AC-Tagger

Requirements

Python 2.7 and PyTorch (http://pytorch.org/).

The model is implemented and tested on PyTorch version 0.3.1 (http://pytorch.org/docs/0.3.1/).

Hardware Requirements

The model is fast on a GPU unit with CUDA + cuDNN deep learning libraries.

Data Requirements

First you need to obtain word embeddings.

For English, we use 100-dimensions Glove embeddings (https://nlp.stanford.edu/projects/glove/).

The preprocessed version of the embeddings can be downloaded from the following link: https://goo.gl/8D87oP

For German, we obtain and utilize the 64-dimensions German embeddings of https://arxiv.org/abs/1603.01360.

The preprocessed version of the embeddings can be downloaded from the following link: https://goo.gl/U8dQAJ

Running Configurations

All configurations are manually set via the config.py file.

Training Instructions

python tagger.py train <path to save model>

Example:

> mkdir ./saved_models
> python tagger.py train ./saved_models/

Testing Instructions

python tagger.py test <path to restore model> <input file path> <output file path>

Example:

> python tagger.py test ./saved_models/ ./data/dev.raw ./saved_models/dev.predicted
> python tagger.py test ./saved_models/ ./data/test.raw ./saved_models/test.predicted

License

MIT license.

Please Cite

@article{DBLP:journals/corr/abs-1810-00428,
  author    = {Saeed Najafi and
               Colin Cherry and
               Grzegorz Kondrak},
  title     = {Efficient Sequence Labeling with Actor-Critic Training},
  journal   = {CoRR},
  volume    = {abs/1810.00428},
  year      = {2018},
  url       = {http://arxiv.org/abs/1810.00428},
  archivePrefix = {arXiv},
  eprint    = {1810.00428},
  timestamp = {Tue, 30 Oct 2018 10:49:09 +0100},
  biburl    = {https://dblp.org/rec/bib/journals/corr/abs-1810-00428},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}