Source code for the paper "Morphological Inflection Generation with Hard Monotonic Attention"
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.


Source code for the paper: Morphological Inflection Generation with Hard Monotonic Attention.

Requires dynet. You should compile the aligner before running by running make while in the src directory.

Usage: [--dynet-mem MEM][--input=INPUT] [--hidden=HIDDEN] [--feat-input=FEAT] [--epochs=EPOCHS] [--layers=LAYERS] [--optimization=OPTIMIZATION] [--reg=REGULARIZATION][--learning=LEARNING] [--plot] [--eval] [--ensemble=ENSEMBLE] TRAIN_PATH DEV_PATH TEST_PATH RESULTS_PATH SIGMORPHON_PATH...


  • TRAIN_PATH train set path
  • DEV_PATH development set path
  • TEST_PATH test set path
  • RESULTS_PATH results file (to be written)
  • SIGMORPHON_PATH sigmorphon repository root containing data, src dirs (available here:


  • -h --help show this help message and exit
  • --dynet-mem MEM allocates MEM bytes for dynet
  • --input=INPUT input embeddings dimension
  • --hidden=HIDDEN lstm hidden layer dimension
  • --feat-input=FEAT feature embeddings dimension
  • --epochs=EPOCHS number of training epochs
  • --layers=LAYERS number of layers in lstm
  • --optimization=OPTIMIZATION chosen optimization method ADAM/SGD/ADAGRAD/MOMENTUM/ADADELTA
  • --reg=REGULARIZATION regularization parameter for optimization
  • --learning=LEARNING learning rate parameter for optimization
  • --plot draw a learning curve plot while training each model
  • --eval run evaluation on existing model (without training)
  • --ensemble=ENSEMBLE ensemble model paths, separated by comma

For example:

python --dynet-mem 4096 --input=100 --hidden=100 --feat-input=20 --epochs=100 --layers=2 --optimization=ADADELTA  /Users/roeeaharoni/research_data/sigmorphon2016-master/data/navajo-task1-train /Users/roeeaharoni/research_data/sigmorphon2016-master/data/navajo-task1-dev /Users/roeeaharoni/research_data/sigmorphon2016-master/data/navajo-task1-test /Users/roeeaharoni/Dropbox/phd/research/morphology/inflection_generation/results/navajo_results.txt /Users/roeeaharoni/research_data/sigmorphon2016-master/

If you use this code for research purposes, please use the following citation:

  author    = {Aharoni, Roee  and  Goldberg, Yoav},
  title     = {Morphological Inflection Generation with Hard Monotonic Attention},
  booktitle = {Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  month     = {July},
  year      = {2017},
  address   = {Vancouver, Canada},
  publisher = {Association for Computational Linguistics},
  pages     = {2004--2015},
  abstract  = {We present a neural model for morphological inflection generation which employs
    a hard attention mechanism, inspired by the nearly-monotonic alignment commonly
    found between the characters in a word and the characters in its inflection. We
    evaluate the model on three previously studied morphological inflection
    generation datasets and show that it provides state of the art results in
    various setups compared to previous neural and non-neural approaches. Finally
    we present an analysis of the continuous representations learned by both the
    hard and soft (Bahdanau, 2014) attention models for the task, shedding some
    light on the features such models extract.},
  url       = {}