dependency_parsing_tf

Tensorflow implementation of "A Fast and Accurate Dependency Parser using Neural Networks" https://cs.stanford.edu/~danqi/papers/emnlp2014.pdf

Tensorboard

tensorboard --logdir=path of model variables' folder

example: tensorboard --logdir=/dependency_parsing_tf/data/params_2017-09-18

Recent changes

1. transition to tf 1.2
2. added cube activation function (ref: paper)
3. trainable word embeddings - initialized with 50d word2vec
4. l2 loss for regularization (ref: paper)
5. tensorboard visualization
6. Dev UAS: 90.03 Test UAS: 90.42
7. No functionality for LAS currently. it can be done with few changes in feature_extraction.py. I will try to add it.

training (exisiting dataset)

python parser_model.py

For new dataset

1. Build new vocabulary & embedding matrices -> set "load_existing_dump=False" in parser_model.py. This will overwrite existing "data/dump" directory content
2. python parser_model.py

training dataset

CONLL format