Rule Extraction from A Trained Conditional Random Field Model
Bhuban Mohan Seth, Pushpak Bhattacharyya · 2011
Conditional Random Field (CRF) has proven to be highly successful for sequence labeling problems like part of speech tagging, segmentation etc. However, the model acts like a black box, providing no insight into what is learned. We propose a system for rule extraction from CRF to assist comprehensibility of the model. Experiments on POS tagging and chunking problem in English are performed as case studies. We test the quality of the extracted rule base by implementing a majority voting rule based tagger, which shows maximum precision of 93.9% for POS tagging and 77% for chunking. The obtained rules conform to our linguistic knowledge of English. We also give quantitative comparison of our approach with PART decision list and C4.5 decision tree learner. Comprehensibility of statistical models is our guiding principle.