def oracle_moves(gold_heads):
dependents = defaultdict(int)
for head in gold_heads:
dependents[head] += 1
m = None
config = Parser.initial_config(len(gold_heads))
i, stack, heads = config
while not Parser.is_final_config(config):
valid_moves = Parser.valid_moves(config)
second_topmost = stack[-2] if len(stack) >= 2 else None
topmost = stack[-1] if len(stack) >= 1 else None
if Parser.LA in valid_moves and gold_heads[
second_topmost] == topmost and dependents[second_topmost] == 0:
m = Parser.LA
dependents[topmost] -= 1
elif Parser.RA in valid_moves and gold_heads[
topmost] == second_topmost and dependents[topmost] == 0:
m = Parser.RA
dependents[second_topmost] -= 1
elif Parser.SH in valid_moves:
m = Parser.SH
else:
raise "Oracle failed, no moves valid"
yield config, m
config = Parser.next_config(config, m)
def predict(self, words, tags):
pred_heads = []
config = Parser.initial_config(len(words))
while not Parser.is_final_config(config):
valid_moves = Parser.valid_moves(config)
features = self.featurize(words, tags, config)
pred_moves = self.model.forward(features)
best_m_s = [valid_moves[0], pred_moves[valid_moves[0]]]
for m in valid_moves:
if pred_moves[m] > best_m_s[1]:
best_m_s = [m, pred_moves[m]]
config = Parser.next_config(config, best_m_s[0])
return config[2]
def predict(self, words, tags):
parser = Parser()
# 1. Start in the initial configuration for the input sentence.
config = parser.initial_config(len(words))
# 2. As long as there are valid moves, ask the averaged perceptron for the next move to take.
while len(self.valid_moves(config)) != 0:
features = self.featurize(words, tags, config)
output_vector = self.model.forward(features)
move = max(output_vector, key=output_vector.get)
config = self.next_config(config, move)
# 3. Return the list of heads associated with the final configuration.
return config[2]