def oracle_trace(self, document):
assert len(document.gold) > 0, "No gold actions"
state = ParserState(document, self)
for gold in document.gold:
print("Taking gold action", gold)
print("On state:", state)
gold_index = self.actions.indices.get(gold, None)
assert gold_index is not None, "Unknown gold action: %r" % gold
assert state.is_allowed(gold_index), "Disallowed gold action: %r" % gold
state.advance(gold)
print("Final state after", len(document.gold), "actions:", state)
def __main__():
trainingCorpus = ParsedConllFile(keepMalformed=False, projectivize=True)
trainingCorpus.read( \
open(trainingFile, 'r',
encoding='utf-8').read())
# make fake model params, enough for lexicon builder
# we still need feature_maps to use ParserState
modelParams = ModelParameters('')
modelParams.trainingFile = trainingFile
modelParams.cfg = {'projectivizeTrainingSet': True}
lexicon = Lexicon(modelParams)
lexicon.compute()
sentence = trainingCorpus.sentences[0]
parser_state = ParserState(sentence, lexicon.getFeatureMaps())
# necessary for initializing and pushing root
# (only initialize transition_state_class once!)
# keep arc_state in sync with parser_state
arc_state = transition_state_class(parser_state)
dynamicOracleTrainTest(parser_state)
def test_parse():
"""Simple tests for the PartialParse.parse function.
Warning: these are not exhaustive.
"""
sentence = [
Token(i + 1, f) for i, f in enumerate(["parse", "this", "sentence"])
]
state = ParserState(stack=[ROOT], buffer=sentence)
dependencies = state.parse(["S", "S", "S", "LA", "RA", "RA"])
dependencies = [(a[0].form, a[1].form) for a in sorted(dependencies)]
expected = [('ROOT', 'parse'), ('parse', 'sentence'), ('sentence', 'this')]
assert dependencies == expected, \
f"parse test resulted in dependencies {dependencies}, expected {expected}"
assert [t.form for t in sentence] == ["parse", "this", "sentence"], \
f"parse test failed: the input sentence should not be modified"
print("parse test passed!")
def parse(self, sentences, model, conllu=False):
"""
@param sentences: a list of (list Token).
@param model: a trained parser model.
@param conllu: if True prints the parsed sentences in CoNLL-U format.
"""
vsentences = self.vectorize(sentences)
UAS = LAS = all_tokens = 0.0
for sent, vsent in zip(sentences, vsentences):
if not conllu: print('.', end='') # show progress
state = ParserState([self.root_token], vsent, []) # FIXME
while state.buffer or len(state.stack) > 1:
feats = state.extract_features(self)
trans = model.predict([feats])[0].argmax()
if not state.step(trans):
break # if transition is not feasible
if conllu:
for j, t in enumerate(sent):
head = deprel = 0
for arc in state.arcs:
if arc[1].id == t.id:
head = arc[0].id
deprel = arc[2]
break
print('\t'.join([
str(j + 1), t.form, '_', t.pos, '_', '_',
str(head), self.id2dep[deprel], '_', '_'
]))
print()
for arc in state.arcs:
pred_h = arc[0].id
gold_h = arc[1].head
UAS += pred_h == gold_h
pred_l = arc[2]
gold_l = arc[1].deprel
LAS += pred_h == gold_h and pred_l == gold_l
all_tokens += 1
UAS /= all_tokens
LAS /= all_tokens
return UAS, LAS
def advanceSentence(self, i):
self.logger.debug('Slot(%d): advance sentence' % i)
assert i >= 0 and i < self.batch_size
if (self.sentence_batch.advanceSentence(i)):
self.parser_states[i] = ParserState(
self.sentence_batch.sentence(i), self.feature_maps)
# necessary for initializing and pushing root
# keep arc_states in sync with parser_states
self.arc_states[i] = \
self.transition_state_class(self.parser_states[i])
else:
self.parser_states[i] = None
self.arc_states[i] = None
def advanceSentence(self, i):
assert i >= 0 and i < self.batch_size
if (self.sentence_batch.advanceSentence(i)):
self.parser_states[i] = ParserState(
self.sentence_batch.sentence(i), self.feature_maps)
# necessary for initializing and pushing root
# keep arc_states in sync with parser_states
self.arc_states[i] = \
self.transition_state_class(self.parser_states[i])
else:
self.parser_states[i] = None
self.arc_states[i] = None
if self.state(i) != None:
self.docids_.insert(0, self.state(i).sentence().docid())
def create_features(self, sentences):
"""
Build training instances.
@return: list(features), list(action) for each state while parsing each sentence
"""
train_x, train_y = [], []
with tqdm(total=len(sentences)) as prog:
for sent in sentences:
# arcs = [(head, dependent, deprel)]
state = ParserState([self.root_token], sent, []) # FIXME
while state.buffer or len(state.stack) > 1:
gold_t = state.get_oracle()
if gold_t is None:
break
train_x.append(state.extract_features(self))
train_y.append(gold_t)
state.step(gold_t) # perform transition
prog.update(1)
return train_x, train_y
def forward(self, document, train=False, debug=False):
# Compute LSTM outputs for all tokens.
lr_out, rl_out, lstm_features = self._lstm_outputs(document)
# Run FF unit.
state = ParserState(document, self.spec)
actions = self.spec.actions
cascade = self.spec.cascade
ff_activations = []
if train:
losses = Losses()
# Translate the gold actions into their cascade equivalents.
cascade_gold = cascade.translate(document.gold)
gold_index = 0
while not state.done:
# Compute the hidden layer once for all cascade delegates.
ff_activation, _ = self._ff_activation(
lr_out, rl_out, ff_activations, state)
cascading = True
delegate_index = 0 # assume we start the cascade at delegate 0
while cascading:
# Get the gold action for the delegate and compute loss w.r.t. it.
gold = cascade_gold[gold_index]
step_loss = cascade.loss(delegate_index, state, ff_activation, gold)
losses.add(delegate_index, step_loss)
# If the gold action was a CASCADE, move to the next delegate.
if gold.is_cascade():
delegate_index = gold.delegate
else:
state.advance(gold)
cascading = False
gold_index += 1
return losses
else:
if document.size() == 0: return state
shift = actions.action(actions.shift())
stop = actions.action(actions.stop())
disallowed_counts = [0] * cascade.size()
total_counts = [0] * cascade.size()
trace = Trace(self.spec, state, lstm_features) if debug else None
while not state.done:
# Compute the FF activation once for all cascade delegates.
ff_activation, ff_features = self._ff_activation(
lr_out, rl_out, ff_activations, state, debug=debug)
if trace:
trace.start_step(state, ff_features)
# Store the last CASCADE action in a cascade.
delegate_index = 0
last = None
while True:
# Get the highest scoring action from the cascade delegate.
# Note: We don't have to do any filtering or checking here, we
# can just return the top-scoring action.
best = cascade.predict(delegate_index, state, last, ff_activation)
final = best
if best.is_cascade():
delegate_index = best.delegate
last = best
else:
# If the action isn't allowed or can't be applied to the state,
# then default to SHIFT or STOP.
index = actions.index(best)
total_counts[delegate_index] += 1
if actions.disallowed[index] or not state.is_allowed(index):
disallowed_counts[delegate_index] += 1
final = shift
if state.current == state.end:
final = stop
if trace:
trace.action(best, final)
if not final.is_cascade():
# Apply the action and stop the cascade.
state.advance(final)
break
return state, disallowed_counts, total_counts, trace
def setupParser(self, mode):
hiddenLayerSizes = self.modelParams.cfg['hiddenLayerSizes']
featureStrings = self.modelParams.cfg['featureStrings']
embeddingSizes = self.modelParams.cfg['embeddingSizes']
batchSize = self.modelParams.cfg['batchSize']
transitionSystem = self.modelParams.cfg['transitionSystem']
if transitionSystem == 'arc-standard':
self.transitionSystem = ArcStandardTransitionSystem()
elif transitionSystem == 'arc-eager':
self.transitionSystem = ArcEagerTransitionSystem()
else:
assert None, 'transition system must be arc-standard or arc-eager'
assert len(hiddenLayerSizes) > 0, 'must have at least one hidden layer'
assert len(featureStrings) == len(set(featureStrings)), \
'duplicate feature string detected'
if mode == 'train':
# determine if we have to compute or read the lexicon
self.logger.info('Computing lexicon from training corpus...')
self.modelParams.lexicon.compute()
self.logger.info('Done building lexicon')
self.modelParams.lexicon.write()
elif mode == 'evaluate':
self.logger.info('Reading lexicon from trained model...')
self.modelParams.lexicon.read()
else:
assert None, 'invalid mode: ' + mode
self.featureMaps = self.modelParams.lexicon.getFeatureMaps()
self.logger.info('Feature strings: ' + str(featureStrings))
# Get major type groups in sorted order by contructing null parser
# state and extracting features, and then concatenating the similar
# types
fvec = SparseFeatureExtractor(featureStrings, self.featureMaps) \
.extract(ParserState(ParsedConllSentence(docid=None),
self.featureMaps), doLogging=False)
featureTypeInstances = fvec.types
self.featureMajorTypeGroups, _ = fvec.concatenateSimilarTypes()
# index: major feature type index
# values: feature names under that type
self.featureNames = [[] for t in self.featureMajorTypeGroups]
self.logger.info('Detected major feature groups (in alphabetical '
'order): ' + str(self.featureMajorTypeGroups))
self.featureDomainSizes = []
#self.featureEmbeddings = []
# For now, use all same embedding sizes
self.featureEmbeddingSizes = \
[embeddingSizes[t] for t in self.featureMajorTypeGroups]
self.BAG_OF_FEATURES_LEN = 0
for i in range(len(featureTypeInstances)):
major_type = featureTypeInstances[i].major_type
major_type_index = self.featureMajorTypeGroups.index(major_type)
self.featureNames[major_type_index].append(
featureTypeInstances[i].name)
self.BAG_OF_FEATURES_LEN += \
(self.featureEmbeddingSizes[major_type_index])
for i in range(len(self.featureMajorTypeGroups)):
major_type = self.featureMajorTypeGroups[i]
self.logger.info('')
self.logger.info('Feature group \'%s\'' % major_type)
self.logger.info('... domain size: %d' % \
(self.featureMaps[major_type].getDomainSize( \
includeSpecial=True)))
self.logger.info('... embedding size: %d' % \
(self.featureEmbeddingSizes[i]))
#self.logger.info('... feature count: %d' % \
# (len(self.featureNames[i])))
self.logger.info('... features')
for fname in self.featureNames[i]:
self.logger.info('....... %s' % (fname))
self.logger.info('... total group embedding size: %d' % \
(len(self.featureNames[i]) * self.featureEmbeddingSizes[i]))
self.logger.info('... initializing random normal embeddings...')
self.featureDomainSizes.append(
self.featureMaps[major_type].getDomainSize( \
includeSpecial=True))
assert len(self.featureDomainSizes) == len(self.featureEmbeddingSizes)
#assert len(self.featureDomainSizes) == len(self.featureEmbeddings)
assert len(self.featureDomainSizes) == len(self.featureNames)
self.logger.info('')
self.logger.info('Batch size (number of parser states): %d' %
batchSize)
self.logger.info('Total feature count: %d' % \
(len(featureTypeInstances)))
self.logger.info('Total bag of features length per state: %d' % \
(self.BAG_OF_FEATURES_LEN))
self.logger.info('Total features input size: %d' % \
(batchSize*self.BAG_OF_FEATURES_LEN))
# for actions, we don't encode UNKNOWN, ROOT, or OUTSIDE
# we only encode the number of base values
self.ACTION_COUNT = self.transitionSystem.numActions(
self.featureMaps['label'].getDomainSize(includeSpecial=False))
self.logger.info('Total action count: %d' % self.ACTION_COUNT)