def compare_templates(self):
for i, t in enumerate([
brill.nltkdemo18(),
brill.nltkdemo18plus(),
brill.brill24(),
brill.fntbl37()
]):
print "\nTEMPLATE {}==================\n".format(i)
self.train(templates=t)
def get_braupt_tagger():
conll_sents = nltk.corpus.conll2000.tagged_sents()
# conll_sents = nltk.corpus.conll2002.tagged_sents()
word_patterns = get_word_patterns()
raubt_tagger = backoff_tagger(conll_sents, [
nltk.tag.AffixTagger, nltk.tag.UnigramTagger, nltk.tag.BigramTagger,
nltk.tag.TrigramTagger
],
backoff=nltk.tag.RegexpTagger(word_patterns))
templates = brill.brill24()
trainer = BrillTaggerTrainer(raubt_tagger, templates)
braubt_tagger = trainer.train(conll_sents, max_rules=100, min_score=3)
return braubt_tagger
def postag(
templates=None,
tagged_data=None,
num_sents=1000,
max_rules=300,
min_score=3,
min_acc=None,
train=0.8,
trace=3,
randomize=False,
ruleformat="str",
incremental_stats=False,
template_stats=False,
error_output=None,
serialize_output=None,
learning_curve_output=None,
learning_curve_take=300,
baseline_backoff_tagger=None,
separate_baseline_data=False,
cache_baseline_tagger=None):
"""
Brill Tagger Demonstration
:param templates: how many sentences of training and testing data to use
:type templates: list of Template
:param tagged_data: maximum number of rule instances to create
:type tagged_data: C{int}
:param num_sents: how many sentences of training and testing data to use
:type num_sents: C{int}
:param max_rules: maximum number of rule instances to create
:type max_rules: C{int}
:param min_score: the minimum score for a rule in order for it to be considered
:type min_score: C{int}
:param min_acc: the minimum score for a rule in order for it to be considered
:type min_acc: C{float}
:param train: the fraction of the the corpus to be used for training (1=all)
:type train: C{float}
:param trace: the level of diagnostic tracing output to produce (0-4)
:type trace: C{int}
:param randomize: whether the training data should be a random subset of the corpus
:type randomize: C{bool}
:param ruleformat: rule output format, one of "str", "repr", "verbose"
:type ruleformat: C{str}
:param incremental_stats: if true, will tag incrementally and collect stats for each rule (rather slow)
:type incremental_stats: C{bool}
:param template_stats: if true, will print per-template statistics collected in training and (optionally) testing
:type template_stats: C{bool}
:param error_output: the file where errors will be saved
:type error_output: C{string}
:param serialize_output: the file where the learned tbl tagger will be saved
:type serialize_output: C{string}
:param learning_curve_output: filename of plot of learning curve(s) (train and also test, if available)
:type learning_curve_output: C{string}
:param learning_curve_take: how many rules plotted
:type learning_curve_take: C{int}
:param baseline_backoff_tagger: the file where rules will be saved
:type baseline_backoff_tagger: tagger
:param separate_baseline_data: use a fraction of the training data exclusively for training baseline
:type separate_baseline_data: C{bool}
:param cache_baseline_tagger: cache baseline tagger to this file (only interesting as a temporary workaround to get
deterministic output from the baseline unigram tagger between python versions)
:type cache_baseline_tagger: C{string}
Note on separate_baseline_data: if True, reuse training data both for baseline and rule learner. This
is fast and fine for a demo, but is likely to generalize worse on unseen data.
Also cannot be sensibly used for learning curves on training data (the baseline will be artificially high).
"""
# defaults
baseline_backoff_tagger = baseline_backoff_tagger or REGEXP_TAGGER
if templates is None:
from nltk.tag.brill import describe_template_sets, brill24
# some pre-built template sets taken from typical systems or publications are
# available. Print a list with describe_template_sets()
# for instance:
templates = brill24()
(training_data, baseline_data, gold_data, testing_data) = \
_demo_prepare_data(tagged_data, train, num_sents, randomize, separate_baseline_data)
# creating (or reloading from cache) a baseline tagger (unigram tagger)
# this is just a mechanism for getting deterministic output from the baseline between
# python versions
if cache_baseline_tagger:
if not os.path.exists(cache_baseline_tagger):
baseline_tagger = UnigramTagger(baseline_data, backoff=baseline_backoff_tagger)
with open(cache_baseline_tagger, 'w') as print_rules:
pickle.dump(baseline_tagger, print_rules)
print("Trained baseline tagger, pickled it to {0}".format(cache_baseline_tagger))
with open(cache_baseline_tagger, "r") as print_rules:
baseline_tagger= pickle.load(print_rules)
print("Reloaded pickled tagger from {0}".format(cache_baseline_tagger))
else:
baseline_tagger = UnigramTagger(baseline_data, backoff=baseline_backoff_tagger)
print("Trained baseline tagger")
if gold_data:
print(" Accuracy on test set: {0:0.4f}".format(baseline_tagger.evaluate(gold_data)))
# creating a Brill tagger
tbrill = time.time()
trainer = BrillTaggerTrainer(baseline_tagger, templates, trace, ruleformat=ruleformat)
print("Training tbl tagger...")
brill_tagger = trainer.train(training_data, max_rules, min_score, min_acc)
print("Trained tbl tagger in {0:0.2f} seconds".format(time.time() - tbrill))
if gold_data:
print(" Accuracy on test set: %.4f" % brill_tagger.evaluate(gold_data))
# printing the learned rules, if learned silently
if trace == 1:
print("\nLearned rules: ")
for (ruleno, rule) in enumerate(brill_tagger.rules(),1):
print("{0:4d} {1:s}".format(ruleno, rule.format(ruleformat)))
# printing template statistics (optionally including comparison with the training data)
# note: if not separate_baseline_data, then baseline accuracy will be artificially high
if incremental_stats:
print("Incrementally tagging the test data, collecting individual rule statistics")
(taggedtest, teststats) = brill_tagger.batch_tag_incremental(testing_data, gold_data)
print(" Rule statistics collected")
if not separate_baseline_data:
print("WARNING: train_stats asked for separate_baseline_data=True; the baseline "
"will be artificially high")
trainstats = brill_tagger.train_stats()
if template_stats:
brill_tagger.print_template_statistics(teststats)
if learning_curve_output:
_demo_plot(learning_curve_output, teststats, trainstats, take=learning_curve_take)
print("Wrote plot of learning curve to {0}".format(learning_curve_output))
else:
print("Tagging the test data")
taggedtest = brill_tagger.batch_tag(testing_data)
if template_stats:
brill_tagger.print_template_statistics()
# writing error analysis to file
if error_output is not None:
with open(error_output, 'w') as f:
f.write('Errors for Brill Tagger %r\n\n' % serialize_output)
for e in error_list(gold_data, taggedtest):
f.write(e+'\n')
print("Wrote tagger errors including context to {0}".format(error_output))
# serializing the tagger to a pickle file and reloading (just to see it works)
if serialize_output is not None:
taggedtest = brill_tagger.batch_tag(testing_data)
with open(serialize_output, 'w') as print_rules:
pickle.dump(brill_tagger, print_rules)
print("Wrote pickled tagger to {0}".format(serialize_output))
with open(serialize_output, "r") as print_rules:
brill_tagger_reloaded = pickle.load(print_rules)
print("Reloaded pickled tagger from {0}".format(serialize_output))
taggedtest_reloaded = brill_tagger.batch_tag(testing_data)
if taggedtest == taggedtest_reloaded:
print("Reloaded tagger tried on test set, results identical")
else:
print("PROBLEM: Reloaded tagger gave different results on test set")
def postag(
templates=None,
tagged_data=None,
num_sents=1000,
max_rules=300,
min_score=3,
min_acc=None,
train=0.8,
trace=3,
randomize=False,
ruleformat="str",
incremental_stats=False,
template_stats=False,
error_output=None,
serialize_output=None,
learning_curve_output=None,
learning_curve_take=300,
baseline_backoff_tagger=None,
separate_baseline_data=False,
cache_baseline_tagger=None):
"""
Brill Tagger Demonstration
:param templates: how many sentences of training and testing data to use
:type templates: list of Template
:param tagged_data: maximum number of rule instances to create
:type tagged_data: C{int}
:param num_sents: how many sentences of training and testing data to use
:type num_sents: C{int}
:param max_rules: maximum number of rule instances to create
:type max_rules: C{int}
:param min_score: the minimum score for a rule in order for it to be considered
:type min_score: C{int}
:param min_acc: the minimum score for a rule in order for it to be considered
:type min_acc: C{float}
:param train: the fraction of the the corpus to be used for training (1=all)
:type train: C{float}
:param trace: the level of diagnostic tracing output to produce (0-4)
:type trace: C{int}
:param randomize: whether the training data should be a random subset of the corpus
:type randomize: C{bool}
:param ruleformat: rule output format, one of "str", "repr", "verbose"
:type ruleformat: C{str}
:param incremental_stats: if true, will tag incrementally and collect stats for each rule (rather slow)
:type incremental_stats: C{bool}
:param template_stats: if true, will print per-template statistics collected in training and (optionally) testing
:type template_stats: C{bool}
:param error_output: the file where errors will be saved
:type error_output: C{string}
:param serialize_output: the file where the learned tbl tagger will be saved
:type serialize_output: C{string}
:param learning_curve_output: filename of plot of learning curve(s) (train and also test, if available)
:type learning_curve_output: C{string}
:param learning_curve_take: how many rules plotted
:type learning_curve_take: C{int}
:param baseline_backoff_tagger: the file where rules will be saved
:type baseline_backoff_tagger: tagger
:param separate_baseline_data: use a fraction of the training data exclusively for training baseline
:type separate_baseline_data: C{bool}
:param cache_baseline_tagger: cache baseline tagger to this file (only interesting as a temporary workaround to get
deterministic output from the baseline unigram tagger between python versions)
:type cache_baseline_tagger: C{string}
Note on separate_baseline_data: if True, reuse training data both for baseline and rule learner. This
is fast and fine for a demo, but is likely to generalize worse on unseen data.
Also cannot be sensibly used for learning curves on training data (the baseline will be artificially high).
"""
# defaults
baseline_backoff_tagger = baseline_backoff_tagger or REGEXP_TAGGER
if templates is None:
from nltk.tag.brill import describe_template_sets, brill24
# some pre-built template sets taken from typical systems or publications are
# available. Print a list with describe_template_sets()
# for instance:
templates = brill24()
(training_data, baseline_data, gold_data, testing_data) = \
_demo_prepare_data(tagged_data, train, num_sents, randomize, separate_baseline_data)
# creating (or reloading from cache) a baseline tagger (unigram tagger)
# this is just a mechanism for getting deterministic output from the baseline between
# python versions
if cache_baseline_tagger:
if not os.path.exists(cache_baseline_tagger):
baseline_tagger = UnigramTagger(baseline_data, backoff=baseline_backoff_tagger)
with open(cache_baseline_tagger, 'w') as print_rules:
pickle.dump(baseline_tagger, print_rules)
print("Trained baseline tagger, pickled it to {0}".format(cache_baseline_tagger))
with open(cache_baseline_tagger, "r") as print_rules:
baseline_tagger= pickle.load(print_rules)
print("Reloaded pickled tagger from {0}".format(cache_baseline_tagger))
else:
baseline_tagger = UnigramTagger(baseline_data, backoff=baseline_backoff_tagger)
print("Trained baseline tagger")
if gold_data:
print(" Accuracy on test set: {0:0.4f}".format(baseline_tagger.evaluate(gold_data)))
# creating a Brill tagger
tbrill = time.time()
trainer = BrillTaggerTrainer(baseline_tagger, templates, trace, ruleformat=ruleformat)
print("Training tbl tagger...")
brill_tagger = trainer.train(training_data, max_rules, min_score, min_acc)
print("Trained tbl tagger in {0:0.2f} seconds".format(time.time() - tbrill))
if gold_data:
print(" Accuracy on test set: %.4f" % brill_tagger.evaluate(gold_data))
# printing the learned rules, if learned silently
if trace == 1:
print("\nLearned rules: ")
for (ruleno, rule) in enumerate(brill_tagger.rules(),1):
print("{0:4d} {1:s}".format(ruleno, rule.format(ruleformat)))
# printing template statistics (optionally including comparison with the training data)
# note: if not separate_baseline_data, then baseline accuracy will be artificially high
if incremental_stats:
print("Incrementally tagging the test data, collecting individual rule statistics")
(taggedtest, teststats) = brill_tagger.batch_tag_incremental(testing_data, gold_data)
print(" Rule statistics collected")
if not separate_baseline_data:
print("WARNING: train_stats asked for separate_baseline_data=True; the baseline "
"will be artificially high")
trainstats = brill_tagger.train_stats()
if template_stats:
brill_tagger.print_template_statistics(teststats)
if learning_curve_output:
_demo_plot(learning_curve_output, teststats, trainstats, take=learning_curve_take)
print("Wrote plot of learning curve to {0}".format(learning_curve_output))
else:
print("Tagging the test data")
taggedtest = brill_tagger.tag_sents(testing_data)
if template_stats:
brill_tagger.print_template_statistics()
# writing error analysis to file
if error_output is not None:
with open(error_output, 'w') as f:
f.write('Errors for Brill Tagger %r\n\n' % serialize_output)
f.write(u'\n'.join(error_list(gold_data, taggedtest)).encode('utf-8') + '\n')
print("Wrote tagger errors including context to {0}".format(error_output))
# serializing the tagger to a pickle file and reloading (just to see it works)
if serialize_output is not None:
taggedtest = brill_tagger.tag_sents(testing_data)
with open(serialize_output, 'w') as print_rules:
pickle.dump(brill_tagger, print_rules)
print("Wrote pickled tagger to {0}".format(serialize_output))
with open(serialize_output, "r") as print_rules:
brill_tagger_reloaded = pickle.load(print_rules)
print("Reloaded pickled tagger from {0}".format(serialize_output))
taggedtest_reloaded = brill_tagger.tag_sents(testing_data)
if taggedtest == taggedtest_reloaded:
print("Reloaded tagger tried on test set, results identical")
else:
print("PROBLEM: Reloaded tagger gave different results on test set")
(r'.*ed$', 'VBD'),
(r'.*ness$', 'NN'),
(r'.*ment$', 'NN'),
(r'.*ful$', 'JJ'),
(r'.*ious$', 'JJ'),
(r'.*ble$', 'JJ'),
(r'.*ic$', 'JJ'),
(r'.*ive$', 'JJ'),
(r'.*ic$', 'JJ'),
(r'.*est$', 'JJ'),
(r'^a$', 'PREP'),
]
raubt_tagger = backoff_tagger(train_sents, [nltk.tag.AffixTagger, nltk.tag.UnigramTagger, nltk.tag.BigramTagger, nltk.tag.TrigramTagger], backoff=nltk.tag.RegexpTagger(word_patterns))
Service.logger.debug("Raubt tagger ready!")
trainer = brill_trainer.BrillTaggerTrainer(raubt_tagger, brill.brill24())
pos_tagger = trainer.train(train_sents, max_rules=500, min_score=3)
Service.logger.debug("Storing POS tagger model to %s" % storedModel)
modelFile = open(storedModel, "wb")
cPickle.dump(pos_tagger, modelFile, -1)
modelFile.close()
# Define harvesting process
# 'content' is the text to harvest terms from
# 'lang' is the target language
# 'prods' is a list of product codes on NeXLT
# example for a string that is in an Inventor product: Getterms("Define user-specific information %d", 'jpn', 'INV'])
def Getterms(content, lang, prods, returnJSON):
def entrenar_bill(initial_tagger,tagger_name):
brill_tagger = BrillTaggerTrainer(initial_tagger=initial_tagger,templates=brill.brill24())
tagger1 = brill_tagger.train(train_reducido[:1000])
evaluacion = tagger1.evaluate(test_reducido[:1000])
xlabels.append("Brill Tagger "+tagger_name)
accuracys.append(evaluacion)
def compare_templates(self):
for i, t in enumerate([brill.nltkdemo18(), brill.nltkdemo18plus(),
brill.brill24(), brill.fntbl37()]):
print "\nTEMPLATE {}==================\n".format(i)
self.train(templates=t)
