Abstract

Universal Dependencies is a framework for annotating syntactic trees consistently across languages to facilitate multilingual NLP and cross-lingual transfer. However, trees of equivalent sentences might assume non-overlapping shapes because of inherent typological variation. In particular, this anisomorphism is driven by the variation in 1) morphological assets and 2) in clause-level constructions (such as polar questions, predicative possession, relative clauses, etc.). In this work, we demonstrate that reducing the level of anisomorphism yields consistent gains for cross-lingual transfer tasks. First, we show how measuring anisomorphism improves the selection of the source in Dependency Parsing transfer. Second, we put forth a method to preprocess source trees matching their shapes with target trees inspired by typological documentation. This yields improvements in the BLEU scores of syntax-based Neural Machine Translation from Arabic to Dutch, and from Indonesian to Portuguese: we release these new datasets with the code. Our results indicate that the compatibility of the shapes of syntactic trees is crucial for source selection and for boosting cross-lingual transfer.