CRFs are best known to provide close to state-of-the-art performance for named-entity tagging. This recipe will tell us how to build one of these systems. The recipe assumes that you have read, understood, and played with the Conditional r andom fields – CRF for word/token tagging recipe in Chapter 4, Tagging Words and Tokens, which addresses the underlying technology. Like HMMs, CRFs treat named entity detection as a word-tagging problem, with an interpretation layer that provides chunkings. Unlike HMMs, CRFs use a logistic-regression-based classification approach, which, in turn, allows for random features to be included. Also, there is an excellent tutorial on CRFs that this recipe follows closely (but omits details) at http://alias-i.com/lingpipe/demos/tutorial/crf/read-me.html. There is also a lot of information in the Javadoc.