Abstract

Even for common NLP tasks, sufficient supervision is not available in manylanguages -- morphological tagging is no exception. In the work presented here,we explore a transfer learning scheme, whereby we train character-levelrecurrent neural taggers to predict morphological taggings for high-resourcelanguages and low-resource languages together. Learning joint characterrepresentations among multiple related languages successfully enables knowledgetransfer from the high-resource languages to the low-resource ones, improvingaccuracy by up to 30%

 

Quick Read (beta)