Rebuild Brown Corpus tags and rerun tagger evaluations. If you train on treebank it should be fairly accurate when you test on brown.

import nltk.tag as tag word-patterns [ (r'^-2 [0-9]+(. [0-9]+)25', (r ' (The l the l Ala l An I an) $ ' ' 'CD'), ' AT ' ) , # cardinal numbers # articles # adjectives # nouns formed from adjectives # adverbs # plural nouns # gerunds (r'.*ing$ ' , 'VBG ' ), past tense verbs nouns (default) # From the NLTK3 cookbook tag-util.py file def backoff tagger (train sents, taqger classes, backoff-None): for cls in tagger classes: backoff cls (train_sents, backoff-backoff) return backoff inner taggerbackoff tagger (brown train, [tag.AffixTagger, tag.UnigramTagger, Cag Bigramlagger tag.TrigramTagger], backoff-tag.RegexpTagger (word_patterns) ) brill_trainer tag.BrillTaggerTrainer (inner_tagger, tag.brill.bril124 )) brill tagger = brill trainer.train (brown train, max rules=400, min sc re=3) print (brill_tagger.evaluate (treebank_test)) print (brill_tagger.evaluate (brown test)) print (brill tagger.evaluate (conll test)) 0.582085042089359 0.9284236339949421 0.6032014228546021 import nltk.tag as tag word-patterns [ (r'^-2 [0-9]+(. [0-9]+)25', (r ' (The l the l Ala l An I an) $ ' ' 'CD'), ' AT ' ) , # cardinal numbers # articles # adjectives # nouns formed from adjectives # adverbs # plural nouns # gerunds (r'.*ing$ ' , 'VBG ' ), past tense verbs nouns (default) # From the NLTK3 cookbook tag-util.py file def backoff tagger (train sents, taqger classes, backoff-None): for cls in tagger classes: backoff cls (train_sents, backoff-backoff) return backoff inner taggerbackoff tagger (brown train, [tag.AffixTagger, tag.UnigramTagger, Cag Bigramlagger tag.TrigramTagger], backoff-tag.RegexpTagger (word_patterns) ) brill_trainer tag.BrillTaggerTrainer (inner_tagger, tag.brill.bril124 )) brill tagger = brill trainer.train (brown train, max rules=400, min sc re=3) print (brill_tagger.evaluate (treebank_test)) print (brill_tagger.evaluate (brown test)) print (brill tagger.evaluate (conll test)) 0.582085042089359 0.9284236339949421 0.6032014228546021