A implementation of Conditional Random Fields (CRFs) with Deep Learning Method.
DeepCRF is a sequene labeling library that uses neural networks and CRFs in Python using Chainer, a flexible deep learning framework.
git clone https://github.com/aonotas/deep-crf.git
cd deep-crf
python setup.py install
train Ma and Hovy (2016) model
$ deep-crf train input_file.txt --delimiter ' ' --model_name bilstm-cnn-crf
$ deep-crf train input_file.txt --delimiter ' ' --model_name bilstm-cnn-crf --n_layer 3
$ 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
$ deep-crf train input_file_multi.txt --delimiter ' ' --model_name bilstm-cnn-crf −−input idx 0,1 −−output idx 2
$ 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 --model_name bilstm-cnn-crf --model_filename bilstm-cnn-crf_adam.model
DeepCRF provides following features.
- Bi-LSTM / Bi-GRU / Bi-RNN
- CNN for character-level representation
- Pre-trained word embedding
- Pre-trained character embedding
- CRFs at output layer
- CoNLL format input/output
- Raw text data input/output
- Training : Your variable files
- Test : Raw text file at command-line
- Evaluation : F-measure, Accuracy
| Model | Accuracy |
|---|---|
| CRFsuit | 96.39 |
| deep-crf | 97.45 |
| dos Santos and Zadrozny (2014) | 97.32 |
| Ma and Hovy (2016) | 97.55 |
| Model | Prec. | Recall | F1 |
|---|---|---|---|
| CRFsuit | 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 |
|---|---|---|---|
| CRFsuit | 93.77 | 93.45 | 93.61 |
| deep-crf | 94.67 | 94.43 | 94.55 |
| Huang et al. (2015) | - | - | 94.46 |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent