This is a python wrapper for crfsuite, a fast implementation of Conditional Random Fields

• Jake Vanderplas <vanderplas@astro.washington.edu>

Currently the package is set-up only for in-place installation. It requires the crfsuite library to be installed: see https://www.chokkan.org/software/crfsuite/

Once this is installed, simply type make in the head directory.

There are a few basic test scripts in the head directory. test.py will read a small dataset from example_files, then run a basic training and tagging operation. crfsuite_test.sh runs the same operation using the command-line frontend provided by crfsuite. To compare the results of the training and tagging, run compare_output.sh. This will print all the places where the tagging results differ.

This is still a very incomplete wrapper. Search TODO within src/crfsuite.pyx to see some issues that need to be addressed.

There are a few 'features' in crfsuite that make efficient python wrapping difficult.

• Model File Output: as currently written, crfsuite writes the result of a training directly to a binary file. The library is not configured to allow writing the model to memory. This means that a python wrapper must write the model to disk, then read the model into memory before performing any tagging operation. It would be better if the model could be saved directly to a CRFsuite model structure, though when dealing with the very large datasets for which crfsuite is designed, it's clear why the author made the choice he did.
• Memory mapping: as currently written, crfsuite data is not stored in contiguous arrays. This means that there is no way to map a crfsuite data structure to a numpy array, and any input to crfsuite will need to be copied in memory. Addressing this would require significant upstream changes: the crfsuite_item_t structure would have to use an array of floats and an array of ints rather than an array of attribute structures.