Bidirectional Long-Short Term Memory sequence tagger
This is an extended version (structbilty
) of the earlier bi-LSTM tagger by Plank et al., (2016).
If you use this tagger please cite:
@inproceedings{plank-etal-2016,
title = "Multilingual Part-of-Speech Tagging with Bidirectional Long Short-Term Memory Models and Auxiliary Loss",
author = "Plank, Barbara and
S{\o}gaard, Anders and
Goldberg, Yoav",
booktitle = "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)",
month = aug,
year = "2016",
address = "Berlin, Germany",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/P16-2067",
doi = "10.18653/v1/P16-2067",
pages = "412--418",
}
For the version called DsDs, please cite: https://aclanthology.coli.uni-saarland.de/papers/D18-1061/d18-1061
pip3 install --user -r requirements.txt
Training the tagger:
python src/structbilty.py --dynet-mem 1500 --train data/da-ud-train.conllu --iters 10 --model da
Training with patience (requires a dev set):
python src/structbilty.py --dynet-mem 1500 --train data/da-ud-train.conllu --dev data/da-ud-dev.conllu --iters 50 --model da --patience 2
Testing and getting the output predictions:
python src/structbilty.py --model da --test data/da-ud-test.conllu --output predictions/test-da.out
Training and testing in two steps (--model
for both saving and loading):
mkdir -p predictions
python src/structbilty.py --dynet-mem 1500 --train data/da-ud-train.conllu --iters 10 --model da
python src/structbilty.py --model da --test data/da-ud-test.conllu --output predictions/test-da.out
By default, the model uses a softmax
decoder. You can use a CRF for BIO sequence tagging with the --crf
option.
The model uses accuracy as default output. If you use the tagger for NER or similar, make sure to not rely on accuracy but use span-F1 or similar.
The Polyglot embeddings (Al-Rfou et al., 2013) can be downloaded from here (0.6GB)
You can load generic word embeddings by using --embeds WORD_EMBEDS_FILE
(as the Polyglot ones above).
Note that the dimensions of embeddings should match the --in_dim
option.
Bilty also supports loading additional embeddings from the input files. This can be enabled by --embeds_in_file FILE
.
It expects the train/dev/test files to be in the following format:
word1<tab>tag1<tab>emb=val1,val2,val3,...
word2<tab>tag1<tab>emb=val1,val2,val3,...
...
Note that the dimensions of embeddings should match the --embeds_in_file_dim
option.
We also provide scripts to generate these files for four commonly used embeddings types (Polyglot, Fasttext, ELMo and BERT), which can be found in the embeds
folder. If we for example want to use BERT embeddings we need to run the following commands:
python3 embeds/transf.py bert-base-multilingual-cased data/da-ud-train.conllu
python3 embeds/transf.py bert-base-multilingual-cased data/da-ud-dev.conllu
python3 embeds/transf.py bert-base-multilingual-cased data/da-ud-test.conllu
This creates .bert files which can be used as input to Bilty when --embeds_in_file
is enabled.
Similar scripts for Poly are in the embeds
folder. For now the language for most of these is hardcoded in the scripts, please modify *.prep.py
accordingly.
Please note that this option does not support the --raw
option.
You can see the options by running:
python src/structbilty.py --help
A great option is DyNet autobatching (Neubig et al., 2017). It speeds up training considerably ( ~20%). You can activate it with:
python src/structbilty.sh --dynet-autobatch 1
- major refactoring of internal data handling
- renaming to
structbilty
--pred-layer
is no longer required- a single
--model
options handles both saving and loading model parameters - the option of running a CRF has been added
- the tagger can handle additional lexical features (see our DsDs paper, EMNLP 2018) below
- grouping of arguments
simplebilty
is deprecated (still available in the former release)- best to run it on a simple CPU
# default reference
@inproceedings{plank-etal-2016,
title = "Multilingual Part-of-Speech Tagging with Bidirectional Long Short-Term Memory Models and Auxiliary Loss",
author = "Plank, Barbara and
S{\o}gaard, Anders and
Goldberg, Yoav",
booktitle = "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)",
month = aug,
year = "2016",
address = "Berlin, Germany",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/P16-2067",
doi = "10.18653/v1/P16-2067",
pages = "412--418",
}
# for DdDs
@InProceedings{plank-agic:2018,
author = "Plank, Barbara
and Agi{\'{c}}, {\v{Z}}eljko",
title = "Distant Supervision from Disparate Sources for Low-Resource Part-of-Speech Tagging",
booktitle = "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
year = "2018",
publisher = "Association for Computational Linguistics",
pages = "614--620",
location = "Brussels, Belgium",
url = "http://aclweb.org/anthology/D18-1061"
}
You can compile dynet from source. Clone it into a directory of your choice called DYNETDIR
:
mkdir $DYNETDIR
git clone https://github.com/clab/dynet
Follow the instructions in the Dynet documentation (use -DPYTHON
,
see http://dynet.readthedocs.io/en/latest/python.html).
And compile dynet:
cmake .. -DEIGEN3_INCLUDE_DIR=$HOME/tools/eigen/ -DPYTHON=`which python`
(if you have a GPU, use: [note: non-deterministic behavior]):
cmake .. -DEIGEN3_INCLUDE_DIR=$HOME/tools/eigen/ -DPYTHON=`which python` -DBACKEND=cuda
(You may need to set you PYTHONPATH to include Dynet's build/python
)
After successful installation open python and import dynet, you can test if the installation worked with:
>>> import dynet
[dynet] random seed: 2809331847
[dynet] allocating memory: 512MB
[dynet] memory allocation done.
>>> dynet.__version__
2.0