Yes, it’s possible to train a recognizer for whatever entity types you need. First you must create your own training corpus, then you can train a classifier on it that will recognize the entities you need.
The nltk does not provide any facilities for annotating a new corpus-- unless you count the existing POS tagger and named entity recognizer, which you can use to pre-process a corpus before you correct the results and annotate your own categories
by hand. Once you have a suitable training corpus in IOB format, take a look at chapters 6 and 7 of the NLTK book. Chapter 7 shows you in detail how to use a recognizer for a chunking task. In chapter 6, you’ll read about how to select and apply features for
any statistical classifier.
Alexis