lancifollia / crf

A pure-Python implementation of the Linear-Chain Conditional Random Fields

crf

A pure python implementation of the Linear-Chain Conditional Random Fields

Dependencies

• Numpy
• Scipy

Usage

You can test this code with CoNLL 2000 Chunking Data.

Training

# format
python3 crf_train.py <train_file> <model_file>

# example
python3 crf_train.py data/chunking_small/small_train.data small_model.json

Test

# format
python3 crf_test.py <test_file> <trained_model_file>

# example
python3 crf_test.py data/chunking_small/small_test.data small_model.json

Benchmark Result

• Data: CoNLL corpus
  • data/chunking_full: original data (8936 sentences)
  • data/chunking_small: sampled data (77 sentences)
• Compared with CRF++
• Use feature set

Accuracy

	crf	CRF++
chunking_full	0.960128	0.960128
chunking_small	0.899072	0.889474

License

MIT

Reference

An Introduction to Conditional Random Fields / Charles Sutton, Andrew McCallum/ 2010

crf's Issues

OOM

I try to train model with myself data(about 20M), but failed with out of memory.

Speed is slow

The speed seems to be pretty slow.