- All code here is distributed with Apache 2.0 license but
the following dependencies have their own licensing restrictions:
- Depends on Mallet 2.0RC4 or later. The mallet jar is bundled with this but you should see http://mallet.cs.umass.edu for license details etc.
- The sample NER code using this API depends on Apache commons CLI parser (also bundled here but with its own license)
This code allows arbitrary sequence tagging tasks to be defined by introducing an abstract FeatureFactory. If you're implementing a new sequence tagging task all you have to do is to extend the FeatureFactory for that task and call the setFeatureFactory() method on the CRFSequenceLearner object without having to figure out any of the obscure Mallet details. For an example, see how the named entity recognition task is defined in:
src/edu/jhu/nlp/tools/NamedEntityTagger.java (We get around 89 F-score for the CONLL'03 English NER data using the BasicNERFeatureFactory)
Also this code simplifies setting up parallel training in CRFs. All you've to do is call setNumThreads() on the CRFSequenceLearner object. To give an idea, by using four threads on four cores the CRF training time is well under 15 mins for English CoNLL data (non-parallel takes more than 4 hours).
If you're doing something special to the CRF output other than printing it, you've to simply implement the OutputCallback. (DefaultOutputCallback just prints)