- In NER task, recently, BiLSTM-CRF Neural Networks are often used, and get the best performance. but I use a simple Neural Networks(
BiLSTM
) and context feature to train data, and get a good performance close to theBiLSTM-CRF
.
pyorch: 0.3.1
python: 3.6.1
cuda: 8.0 (support cuda speed up, can chose)
modify the config file, detail see the Config directory
(1) sh run.sh
(2) python -u main_hyperparams.py --config_file ./Config/config.cfg
BiLSTM + context feature
BiLSTM-CRF
- Now, only support
BiLSTM + context feature
, BiLSTM-CRF will be support later.
- A simple test of the training speed and decoding time on the
CPU
andGPU
,requires only4 seconds
for the decoding time on the GPU. why so fast ? In terms of decoding, batch calculation is performed in some places, so the decoding time is much faster than one sentence.
- updating......
-
if you have any question, you can open a issue or email
bamtercelboo@{gmail.com, 163.com}
. -
if you have any good suggestions, you can PR or email me.