nltk.tag.brill_trainer.BrillTaggerTrainer

Construct a Brill tagger from a baseline tagger and a set of templates

Trains the Brill tagger on the corpus train_sents, producing at most max_rules transformations, each of which reduces the net number of errors in the corpus by at least min_score, and each of which has accuracy not lower than min_acc.

#imports >>> from nltk.tbl.template import Template >>> from nltk.tag.brill import Pos, Word >>> from nltk.tag import untag, RegexpTagger, BrillTaggerTrainer

#some data >>> from nltk.corpus import treebank >>> training_data = treebank.tagged_sents()[:100] >>> baseline_data = treebank.tagged_sents()[100:200] >>> gold_data = treebank.tagged_sents()[200:300] >>> testing_data = [untag(s) for s in gold_data]

>>> backoff = RegexpTagger([
... (r'^-?[0-9]+(.[0-9]+)?$', 'CD'),   # cardinal numbers
... (r'(The|the|A|a|An|an)$', 'AT'),   # articles
... (r'.*able$', 'JJ'),                # adjectives
... (r'.*ness$', 'NN'),                # nouns formed from adjectives
... (r'.*ly$', 'RB'),                  # adverbs
... (r'.*s$', 'NNS'),                  # plural nouns
... (r'.*ing$', 'VBG'),                # gerunds
... (r'.*ed$', 'VBD'),                 # past tense verbs
... (r'.*', 'NN')                      # nouns (default)
... ])

>>> baseline = backoff #see NOTE1

>>> baseline.evaluate(gold_data) #doctest: +ELLIPSIS
0.2450142...

#templates >>> Template._cleartemplates() #clear any templates created in earlier tests >>> templates = [Template(Pos([-1])), Template(Pos([-1]), Word([0]))]

#construct a BrillTaggerTrainer >>> tt = BrillTaggerTrainer(baseline, templates, trace=3)

>>> tagger1 = tt.train(training_data, max_rules=10)
TBL train (fast) (seqs: 100; tokens: 2417; tpls: 2; min score: 2; min acc: None)
Finding initial useful rules...
    Found 845 useful rules.
<BLANKLINE>
           B      |
   S   F   r   O  |        Score = Fixed - Broken
   c   i   o   t  |  R     Fixed = num tags changed incorrect -> correct
   o   x   k   h  |  u     Broken = num tags changed correct -> incorrect
   r   e   e   e  |  l     Other = num tags changed incorrect -> incorrect
   e   d   n   r  |  e
------------------+-------------------------------------------------------
 132 132   0   0  | AT->DT if Pos:NN@[-1]
  85  85   0   0  | NN->, if Pos:NN@[-1] & Word:,@[0]
  69  69   0   0  | NN->. if Pos:NN@[-1] & Word:.@[0]
  51  51   0   0  | NN->IN if Pos:NN@[-1] & Word:of@[0]
  47  63  16 161  | NN->IN if Pos:NNS@[-1]
  33  33   0   0  | NN->TO if Pos:NN@[-1] & Word:to@[0]
  26  26   0   0  | IN->. if Pos:NNS@[-1] & Word:.@[0]
  24  24   0   0  | IN->, if Pos:NNS@[-1] & Word:,@[0]
  22  27   5  24  | NN->-NONE- if Pos:VBD@[-1]
  17  17   0   0  | NN->CC if Pos:NN@[-1] & Word:and@[0]

>>> tagger1.rules()[1:3]
(Rule('001', 'NN', ',', [(Pos([-1]),'NN'), (Word([0]),',')]), Rule('001', 'NN', '.', [(Pos([-1]),'NN'), (Word([0]),'.')]))

>>> train_stats = tagger1.train_stats()
>>> [train_stats[stat] for stat in ['initialerrors', 'finalerrors', 'rulescores']]
[1775, 1269, [132, 85, 69, 51, 47, 33, 26, 24, 22, 17]]

>>> tagger1.print_template_statistics(printunused=False)
TEMPLATE STATISTICS (TRAIN)  2 templates, 10 rules)
TRAIN (   2417 tokens) initial  1775 0.2656 final:  1269 0.4750
#ID | Score (train) |  #Rules     | Template
--------------------------------------------
001 |   305   0.603 |   7   0.700 | Template(Pos([-1]),Word([0]))
000 |   201   0.397 |   3   0.300 | Template(Pos([-1]))
<BLANKLINE>
<BLANKLINE>

>>> tagger1.evaluate(gold_data) # doctest: +ELLIPSIS
0.43996...

>>> tagged, test_stats = tagger1.batch_tag_incremental(testing_data, gold_data)

>>> tagged[33][12:] == [('foreign', 'IN'), ('debt', 'NN'), ('of', 'IN'), ('$', 'NN'), ('64', 'CD'),
... ('billion', 'NN'), ('*U*', 'NN'), ('--', 'NN'), ('the', 'DT'), ('third-highest', 'NN'), ('in', 'NN'),
... ('the', 'DT'), ('developing', 'VBG'), ('world', 'NN'), ('.', '.')]
True

>>> [test_stats[stat] for stat in ['initialerrors', 'finalerrors', 'rulescores']]
[1855, 1376, [100, 85, 67, 58, 27, 36, 27, 16, 31, 32]]

# a high-accuracy tagger >>> tagger2 = tt.train(training_data, max_rules=10, min_acc=0.99) TBL train (fast) (seqs: 100; tokens: 2417; tpls: 2; min score: 2; min acc: 0.99) Finding initial useful rules...

Found 845 useful rules.

<BLANKLINE>

B |

S F r O | Score = Fixed - Broken c i o t | R Fixed = num tags changed incorrect -> correct o x k h | u Broken = num tags changed correct -> incorrect r e e e | l Other = num tags changed incorrect -> incorrect e d n r | e

------------------+-------------------------------------------------------

132 132 0 0 | AT->DT if Pos:NN@[-1]

85 85 0 0 | NN->, if Pos:NN@[-1] & Word:,@[0] 69 69 0 0 | NN->. if Pos:NN@[-1] & Word:.@[0] 51 51 0 0 | NN->IN if Pos:NN@[-1] & Word:of@[0] 36 36 0 0 | NN->TO if Pos:NN@[-1] & Word:to@[0] 26 26 0 0 | NN->. if Pos:NNS@[-1] & Word:.@[0] 24 24 0 0 | NN->, if Pos:NNS@[-1] & Word:,@[0] 19 19 0 6 | NN->VB if Pos:TO@[-1] 18 18 0 0 | CD->-NONE- if Pos:NN@[-1] & Word:0@[0] 18 18 0 0 | NN->CC if Pos:NN@[-1] & Word:and@[0]

>>> tagger2.evaluate(gold_data)  # doctest: +ELLIPSIS
0.44159544...
>>> tagger2.rules()[2:4]
(Rule('001', 'NN', '.', [(Pos([-1]),'NN'), (Word([0]),'.')]), Rule('001', 'NN', 'IN', [(Pos([-1]),'NN'), (Word([0]),'of')]))

# NOTE1: (!!FIXME) A far better baseline uses nltk.tag.UnigramTagger, # with a RegexpTagger only as backoff. For instance, # >>> baseline = UnigramTagger(baseline_data, backoff=backoff) # However, as of Nov 2013, nltk.tag.UnigramTagger does not yield consistent results # between python versions. The simplistic backoff above is a workaround to make doctests # get consistent input.