Multilingual dependency parsing from universal dependencies to sesame street

Abstract

Research on dependency parsing has always had a strong multilingual orientation, but the lack of standardized annotations for a long time made it difficult both to meaningfully compare results across languages and to develop truly multilingual systems. The Universal Dependencies project has during the last five years tried to overcome this obstacle by developing cross-linguistically consistent morphosyntactic annotation for many languages. During the same period, dependency parsing (like the rest of NLP) has been transformed by the adoption of continuous vector representations and neural network techniques. In this paper, I will introduce the framework and resources of Universal Dependencies, and discuss advances in dependency parsing enabled by these resources in combination with deep learning techniques, ranging from traditional word and character embeddings to deep contextualized word representations like ELMo and BERT.