proposed methods have a measurable impact on neural network dependency parsing accuracy