def train_sent(self, sent):
# take sent from corpus and update weight vector
# according to this example
self.model.update_perceptron_counter()
# init state
sent = [ROOT] + sent
init_state = get_state(sent)
# build gold tree
gold_tree = get_tree(sent)
beam = Beam(self.model.beam_size)
top_valid_state = None
beam.add(init_state)
# loop n-1 step
step = 0
# above check
for i in xrange(len(sent) - 1):
step += 1
# extend current beam
beam, top_valid_state = self.extend_beam_for_train(beam, gold_tree)
try:
if not beam.has_element(top_valid_state.top()):
self.update_paramaters(beam.top(), -1)
self.update_paramaters(top_valid_state.top(), 1)
break
except Exception:
raise
if step == len(sent) - 1:
top_beam = beam.top()
predict_arcs = top_beam['arcs']
if not compare_arcs(predict_arcs, gold_tree):
self.update_paramaters(top_beam, -1)
self.update_paramaters(top_valid_state.top(), 1)
def train(self, sent):
# update paramaters with one sent
# ROOT token at begining of pending
sent = [ROOT] + sent
# oracle object to check valid action
oracle = Oracle(sent)
# gold_deps for full update
gold_deps = self._build_gold(sent)
# create start state
init_state = self._get_state(sent)
# create beam
beam = Beam(self.model.beam_size)
# add state to beam
beam.add(init_state)
# correct action with highest score at one step
valid_action = None
for step in range(len(sent) - 1):
beam, valid_action = self._extend_beam(beam, oracle)
# if beam not contain valid action in it, update
if not beam.has_element(valid_action):
beam_top = beam.top()
self.model.update(beam_top, valid_action)
break
else:
beam_top = beam.top()
beam_deps = beam_top['deps']
# if final deps is not like gold_deps, do full update
if not self._check_equal(gold_deps, beam_deps):
self.model.update(beam_top, valid_action)
def _extend_beam(self, beam, oracle):
new_beam = Beam(self.model.beam_size)
valid_action = Beam(beam_size=1)
for state in beam:
pending, prev_score, prev_feats, deps, stt = self._extract_state(
state)
for i, (tok1, tok2) in enumerate(zip(pending, pending[1:])):
lc_feats = self.model.featex(pending, deps, i)
scores = self.model.get_score(lc_feats)
go_feats = prev_feats + lc_feats
for clas, score in scores.iteritems():
arc = self._get_action(clas, tok1, tok2)
# stt ensure all action before in state is valid
if stt:
is_valid = self._check_valid(arc, deps, oracle)
n_pending, n_deps = self._apply_action(arc, state)
if prev_score == float('-inf'):
n_score = score
else:
n_score = prev_score + score
new_state = self._get_state(n_pending, go_feats, n_score,
clas, n_deps, is_valid)
new_beam.add(new_state)
if is_valid:
valid_action.add(new_state)
return new_beam, valid_action.top()
def parse(self, sent):
sent = [ROOT] + sent
init_state = get_state(sent)
beam = Beam(self.model.beam_size)
beam.add(init_state)
for i in xrange(len(sent) - 1):
beam = self.extend_beam_for_parse(beam)
final_state = beam.top()
return final_state['arcs']
def parse(self, sent):
# parse one sent according to current model paramaters
# ROOT token at begining of pending
sent = [ROOT] + sent
# start state
init_state = self._get_state(sent)
# create a beam
beam = Beam(self.model.beam_size)
# add state to beam
beam.add(init_state)
# loop until only one tree left
for step in range(len(sent) - 1):
# beam of next step
beam = self._extend_beam_for_test(beam)
# result of parse
deps = beam.top()['deps']
return deps
