def batch_tag(self, sentences):
# Write the test corpus to a temporary file
(fd, test_file) = mkstemp('.txt', 'test')
self.write_test_corpus(sentences, os.fdopen(fd, 'w'))
try:
# Run mallet on the test file.
stdout, stderr = call_mallet([
self._RUN_CRF, '--model-file',
os.path.abspath(self.crf_info.model_filename), '--test-file',
test_file
],
stdout='pipe')
# Decode the output
labels = self.parse_mallet_output(stdout)
# strip __start__ and __end__
if self.crf_info.add_start_state and self.crf_info.add_end_state:
labels = [labs[1:-1] for labs in labels]
elif self.crf_info.add_start_state:
labels = [labs[1:] for labs in labels]
elif self.crf_info.add_end_state:
labels = [labs[:-1] for labs in labels]
# Combine the labels and the original sentences.
return [
zip(sent, label) for (sent, label) in zip(sentences, labels)
]
finally:
os.remove(test_file)
def batch_tag(self, sentences):
# Write the test corpus to a temporary file
(fd, test_file) = mkstemp('.txt', 'test')
self.write_test_corpus(sentences, os.fdopen(fd, 'w'))
try:
# Run mallet on the test file.
stdout, stderr = call_mallet([self._RUN_CRF,
'--model-file', os.path.abspath(self.crf_info.model_filename),
'--test-file', test_file], stdout='pipe')
# Decode the output
labels = self.parse_mallet_output(stdout)
# strip __start__ and __end__
if self.crf_info.add_start_state and self.crf_info.add_end_state:
labels = [labs[1:-1] for labs in labels]
elif self.crf_info.add_start_state:
labels = [labs[1:] for labs in labels]
elif self.crf_info.add_end_state:
labels = [labs[:-1] for labs in labels]
# Combine the labels and the original sentences.
return [zip(sent, label) for (sent,label) in
zip(sentences, labels)]
finally:
os.remove(test_file)
def train(cls,
feature_detector,
corpus,
filename=None,
weight_groups=None,
gaussian_variance=1,
default_label='O',
transduction_type='VITERBI',
max_iterations=500,
add_start_state=True,
add_end_state=True,
trace=1):
"""
Train a new linear chain CRF tagger based on the given corpus
of training sequences. This tagger will be backed by a I{crf
model file}, containing both a serialized Mallet model and
information about the CRF's structure. This crf model file
will I{not} be automatically deleted -- if you wish to delete
it, you must delete it manually. The filename of the model
file for a MalletCRF C{crf} is available as C{crf.filename}.
@type corpus: C{list} of C{tuple}
@param corpus: Training data, represented as a list of
sentences, where each sentence is a list of (token, tag)
tuples.
@type filename: C{str}
@param filename: The filename that should be used for the crf
model file that backs the new C{MalletCRF}. If no
filename is given, then a new filename will be chosen
automatically.
@type weight_groups: C{list} of L{CRFInfo.WeightGroup}
@param weight_groups: Specifies how input-features should
be mapped to joint-features. See L{CRFInfo.WeightGroup}
for more information.
@type gaussian_variance: C{float}
@param gaussian_variance: The gaussian variance of the prior
that should be used to train the new CRF.
@type default_label: C{str}
@param default_label: The "label for initial context and
uninteresting tokens" (from Mallet's SimpleTagger.java.)
It's unclear whether this currently has any effect.
@type transduction_type: C{str}
@param transduction_type: The type of transduction used by
the CRF. Can be VITERBI, VITERBI_FBEAM, VITERBI_BBEAM,
VITERBI_FBBEAM, or VITERBI_FBEAMKL.
@type max_iterations: C{int}
@param max_iterations: The maximum number of iterations that
should be used for training the CRF.
@type add_start_state: C{bool}
@param add_start_state: If true, then NLTK will add a special
start state, named C{'__start__'}. The initial cost for
the start state will be set to 0; and the initial cost for
all other states will be set to +inf.
@type add_end_state: C{bool}
@param add_end_state: If true, then NLTK will add a special
end state, named C{'__end__'}. The final cost for the end
state will be set to 0; and the final cost for all other
states will be set to +inf.
@type trace: C{int}
@param trace: Controls the verbosity of trace output generated
while training the CRF. Higher numbers generate more verbose
output.
"""
t0 = time.time() # Record starting time.
# If they did not supply a model filename, then choose one.
if filename is None:
(fd, filename) = mkstemp('.crf', 'model')
os.fdopen(fd).close()
# Ensure that the filename ends with '.zip'
if not filename.endswith('.crf'):
filename += '.crf'
if trace >= 1:
print '[MalletCRF] Training a new CRF: %s' % filename
# Create crf-info object describing the new CRF.
crf_info = MalletCRF._build_crf_info(corpus, gaussian_variance,
default_label, max_iterations,
transduction_type, weight_groups,
add_start_state, add_end_state,
filename, feature_detector)
# Create a zipfile, and write crf-info to it.
if trace >= 2:
print '[MalletCRF] Adding crf-info.xml to %s' % filename
zf = zipfile.ZipFile(filename, mode='w')
zf.writestr('crf-info.xml', crf_info.toxml() + '\n')
zf.close()
# Create the CRF object.
crf = MalletCRF(filename, feature_detector)
# Write the Training corpus to a temporary file.
if trace >= 2:
print '[MalletCRF] Writing training corpus...'
(fd, train_file) = mkstemp('.txt', 'train')
crf.write_training_corpus(corpus, os.fdopen(fd, 'w'))
try:
if trace >= 1:
print '[MalletCRF] Calling mallet to train CRF...'
cmd = [
MalletCRF._TRAIN_CRF, '--model-file',
os.path.abspath(filename), '--train-file', train_file
]
if trace > 3:
call_mallet(cmd)
else:
p = call_mallet(cmd,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
blocking=False)
MalletCRF._filter_training_output(p, trace)
finally:
# Delete the temp file containing the training corpus.
os.remove(train_file)
if trace >= 1:
print '[MalletCRF] Training complete.'
print '[MalletCRF] Model stored in: %s' % filename
if trace >= 2:
dt = time.time() - t0
print '[MalletCRF] Total training time: %d seconds' % dt
# Return the completed CRF.
return crf
def train(cls, feature_detector, corpus, filename=None,
weight_groups=None, gaussian_variance=1, default_label='O',
transduction_type='VITERBI', max_iterations=500,
add_start_state=True, add_end_state=True, trace=1):
"""
Train a new linear chain CRF tagger based on the given corpus
of training sequences. This tagger will be backed by a I{crf
model file}, containing both a serialized Mallet model and
information about the CRF's structure. This crf model file
will I{not} be automatically deleted -- if you wish to delete
it, you must delete it manually. The filename of the model
file for a MalletCRF C{crf} is available as C{crf.filename}.
@type corpus: C{list} of C{tuple}
@param corpus: Training data, represented as a list of
sentences, where each sentence is a list of (token, tag)
tuples.
@type filename: C{str}
@param filename: The filename that should be used for the crf
model file that backs the new C{MalletCRF}. If no
filename is given, then a new filename will be chosen
automatically.
@type weight_groups: C{list} of L{CRFInfo.WeightGroup}
@param weight_groups: Specifies how input-features should
be mapped to joint-features. See L{CRFInfo.WeightGroup}
for more information.
@type gaussian_variance: C{float}
@param gaussian_variance: The gaussian variance of the prior
that should be used to train the new CRF.
@type default_label: C{str}
@param default_label: The "label for initial context and
uninteresting tokens" (from Mallet's SimpleTagger.java.)
It's unclear whether this currently has any effect.
@type transduction_type: C{str}
@param transduction_type: The type of transduction used by
the CRF. Can be VITERBI, VITERBI_FBEAM, VITERBI_BBEAM,
VITERBI_FBBEAM, or VITERBI_FBEAMKL.
@type max_iterations: C{int}
@param max_iterations: The maximum number of iterations that
should be used for training the CRF.
@type add_start_state: C{bool}
@param add_start_state: If true, then NLTK will add a special
start state, named C{'__start__'}. The initial cost for
the start state will be set to 0; and the initial cost for
all other states will be set to +inf.
@type add_end_state: C{bool}
@param add_end_state: If true, then NLTK will add a special
end state, named C{'__end__'}. The final cost for the end
state will be set to 0; and the final cost for all other
states will be set to +inf.
@type trace: C{int}
@param trace: Controls the verbosity of trace output generated
while training the CRF. Higher numbers generate more verbose
output.
"""
t0 = time.time() # Record starting time.
# If they did not supply a model filename, then choose one.
if filename is None:
(fd, filename) = mkstemp('.crf', 'model')
os.fdopen(fd).close()
# Ensure that the filename ends with '.zip'
if not filename.endswith('.crf'):
filename += '.crf'
if trace >= 1:
print '[MalletCRF] Training a new CRF: %s' % filename
# Create crf-info object describing the new CRF.
crf_info = MalletCRF._build_crf_info(
corpus, gaussian_variance, default_label, max_iterations,
transduction_type, weight_groups, add_start_state,
add_end_state, filename, feature_detector)
# Create a zipfile, and write crf-info to it.
if trace >= 2:
print '[MalletCRF] Adding crf-info.xml to %s' % filename
zf = zipfile.ZipFile(filename, mode='w')
zf.writestr('crf-info.xml', crf_info.toxml()+'\n')
zf.close()
# Create the CRF object.
crf = MalletCRF(filename, feature_detector)
# Write the Training corpus to a temporary file.
if trace >= 2:
print '[MalletCRF] Writing training corpus...'
(fd, train_file) = mkstemp('.txt', 'train')
crf.write_training_corpus(corpus, os.fdopen(fd, 'w'))
try:
if trace >= 1:
print '[MalletCRF] Calling mallet to train CRF...'
cmd = [MalletCRF._TRAIN_CRF,
'--model-file', os.path.abspath(filename),
'--train-file', train_file]
if trace > 3:
call_mallet(cmd)
else:
p = call_mallet(cmd, stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
blocking=False)
MalletCRF._filter_training_output(p, trace)
finally:
# Delete the temp file containing the training corpus.
os.remove(train_file)
if trace >= 1:
print '[MalletCRF] Training complete.'
print '[MalletCRF] Model stored in: %s' % filename
if trace >= 2:
dt = time.time()-t0
print '[MalletCRF] Total training time: %d seconds' % dt
# Return the completed CRF.
return crf
