Universal Dependency Parser: A Single Parser for Many Languages on Arc-Swift

Dependency parsing has been a longstanding task in NLP with a recent boost in performance thanks to neural network models. However, most dependency parsers are monolingual–a single parser is trained per language–and utilize transitionbased systems that are limited to local information. We utilize a novel transitionbased system, arc-swift, proposed in [1] that incorporates both local and global information and achieves improved performance compared to current transitionbased systems. We propose a universal, multilingual dependency parser model built on arc-swift and augment this model by aligning GloVe word vectors across languages and prepending the parser with a character-level language detector model to help the parser explicitly learn language embeddings. Our results show that word vector alignments greatly improve UAS and LAS scores, while language detector predictions show modest increase in parse scores. Our universal dependency parser performs comparably to monolingual parser baselines models, and is potentially able to help advance downstream applications, such as machine translation.