rungsted 1.2.1

Rungsted. An efficient HMM-based structured prediction model for sequential labeling tasks, with extras.

## Rungsted structed perceptron sequential tagger

### Building

Use

python setup.py build_ext --inplace

Building with the above command happens in place, leaving the generated C and C++ files in the source directory for inspection. Changes in dependent modules are unfortunately not picked up by the build system. Whenever you need to start from a clean slate, use the supplied clean.sh script to get rid of the generated files.

The build script requires a recent version of Cython. If you don’t have Cython, it can be installed as below:

pip install cython

### Demo

The repository contains a subset of the part-of-speech tagged Brown corpus. To run the structured perceptron labeler on this dataset, execute:

python src/labeler.py --train data/brown.train --test data/brown.test.vw

Rungsted’s input format is closely modeled on the powerful and flexible format of [Vowpal Wabbit](https://github.com/JohnLangford/vowpal_wabbit/wiki/Input-format), with the exception that Rungsted is perfectly fine with labels that are not integers.

### Datasets

Provided you have a working installation of NLTK, you can recreate the Brown dataset with this command.

python rungsted/datasets/cr_brown_pos_data.py data/brown.train.vw data/brown.test.vw

There is also a script rungsted/datasets/conll_to_vw.py to convert from CONLL-formatted input to Rungsted