Character sequence-to-sequence model with global attention for universal morphological reinflection

Abstract

This paper presents a neural network based approach for the CoNLL-SIGMORPHON-2017 Shared Task 1 on morphological reinflection. We propose an encoder-decoder architecture to model this morphological reinflection problem. For an input word, every character is encoded through a Bi-directional Gated Recurrent Unit (GRU) network. Another GRU network is deployed as a decoder to generate the inflection. We participate in Task 1, which includes 52 languages without using external resources. In each language, three training sets are provided (high, medium, and low respectively represent the amount of training data; Scottish Gaelic only has medium and low), totally 155 training sets. Due to time constraints, we only search for optimized parameters of our model based on the Albanian dataset. The source code of our model is available at https://github.com/valdersoul/conll2017.