We present two neural systems for the Subtask 2 of CONLL shared task on Universal Morphological reinflection. Both are implementations of encoder decoder RNNs. The first system is a standard encoder decoder network with a soft attention mechanism. The second system first tries to predict the morpho-syntactic descriptions(MSDs) for the target word and then inflects the lemma based on the predicted MSD. This uses two separate encoder-decoder networks for the MSD prediction and the subsequent inflection.
The inflection model is based on CU Boulder's submission to Sigmorphon 2017 and is directly adapted from https://github.com/Adamits/conll2018/tree/master/task1/deep
System1 is adapted from the baseline that was provided by the organizers of Sigmorphon 2018. However it uses a slightly different attention model and different hyper-parameters, and was completely rewritten using PyTorch. The original baseline can be found here https://github.com/sigmorphon/conll2018/tree/master/task2
To run the code, use the following command
python3 baseline/system1.py trainsets/en-track1-high devsets/en-track1-covered devsets/en-uncovered