def demo():
"""
Non-interactive demonstration of the clusterers with simple 2-D data.
"""
# use a set of tokens with 2D indices
tokens = [Token(FEATURES=Numeric.array([3, 3])),
Token(FEATURES=Numeric.array([1, 2])),
Token(FEATURES=Numeric.array([4, 2])),
Token(FEATURES=Numeric.array([4, 0])),
Token(FEATURES=Numeric.array([2, 3])),
Token(FEATURES=Numeric.array([3, 1]))]
# test k-means using the euclidean distance metric, 2 means and repeat
# clustering 10 times with random seeds
clusterer = KMeansClusterer(2, euclidean_distance, repeats=10)
clusterer.cluster(tokens, True)
print 'using clusterer', clusterer
print 'clustered', str(tokens)[:60], '...'
# classify a new token
token = Token(FEATURES=Numeric.array([3, 3]))
print 'classify(%s)' % token,
clusterer.classify(token)
print token
# test the GAAC clusterer with 4 clusters
clusterer = GroupAverageAgglomerativeClusterer(4)
print 'using clusterer', clusterer
clusterer.cluster(tokens, True)
#print 'clustered', str(tokens)[:60], '...'
print 'clustered', tokens
# show the dendogram
clusterer.dendogram().show()
# classify a new token
token = Token(FEATURES=Numeric.array([3, 3]))
print 'classify(%s)' % token,
clusterer.classify(token)
print token
print
# test the EM clusterer with means given by k-means (2) and
# dimensionality reduction
clusterer = KMeansClusterer(2, euclidean_distance, svd_dimensions=1)
clusterer.cluster(tokens)
means = clusterer.means()
clusterer = ExpectationMaximizationClusterer(means, svd_dimensions=1)
clusterer.cluster(tokens, True)
print 'using clusterer', clusterer
print 'clustered', str(tokens)[:60], '...'
# classify a new token
token = Token(FEATURES=Numeric.array([3, 3]))
print 'classify(%s)' % token,
clusterer.classify(token)
print token
# show the classification probabilities
token = Token(FEATURES=Numeric.array([2.2, 2]))
print 'classification_probdist(%s)' % token
clusterer.classification_probdist(token)
for sample in token['CLUSTER_PROBDIST'].samples():
print '%s => %.0f%%' % (sample,
token['CLUSTER_PROBDIST'].prob(sample) *100)
def demo_em():
# example from figure 14.10, page 519, Manning and Schutze
tokens = [
Token(FEATURES=Numeric.array(f))
for f in [[0.5, 0.5], [1.5, 0.5], [1, 3]]
]
means = [[4, 2], [4, 2.01]]
clusterer = ExpectationMaximizationClusterer(means, bias=0.1)
clusterer.cluster(tokens, True, trace=True)
print 'clustered', tokens
for c in range(2):
print 'cluster %d' % c
print 'prior', clusterer._priors[c]
print 'mean ', clusterer._means[c]
print 'covar', clusterer._covariance_matrices[c]
# classify a new token
token = Token(FEATURES=Numeric.array([2, 2]))
print 'classify(%s)' % token,
clusterer.classify(token)
print token
# show the classification probabilities
token = Token(FEATURES=Numeric.array([2, 2]))
print 'classification_probdist(%s)' % token
clusterer.classification_probdist(token)
for sample in token['CLUSTER_PROBDIST'].samples():
print '%s => %.0f%%' % (sample,
token['CLUSTER_PROBDIST'].prob(sample) * 100)
def raw_tag(self, words):
SUBTOKENS = self.property('SUBTOKENS')
TEXT = self.property('TEXT')
TAG = self.property('TAG')
subtoks = [Token({TEXT:w}) for w in words]
token = Token({SUBTOKENS:subtoks})
self.tag(token)
return [token[TAG] for token in token[SUBTOKENS]]
def __init__(self, classifier, labeled_tokens):
"""
Entry conf[i][j] is the number of times a document with label i
was given label j.
"""
assert _chktype(1, classifier, ClassifierI)
assert _chktype(2, labeled_tokens, [Token], (Token, ))
try:
import Numeric
except:
raise ImportError('ConfusionMatrix requires Numeric')
# Extract the labels.
ldict = {}
for ltok in labeled_tokens:
ldict[ltok.type().label()] = 1
labels = ldict.keys()
# Construct a label->index dictionary
indices = {}
for i in range(len(labels)):
indices[labels[i]] = i
confusion = Numeric.zeros((len(labels), len(labels)))
for ltok in labeled_tokens:
utok = Token(ltok.type().text(), ltok.loc())
ctok = classifier.classify(utok)
confusion[indices[ltok.type().label()],
indices[ctok.type().label()]] += 1
self._labels = labels
self._confusion = confusion
self._max_conf = max(Numeric.resize(confusion, (len(labels)**2, )))
def parseToken(self,text,
interactive=0,trace=1,draw=0,print_parses=1,cumStats=None,chunker=None,trueTree=None):
if chunker == None: chunker = self.period_chunker
if self.stats:
dirStats = parseStats()
else:
Parses = []
chunker.parse(text)
for sent in text['TREE']:
if not isinstance(sent,Tree): continue
sentToken = Token(WORDS=sent.leaves(),SUBTOKENS=sent.leaves())
print 'parsing',sentToken
if self.stats:
parse_stats(self.Parsers, sentToken, dirStats, trace, trueTree)
print_parse_summary(self.Parsers, dirStats, interactive, draw, print_parses)
# Check for empty parse
if (dirStats.parse_list == [] or not isinstance(dirStats.parse_list[-1][0],Tree)):
chunker2 = None
if chunker != self.period_chunker and sent.count('.')>1: chunker2 = self.period_chunker
elif chunker != self.punct_chunker and sent.count('.')<=1: chunker2 = self.punct_chunker
else: chunker2 = self.unigramTag(dirStats, sentToken)
if chunker2:
if __debug__: print 'No parse, retry',chunker2
dirStats = self.parseToken(sentToken,interactive,trace,draw,print_parses,dirStats,chunker2,trueTree)
###elif __debug__: print 'Good parse, no retry',dirStats.parse_list
else:
parse = self.Parsers[0].get_parse_list(sentToken)
if parse: Parses.append(parse[0])
if self.stats:
if cumStats:
cumStats += dirStats.sum()
return cumStats
else: return dirStats.sum()
else:
return Parses
def log_likelihood(classifier, labeled_tokens):
"""
Evaluate the log likelihood of the given list of labeled
tokens for the given classifier model. This nonpositive float
gives an indication of how well the classifier models the
data. Values closer to zero indicate that it models it more
accurately.
@rtype: C{float}
@return: The log likelihood of C{labeled_tokens} for the given
classifier model.
@param labeled_tokens: The tokens whose log likelihood should
be computed.
@type labeled_tokens: C{list} of (C{Token} with type
C{LabeledText})
"""
assert _chktype(1, classifier, ClassifierI)
assert _chktype(2, labeled_tokens, [Token], (Token, ))
likelihood = 0.0
for ltok in labeled_tokens:
utok = Token(ltok.type().text(), ltok.loc())
label = ltok.type().label()
dist = classifier.distribution_dictionary(utok)
if dist[label] == 0:
# Use some approximation to infinity. What this does
# depends on your system's float implementation.
likelihood -= 1e1000
else:
likelihood += math.log(dist[label])
return likelihood / len(labeled_tokens)
def demo_pos():
from sys import stdout
print 'Training HMM...'
labelled_sequences, tag_set, num_features = load_pos()
trainer = MultiOutputHMMTrainer(tag_set, [[] for x in range(num_features)])
hmm = trainer.train_supervised(
Token(SUBTOKENS=labelled_sequences[100:]),
estimator=lambda fd, bins: LidstoneProbDist(fd, 0.1, bins))
print 'Testing...'
for super_token in labelled_sequences[:3]:
print super_token
print 'HMM >>>'
print hmm.best_path(super_token.exclude('TAG'))
print '-' * 60
count = correct = 0
for super_token in labelled_sequences[:100]:
print '.',
stdout.flush()
pts = hmm.best_path(super_token.exclude('TAG'))
for token, tag in zip(super_token['SUBTOKENS'], pts):
count += 1
if tag == token['TAG']:
correct += 1
print 'accuracy over first', count, 'tokens %.1f' % (100.0 * correct /
count)
def classify(self, unlabeled_token):
# inherit doco
fv_list = self._fd_list.detect(unlabeled_token.type())
fnums = map(lambda x: x[0], fv_list.assignments())
leaf = self._root.traverse(fnums)
label = leaf.label()
return Token(LabeledText(unlabeled_token.type(), label),
unlabeled_token.loc())
def demo():
# demonstrates HMM probability calculation
# example taken from page 381, Huang et al
symbols = ['up', 'down', 'unchanged']
states = ['bull', 'bear', 'static']
def pd(values, samples):
d = {}
for value, item in zip(values, samples):
d[item] = value
return DictionaryProbDist(d)
def cpd(array, conditions, samples):
d = {}
for values, condition in zip(array, conditions):
d[condition] = pd(values, samples)
return DictionaryConditionalProbDist(d)
A = array([[0.6, 0.2, 0.2], [0.5, 0.3, 0.2], [0.4, 0.1, 0.5]], Float64)
A = cpd(A, states, states)
B = array([[0.7, 0.1, 0.2], [0.1, 0.6, 0.3], [0.3, 0.3, 0.4]], Float64)
B = cpd(B, states, symbols)
pi = array([0.5, 0.2, 0.3], Float64)
pi = pd(pi, states)
model = HiddenMarkovModel(symbols=symbols,
states=states,
transitions=A,
outputs=B,
priors=pi)
print 'Testing', model
for test in [['up'] * 2, ['up'] * 5, ['up', 'down', 'up'], ['down'] * 5,
['unchanged'] * 5 + ['up']]:
token = Token(SUBTOKENS=map(lambda t: Token(TEXT=t), test))
print 'Testing with state sequence', test
print 'probability =', model.probability(token)
print 'tagging = ', model.tag(token)
print 'p(tagged) = ', model.probability(token)
print
def random_sample(self, rng, length):
"""
Randomly sample the HMM to generate a sentence of a given length. This
samples the prior distribution then the observation distribution and
transition distribution for each subsequent observation and state.
This will mostly generate unintelligible garbage, but can provide some
amusement.
@return: the randomly created state/observation sequence,
generated according to the HMM's probability
distributions. The SUBTOKENS have TEXT and TAG
properties containing the observation and state
respectively.
@rtype: Token
@param rng: random number generator
@type rng: Random (or any object with a random() method)
@param length: desired output length
@type length: int
"""
assert chktype(2, length, types.IntType)
# load the property names
SUBTOKENS = self._properties.get('SUBTOKENS', 'SUBTOKENS')
TEXT = self._properties.get('TEXT', 'TEXT')
TAG = self._properties.get('TAG', 'TAG')
# sample the starting state and symbol prob dists
tokens = []
state = self._sample_probdist(self._priors, rng.random(), self._states)
symbol = self._sample_probdist(self._outputs[state], rng.random(),
self._symbols)
tokens.append(Token(TEXT=symbol, TAG=state))
for i in range(1, length):
# sample the state transition and symbol prob dists
state = self._sample_probdist(self._transitions[state],
rng.random(), self._states)
symbol = self._sample_probdist(self._outputs[state], rng.random(),
self._symbols)
tokens.append(Token(TEXT=symbol, TAG=state))
return Token(SUBTOKENS=tokens)
def demo_kmeans():
# example from figure 14.9, page 517, Manning and Schutze
tokens = [Token(FEATURES=Numeric.array(f))
for f in [[2, 1], [1, 3], [4, 7], [6, 7]]]
means = [[4, 3], [5, 5]]
clusterer = KMeansClusterer(2, euclidean_distance, initial_means=means)
clusterer.cluster(tokens, True, trace=True)
print 'clustered', tokens
print 'means', clusterer.means()
def demo_pos_bw():
# demonstrates the Baum-Welch algorithm in POS tagging
from nltk.set import MutableSet
print 'Training HMM (supervised)...'
labelled_sequences, tag_set = load_pos()
symbol_set = MutableSet()
for sequence in labelled_sequences:
for token in sequence['SUBTOKENS']:
symbol_set.insert(token['TEXT'])
trainer = HiddenMarkovModelTrainer(tag_set, symbol_set.elements())
hmm = trainer.train_supervised(
Token(SUBTOKENS=labelled_sequences[100:300]),
estimator=lambda fd, bins: LidstoneProbDist(fd, 0.1, bins))
print 'Training (unsupervised)...'
# it's rather slow - so only use 10 samples
unlabelled = Token(SUBTOKENS=_untag(labelled_sequences[301:311]))
hmm = trainer.train_unsupervised(unlabelled, model=hmm, max_iterations=5)
test_pos(hmm, labelled_sequences[:100], True)
def demo_pos():
# demonstrates POS tagging using supervised training
print 'Training HMM...'
labelled_sequences, tag_set, symbols = load_pos()
trainer = HiddenMarkovModelTrainer(tag_set, symbols)
hmm = trainer.train_supervised(
Token(SUBTOKENS=labelled_sequences[100:]),
estimator=lambda fd, bins: LidstoneProbDist(fd, 0.1, bins))
print 'Testing...'
test_pos(hmm, labelled_sequences[:100], True)
def randomtreetok(depth=0, left=0, bf=None):
if bf == None: bf = randint(1,2)
if randint(0,7-depth) == 0 and depth>1:
len = randint(1,5)
return Token('L%d' % randint(0, 10), left, left+len)
else:
numchildren = randint(1,bf)
children = []
for x in range(numchildren):
children.append(randomtreetok(depth+1, left, bf))
left = children[-1].loc().end()
return TreeToken('Node %d' % randint(0,10000), *children)
def load_pos():
from nltk.corpus import brown
from nltk.tagger import TaggedTokenizer
tagged_tokens = []
for item in brown.items()[:5]:
tagged_tokens.append(brown.tokenize(item))
tag_set = [
"'", "''", '(', ')', '*', ',', '.', ':', '--', '``', 'abl', 'abn',
'abx', 'ap', 'ap$', 'at', 'be', 'bed', 'bedz', 'beg', 'bem', 'ben',
'ber', 'bez', 'cc', 'cd', 'cd$', 'cs', 'do', 'dod', 'doz', 'dt', 'dt$',
'dti', 'dts', 'dtx', 'ex', 'fw', 'hv', 'hvd', 'hvg', 'hvn', 'hvz',
'in', 'jj', 'jjr', 'jjs', 'jjt', 'md', 'nn', 'nn$', 'nns', 'nns$',
'np', 'np$', 'nps', 'nps$', 'nr', 'nr$', 'od', 'pn', 'pn$', 'pp$',
'ppl', 'ppls', 'ppo', 'pps', 'ppss', 'ql', 'qlp', 'rb', 'rb$', 'rbr',
'rbt', 'rp', 'to', 'uh', 'vb', 'vbd', 'vbg', 'vbn', 'vbz', 'wdt',
'wp$', 'wpo', 'wps', 'wql', 'wrb'
]
sequences = []
sequence = []
start_re = re.compile(r'[^-*+]*')
for token in tagged_tokens:
# the multi-output allows us to treat each word as a
# tuple of features
for sub_token in token['SUBTOKENS']:
sequence.append(sub_token)
# a feature for words as lower case
features = [sub_token['TEXT'].lower()]
#a feature for word suffixes of length 3
features.append(sub_token['TEXT'][-3:])
# a feature for the length of words
features.append(len(sub_token['TEXT']))
# store the observation as a tuple of features
sub_token['TEXT'] = tuple(features)
m = start_re.match(sub_token['TAG'])
# cleanup the tag
tag = m.group(0)
if tag in tag_set:
sub_token['TAG'] = tag
else:
sub_token['TAG'] = '*'
# split on the period tag
if sub_token['TAG'] == '.':
sequences.append(Token(SUBTOKENS=sequence))
sequence = []
return sequences, tag_set, 3
def _demo_stemmer(stemmer):
# Tokenize a sample text.
from nltk.tokenizer import WhitespaceTokenizer
text = Token(TEXT='John was eating icecream')
WhitespaceTokenizer().tokenize(text)
# Use the stemmer to stem it.
for word in text['SUBTOKENS']:
stemmer.stem(word)
# Print the results.
print stemmer
for word in text['SUBTOKENS']:
print '%20s => %s' % (word['TEXT'], word['STEM'])
print
def classify(self, unlabeled_token):
# Inherit docs from ClassifierI
assert _chktype(1, unlabeled_token, Token)
text = unlabeled_token.type()
# (label, likelihood) pair that maximizes likelihood
max = (None, 0)
# Find the label that maximizes the non-normalized probability
# fv_list_likelihoods.
for label in self._labels:
fv_list = self._fd_list.detect(LabeledText(text, label))
p = self.fv_list_likelihood(fv_list, label)
if p > max[1]: max = (label, p)
return Token(LabeledText(text, max[0]), unlabeled_token.loc())
def tree2frame(Dirs, index = 0, parent = ''):
"""
@return: content frame representation of the surface semantics of the parse tree.
@rtype: C{SurfaceSemanticsStructure}
@return proposition name
@rtype: C{str}
@return index
@rtype: C{int}
"""
Frame = SurfaceSemanticsStructure()
if isinstance(Dirs,Tree):
Prop = Dirs.node.capitalize()
hasSubTree = True in [isinstance(child,Tree) for child in Dirs]
else: Prop = None
if isinstance(Dirs,Tree) and hasSubTree:
for i,child in enumerate(Dirs):
value,prop,index = tree2frame(child,index+1,Dirs.node.capitalize())
filed = False # Account for children with the same names
if value and prop:
prop_name = prop
while not filed:
if not Frame.has_key(prop):
Frame[prop] = value
filed = True
else:
prop= prop_name+'_'+str(i)
elif value:
Frame1 = Frame.unify(value)
if Frame1: Frame = Frame1
else:
while not filed:
if not Frame.has_key('SubFrame'+'_'+str(index)):
Frame['SubFrame'+'_'+str(index)] = value
filed = True
elif ((isinstance(Dirs,Tree) and not hasSubTree and Dirs)
or isinstance(Dirs,Token)):
index += 1
if isinstance(Dirs,Token): token = Dirs
if isinstance(Dirs,Tree):
token = Token(TEXT=' '.join([child['TEXT'] for child in Dirs]))
parent = Dirs.node.capitalize()
Frame['TEXT'] = token['TEXT']
Frame['MEAN'] = extractSurfaceSemantics(token,parent)
Frame['INDEX']=index
return Frame,Prop,index
def demo_bw():
# demo Baum Welch by generating some sequences and then performing
# unsupervised training on them
# example taken from page 381, Huang et al
symbols = ['up', 'down', 'unchanged']
states = ['bull', 'bear', 'static']
def pd(values, samples):
d = {}
for value, item in zip(values, samples):
d[item] = value
return DictionaryProbDist(d)
def cpd(array, conditions, samples):
d = {}
for values, condition in zip(array, conditions):
d[condition] = pd(values, samples)
return DictionaryConditionalProbDist(d)
A = array([[0.6, 0.2, 0.2], [0.5, 0.3, 0.2], [0.4, 0.1, 0.5]], Float64)
A = cpd(A, states, states)
B = array([[0.7, 0.1, 0.2], [0.1, 0.6, 0.3], [0.3, 0.3, 0.4]], Float64)
B = cpd(B, states, symbols)
pi = array([0.5, 0.2, 0.3], Float64)
pi = pd(pi, states)
model = HiddenMarkovModel(symbols=symbols,
states=states,
transitions=A,
outputs=B,
priors=pi)
# generate some random sequences
training = []
import random
rng = random.Random()
for i in range(10):
item = model.random_sample(rng, 5)
training.append(item)
training = Token(SUBTOKENS=training)
# train on those examples, starting with the model that generated them
trainer = HiddenMarkovModelTrainer(states, symbols)
hmm = trainer.train_unsupervised(training,
model=model,
max_iterations=1000)
def _get_toks(file='ca01', debug=0):
"""
Load tokens from the given file.
"""
assert _chktype(1, file, types.StringType)
assert _chktype(2, debug, types.IntType)
_resettime()
if debug: print _timestamp(), 'tokenizing', file
ttoks = brown.tokenize(file)
labeled_tokens = [Token(LabeledText(tok.type().base().lower(),
tok.type().tag()),
tok.loc())
for tok in ttoks]
if debug: print _timestamp(), ' done tokenizing'
return labeled_tokens
def label_tokens(unlabeled_tokens, label):
"""
@return: a list of labeled tokens, whose text and location
correspond to C{unlabeled_tokens}, and whose labels are
C{label}.
@rtype: C{list} of (C{Token} with type C{LabeledText})
@param unlabeled_tokens: The list of tokens for which a labeled
token list should be created.
@type unlabeled_tokens: C{list} of C{Token}
@param label: The label for the new labeled tokens.
@type label: (immutable)
"""
assert _chktype(1, unlabeled_tokens, [Token], (Token, ))
return [
Token(LabeledText(tok.type(), label), tok.loc())
for tok in unlabeled_tokens
]
def stem(self, token):
# inherit docs from StemmerI
# TODO - when the new token comes out, use it to get the
# part-of-speech, thus narrowing the search (and getting eg.
# fly/verb for the query flies, rather than the plural noun).
# This will only match the first POS from the list below...
for pos in [NOUN, VERB, ADJECTIVE, ADVERB]:
stemmed = morphy(token.type().lower(), pos)
if stemmed:
# restore the case
new_string = ''
for index in range(min(len(token.type()), len(stemmed))):
if token.type()[index].isupper():
new_string += stemmed[index].upper()
else:
new_string += stemmed[index]
return Token(new_string, token.loc())
return token
def load_pos():
from nltk.corpus import brown
tagged_tokens = []
for item in brown.items()[:5]:
tagged_tokens.append(brown.read(item))
tag_set = [
"'", "''", '(', ')', '*', ',', '.', ':', '--', '``', 'abl', 'abn',
'abx', 'ap', 'ap$', 'at', 'be', 'bed', 'bedz', 'beg', 'bem', 'ben',
'ber', 'bez', 'cc', 'cd', 'cd$', 'cs', 'do', 'dod', 'doz', 'dt', 'dt$',
'dti', 'dts', 'dtx', 'ex', 'fw', 'hv', 'hvd', 'hvg', 'hvn', 'hvz',
'in', 'jj', 'jjr', 'jjs', 'jjt', 'md', 'nn', 'nn$', 'nns', 'nns$',
'np', 'np$', 'nps', 'nps$', 'nr', 'nr$', 'od', 'pn', 'pn$', 'pp$',
'ppl', 'ppls', 'ppo', 'pps', 'ppss', 'ql', 'qlp', 'rb', 'rb$', 'rbr',
'rbt', 'rp', 'to', 'uh', 'vb', 'vbd', 'vbg', 'vbn', 'vbz', 'wdt',
'wp$', 'wpo', 'wps', 'wql', 'wrb'
]
sequences = []
sequence = []
symbols = {}
start_re = re.compile(r'[^-*+]*')
for token in tagged_tokens:
for sub_token in token['WORDS']:
sequence.append(sub_token)
# make words lower case
sub_token['TEXT'] = sub_token['TEXT'].lower()
symbols[sub_token['TEXT']] = 1
m = start_re.match(sub_token['TAG'])
# cleanup the tag
tag = m.group(0)
if tag in tag_set:
sub_token['TAG'] = tag
else:
sub_token['TAG'] = '*'
# split on the period tag
if sub_token['TAG'] == '.':
sequences.append(Token(SUBTOKENS=sequence))
sequence = []
return sequences, tag_set, symbols.keys()
def demo(best_path, cache_factory):
# demonstrates POS tagging using supervised training
print 'Training HMM...'
labelled_sequences, tag_set, symbols = load_pos()
trainer = HiddenMarkovModelTrainer(tag_set, symbols)
hmm = trainer.train_supervised(
Token(SUBTOKENS=labelled_sequences[100:]),
estimator=lambda fd, bins: LidstoneProbDist(fd, 0.1, bins))
print 'Creating cache', cache_factory
cache = cache_factory(hmm)
print 'Overriding best_path with', best_path
hmm.__class__.best_path = lambda self, seq: best_path(self, seq, cache)
print 'Testing...'
import time
start = time.clock()
test_pos(hmm, labelled_sequences[:100], True)
print 'elapsed time', (time.clock() - start)
def accuracy(classifier, labeled_tokens):
"""
@rtype: C{float}
@return: the given classifier model's accuracy on the given list
of labeled tokens. This float between zero and one indicates
what proportion of the tokens the model would label correctly.
@param labeled_tokens: The tokens for which the model's
accuracy should be computed.
@type labeled_tokens: C{list} of (C{Token} with type
C{LabeledText})
"""
assert _chktype(1, classifier, ClassifierI)
assert _chktype(2, labeled_tokens, [Token], (Token, ))
total = 0
correct = 0
for ltok in labeled_tokens:
utok = Token(ltok.type().text(), ltok.loc())
if classifier.classify(utok) == ltok:
correct += 1
total += 1
return float(correct) / total
def test(numFiles=100,
max_rules=200,
min_score=2,
ruleFile="dump.rules",
errorOutput="errors.out",
ruleOutput="rules.out",
randomize=False,
train=.8,
trace=3):
NN_CD_tagger = RegexpTagger([(r'^[0-9]+(.[0-9]+)?$', 'CD'), (r'.*', 'NN')],
TAG='POS')
# train is the proportion of data used in training; the rest is reserved
# for testing.
print "Loading tagged data..."
taggedData = getWSJTokens(numFiles, randomize)
trainCutoff = int(len(taggedData) * train)
trainingData = Token(SUBTOKENS=taggedData[0:trainCutoff])
goldData = Token(SUBTOKENS=taggedData[trainCutoff:])
testingData = goldData.exclude('POS')
# Unigram tagger
print "Training unigram tagger:",
u = UnigramTagger(TAG='POS')
u.train(trainingData)
backoff = BackoffTagger([u, NN_CD_tagger], TAG='POS')
print("[accuracy: %f]" % tagger_accuracy(backoff, [goldData]))
# Brill tagger
templates = [
SymmetricProximateTokensTemplate(ProximateTagsRule, (1, 1)),
SymmetricProximateTokensTemplate(ProximateTagsRule, (2, 2)),
SymmetricProximateTokensTemplate(ProximateTagsRule, (1, 2)),
SymmetricProximateTokensTemplate(ProximateTagsRule, (1, 3)),
# SymmetricProximateTokensTemplate(ProximateWordsRule, (1,1)),
# SymmetricProximateTokensTemplate(ProximateWordsRule, (2,2)),
# SymmetricProximateTokensTemplate(ProximateWordsRule, (1,2)),
# SymmetricProximateTokensTemplate(ProximateWordsRule, (1,3)),
ProximateTokensTemplate(ProximateTagsRule, (-1, -1), (1, 1)),
# ProximateTokensTemplate(ProximateWordsRule, (-1, -1), (1,1)),
]
#trainer = FastBrillTaggerTrainer(backoff, templates, trace, TAG='POS')
trainer = BrillTaggerTrainer(backoff, templates, trace, TAG='POS')
b = trainer.train(trainingData, max_rules, min_score)
print
print("Brill accuracy: %f" % tagger_accuracy(b, [goldData]))
print("\nRules: ")
printRules = file(ruleOutput, 'w')
for rule in b.rules():
print(str(rule))
printRules.write(str(rule) + "\n\n")
#b.saveRules(ruleFile)
b.tag(testingData)
el = errorList(goldData, testingData)
errorFile = file(errorOutput, 'w')
for e in el:
errorFile.write(e + "\n\n")
errorFile.close()
print("Done.")
return b
def raw_stem(self, text):
TEXT = self.property('TEXT')
STEM = self.property('STEM')
token = Token({TEXT: text})
self.stem(token)
return token[STEM]
def cross_validate(trainer,
labeled_tokens,
n_folds=10,
target=None,
trace=False):
"""
Perform N-fold cross validation on the given classifier. This divides the
tokens into N equally sized groups (subject to rounding), then performs N
training and testing passes. Each pass involves testing on a single fold
and testing on the remaining folds. This way every instance is used
exactly once for testing. The results (predictive accuracy) are averaged
over the N trials. The mean and standard deviation are returned as a
tuple.
"""
assert len(labeled_tokens) >= n_folds
# should randomly reorder labeled_tokens first?
folds = []
n = len(labeled_tokens)
for i in range(n_folds):
start = i * n / n_folds
end = (i + 1) * n / n_folds
folds.append(labeled_tokens[start:end])
if trace:
print 'cross_validate - using %d folds of %d items each approx' \
% (n_folds, len(folds[0]))
accuracies = []
precisions = []
recalls = []
for i in range(n_folds):
training = folds[:]
testing = training[i]
del training[i]
training = reduce(operator.add, training) # flatten
if trace:
print 'cross_validate [%d] - training classifier...' % (i + 1)
import time
start = time.time()
classifier = trainer.train(training)
if trace:
end = time.time()
print 'cross_validate elapsed time %.2f seconds' % (end - start)
print 'cross_validate [%d] - testing classifier...' % (i + 1)
start = end
yes = no = 0
tp = tn = fp = fn = 0
for ltok in testing:
utok = Token(ltok.type().text(), ltok.loc())
if trace >= 2:
print 'cross_validate [%d] - given' % (i + 1), ltok
ctok = classifier.classify(utok)
if trace >= 2:
print 'cross_validate [%d] - classified' % (i + 1),
print ctok.type().label()
if ltok.type().label() == ctok.type().label():
yes += 1
else:
no += 1
if target:
if ltok.type().label() == target:
if ltok.type().label() == ctok.type().label():
tp += 1
else:
fn += 1
else:
if ltok.type().label() == ctok.type().label():
fp += 1
else:
tn += 1
acc = float(yes) / (yes + no)
accuracies.append(acc)
if target:
precision = recall = None
try:
recall = float(tp) / (tp + fn)
recalls.append(recall)
except ZeroDivisionError:
pass
try:
precision = float(tp) / (tp + fp)
precisions.append(precision)
except ZeroDivisionError:
pass
if trace:
end = time.time()
print 'cross_validate elapsed time %.2f seconds' % (end - start)
print 'cross_validate [%d] - accuracy %.3f' % (i + 1, acc)
if target:
print 'cross_validate [%d] - precision %s recall %s' \
% (i + 1, precision, recall)
if trace:
print 'cross_validate - calculating mean and variance'
# find the mean
mean = reduce(operator.add, accuracies) / float(len(accuracies))
if target:
recall = reduce(operator.add, recalls) / float(len(recalls))
if len(precisions) > 0:
precision = reduce(operator.add, precisions) / float(
len(precisions))
else:
precision = None
# find the standard deviation
var = 0
for i in range(n_folds):
var += accuracies[i] * (accuracies[i] - mean)**2
sd = var**0.5
if target:
return mean, sd, precision, recall
else:
return mean, sd
def test(
numFiles=100,
max_rules=200,
min_score=2,
ruleFile="dump.rules",
errorOutput="errors.out",
ruleOutput="rules.out",
randomize=False,
train=0.8,
trace=3,
):
NN_CD_tagger = RegexpTagger([(r"^[0-9]+(.[0-9]+)?$", "CD"), (r".*", "NN")], TAG="POS")
# train is the proportion of data used in training; the rest is reserved
# for testing.
print "Loading tagged data..."
taggedData = getWSJTokens(numFiles, randomize)
trainCutoff = int(len(taggedData) * train)
trainingData = Token(SUBTOKENS=taggedData[0:trainCutoff])
goldData = Token(SUBTOKENS=taggedData[trainCutoff:])
testingData = goldData.exclude("POS")
# Unigram tagger
print "Training unigram tagger:",
u = UnigramTagger(TAG="POS")
u.train(trainingData)
backoff = BackoffTagger([u, NN_CD_tagger], TAG="POS")
print ("[accuracy: %f]" % tagger_accuracy(backoff, [goldData]))
# Brill tagger
templates = [
SymmetricProximateTokensTemplate(ProximateTagsRule, (1, 1)),
SymmetricProximateTokensTemplate(ProximateTagsRule, (2, 2)),
SymmetricProximateTokensTemplate(ProximateTagsRule, (1, 2)),
SymmetricProximateTokensTemplate(ProximateTagsRule, (1, 3)),
# SymmetricProximateTokensTemplate(ProximateWordsRule, (1,1)),
# SymmetricProximateTokensTemplate(ProximateWordsRule, (2,2)),
# SymmetricProximateTokensTemplate(ProximateWordsRule, (1,2)),
# SymmetricProximateTokensTemplate(ProximateWordsRule, (1,3)),
ProximateTokensTemplate(ProximateTagsRule, (-1, -1), (1, 1)),
# ProximateTokensTemplate(ProximateWordsRule, (-1, -1), (1,1)),
]
# trainer = FastBrillTaggerTrainer(backoff, templates, trace, TAG='POS')
trainer = BrillTaggerTrainer(backoff, templates, trace, TAG="POS")
b = trainer.train(trainingData, max_rules, min_score)
print
print ("Brill accuracy: %f" % tagger_accuracy(b, [goldData]))
print ("\nRules: ")
printRules = file(ruleOutput, "w")
for rule in b.rules():
print (str(rule))
printRules.write(str(rule) + "\n\n")
# b.saveRules(ruleFile)
b.tag(testingData)
el = errorList(goldData, testingData)
errorFile = file(errorOutput, "w")
for e in el:
errorFile.write(e + "\n\n")
errorFile.close()
print ("Done.")
return b
##//////////////////////////////////////////////////////
## Demo Code
##//////////////////////////////////////////////////////
import random
if __name__ == '__main__':
def fill(cw):
cw['fill'] = '#%06d' % random.randint(0,999999)
cf = CanvasFrame(width=550, height=450, closeenough=2)
tree = Tree.parse('''
(S (NP the very big cat)
(VP (Adv sorta) (V saw) (NP (Det the) (N dog))))
''', leafparser = lambda t: Token(TEXT=t))
tc = TreeWidget(cf.canvas(), tree, draggable=1,
node_font=('helvetica', -14, 'bold'),
leaf_font=('helvetica', -12, 'italic'),
roof_fill='white', roof_color='black',
leaf_color='green4', node_color='blue2')
cf.add_widget(tc,10,10)
def boxit(canvas, text):
big = ('helvetica', -16, 'bold')
return BoxWidget(canvas, TextWidget(canvas, text,
font=big), fill='green')
def ovalit(canvas, text):
return OvalWidget(canvas, TextWidget(canvas, text),
fill='cyan')
