Neural Factor Graph Models for Cross-lingual Morphological Tagging

  • 2018-05-11 19:27:07
  • Chaitanya Malaviya, Matthew R. Gormley, Graham Neubig
  • 86

Abstract

Morphological analysis involves predicting the syntactic traits of a word(e.g. {POS: Noun, Case: Acc, Gender: Fem}). Previous work in morphologicaltagging improves performance for low-resource languages (LRLs) throughcross-lingual training with a high-resource language (HRL) from the samefamily, but is limited by the strict, often false, assumption that tag setsexactly overlap between the HRL and LRL. In this paper we propose a method forcross-lingual morphological tagging that aims to improve information sharingbetween languages by relaxing this assumption. The proposed model usesfactorial conditional random fields with neural network potentials, making itpossible to (1) utilize the expressive power of neural network representationsto smooth over superficial differences in the surface forms, (2) model pairwiseand transitive relationships between tags, and (3) accurately generate tag setsthat are unseen or rare in the training data. Experiments on four languagesfrom the Universal Dependencies Treebank demonstrate superior taggingaccuracies over existing cross-lingual approaches.

 

Quick Read (beta)

loading the full paper ...