A implementation of Conditional Random Fields (CRFs) with Deep Learning Method.
DeepCRF is a sequence labeling library that uses neural networks and CRFs in Python using Chainer, a flexible deep learning framework.
# if you use Ubuntu
sudo apt install libhdf5-dev
git clone https://github.com/aonotas/deep-crf.git
cd deep-crf
python setup.py install
# if you want to use Chainer v1.24.0
pip install 'chainer==1.24.0'
# if you want to use Chainer v2.1.0
pip install 'chainer==2.1.0'
pip install cupy # if you want to use CUDA
$ deep-crf train input_file.txt --delimiter=' ' --dev_file input_file_dev.txt --save_dir save_model_dir --save_name bilstm-cnn-crf_adam --optimizer adam
Note that --dev_file
means path of development file to use early stopping.
$ cat input_file.txt
Barack B−PERSON
Hussein I−PERSON
Obama E−PERSON
is O
a O
man O
. O
Yuji B−PERSON
Matsumoto E−PERSON
is O
a O
man O
. O
Each line is word
and gold tag
.
One line is represented by word
[ ](space)
gold tag
.
Note that you should put empty line (\n)
between sentences.
This format is called CoNLL format.
$ deep-crf train input_file.txt --delimiter=' ' --n_layer 3 --dev_file input_file_dev.txt --save_dir save_model_dir --save_name bilstm-cnn-crf_adam --optimizer adam
If input file is multiple due to large input files or many lines, please following commands.
Please add this arg : --use_list_files 1
$ deep-crf train input_file_list.txt --delimiter=' ' --n_layer 3 --dev_file input_file_dev.txt --save_dir save_model_dir --save_name bilstm-cnn-crf_adam --optimizer adam --use_list_files 1
$ cat input_file_list.txt
./path_to_file/input_file_1.txt
./path_to_file/input_file_2.txt
./path_to_file/input_file_3.txt
$ deep-crf train input_file.txt --delimiter=' ' --n_layer 3 --word_emb_file ./glove.6B.100d.txt --word_emb_vocab_type replace_all --dev_file input_file_dev.txt
We prepare some vocab mode.
--word_emb_vocab_type
: select from [replace_all, replace_only, additional]replace_all
: Replace training vocab by Glove embeddings's vocab.replace_only
: Replace word embedding exists in training vocab.additional
: Concatenate training vocab and Glove embeddings's vocab.If you want to use word2vec embeddings, please convert Glove format.
$ head glove.6B.100d.txt
the -0.038194 -0.24487 0.72812 -0.39961 0.083172
dog -0.10767 0.11053 0.59812 -0.54361 0.67396
cat -0.33979 0.20941 0.46348 -0.64792 -0.38377
of -0.1529 -0.24279 0.89837 0.16996 0.53516
to -0.1897 0.050024 0.19084 -0.049184 -0.089737
and -0.071953 0.23127 0.023731 -0.50638 0.33923
in 0.085703 -0.22201 0.16569 0.13373 0.38239
$ deep-crf train input_file_multi.txt --delimiter=' ' --input_idx 0,1 --output_idx 2 --dev_file input_file_dev.txt --save_dir save_model_dir --save_name bilstm-cnn-crf_adam_additional --optimizer adam
$ cat input_file_multi.txt
Barack NN B−PERSON
Hussein NN I−PERSON
Obama NN E−PERSON
is VBZ O
a DT O
man NN O
. . O
Yuji NN B−PERSON
Matsumoto NN E−PERSON
is VBZ O
a DT O
man NN O
. . O
Note that --input_idx
means that input features (but word feature must be 0-index) like this example.
(Now developing this multi-task learning mode...)
$ deep-crf train input_file_multi.txt --delimiter ' ' --model_name bilstm-cnn-crf --input idx 0 --output idx 1,2
$ deep-crf predict input_raw_file.txt --delimiter=' ' --model_filename ./save_model_dir/bilstm-cnn-crf_adam_epoch3.model --save_dir save_model_dir --save_name bilstm-cnn-crf_adam --predicted_output predicted.txt
Please use following format when predict
.
$ cat input_raw_file.txt
Barack Hussein Obama is a man .
Yuji Matsumoto is a man .
Note that --model_filename
means saved model file path.
Please set same --save_name
in training step.
$ deep-crf predict input_file_multi.txt --delimiter=' ' --input_idx 0,1 --output_idx 2 --model_filename ./save_model_dir/bilstm-cnn-crf_multi_epoch3.model --save_dir save_model_dir --save_name bilstm-cnn-crf_multi --predicted_output predicted.txt
Note that you must prepare CoNLL format input file when you use additional feature mode in training step.
$ cat input_file_multi.txt
Barack NN B−PERSON
Hussein NN I−PERSON
Obama NN E−PERSON
is VBZ O
a DT O
man NN O
. . O
Yuji NN B−PERSON
Matsumoto NN E−PERSON
is VBZ O
a DT O
man NN O
. . O
$ deep-crf eval gold.txt predicted.txt
$ head gold.txt
O
O
B-LOC
O
O
B-PERSON
cd deep-crf
git pull
python setup.py install
deep-crf train --help
if you got CUDNN ERROR, please let me know in issues.
You can cudnn-off mode with --use_cudnn=0
DeepCRF provides following features.
Model | Accuracy |
---|---|
CRFsuite | 96.39 |
deep-crf | 97.45 |
dos Santos and Zadrozny (2014) | 97.32 |
Ma and Hovy (2016) | 97.55 |
Model | Prec. | Recall | F1 |
---|---|---|---|
CRFsuite | 84.43 | 83.60 | 84.01 |
deep-crf | 90.82 | 91.11 | 90.96 |
Ma and Hovy (2016) | 91.35 | 91.06 | 91.21 |
Model | Prec. | Recall | F1 |
---|---|---|---|
CRFsuite | 93.77 | 93.45 | 93.61 |
deep-crf | 94.67 | 94.43 | 94.55 |
Huang et al. (2015) | - | - | 94.46 |