MorphoGen: Full Inflection Generation Using Recurrent Neural Networks.

Sub-word level alternations during inflection (apophonies) are an common linguistic phenomenon present in morphologically-rich languages, like Romanian. Inflection learning, or predicting the inflection class of a partially regular or fully irregular verb or noun in such a language has been a widely studied task in NLP, but generative models are limited to capturing the most common ending patterns and apophonies. In this paper, we show how to train a character-level Recurrent Neural Network language model to be able to accurately generate the full inflection of verbs in Romanian, Finish, and Spanish and model stem-level phonological alternations triggered by inflection in an unsupervised way. We also introduce a method to evaluate the accuracy of the generated inflections.