def __init__(self, length, classes, weights, gazetteer):
self.length = length
self.classes = classes
self.trellis = [[(0.0, 0) for i in range(length)] for j in range(len(classes))]
self.weights = {}
self.read_weights(weights, self.weights)
self.feature_gen = FeatGenerator(gazetteer)
class Viterbi:
'Viterbi decoder to estimate the most likely sequence'
def __init__(self, length, classes, weights, gazetteer):
self.length = length
self.classes = classes
self.trellis = [[(0.0, 0) for i in range(length)] for j in range(len(classes))]
self.weights = {}
self.read_weights(weights, self.weights)
self.feature_gen = FeatGenerator(gazetteer)
def read_weights(self, weights_fname, weights_dict):
file = open(weights_fname, 'r')
for line in file.readlines():
tokens = line.split()
weights_dict[tokens[0]] = float(tokens[1])
# this method computes the local features given a word and its context
def get_local_score(self, prev_word, prev_pos_tag, curr_word, curr_pos_tag, \
next_word, next_pos_tag, prev_ner_tag, curr_ner_tag, token_nbr):
feats = {}
self.feature_gen.get_feature_vector(prev_word, prev_pos_tag, curr_word, curr_pos_tag, \
next_word, next_pos_tag, prev_ner_tag, curr_ner_tag, token_nbr, feats)
score = 0.0
for key, value in feats.iteritems():
increment = (value * self.weights.get(key, 0.0))
score = score + increment
#if curr_ner_tag is 'O':
#print key, value
return score
# computes the maximums and updates column of the trellis
def update_trellis(self, prev, curr, next, token_nbr):
class_indices = range(len(self.classes))
# extract the word and pos tag from prev, curr and next
prev_word = prev[0]
prev_pos_tag = prev[1]
curr_word = curr[0]
curr_pos_tag = curr[1]
next_word = next[0]
next_pos_tag = next[1]
for i in class_indices:
max_score = 0.0
backpointer = 0
curr_ner_tag = self.classes[i]
# check for STOP and START
for j in class_indices:
prev_ner_tag = self.classes[j]
if token_nbr is 0:
prev_score = 0.0
else:
prev_score = self.trellis[j][token_nbr-1][0]
score = prev_score + self.get_local_score(prev_word, prev_pos_tag, curr_word, curr_pos_tag, \
next_word, next_pos_tag, prev_ner_tag, curr_ner_tag, token_nbr)
if score > max_score:
max_score = score
backpointer = j
self.trellis[i][token_nbr] = (max_score, backpointer)
def get_output_sequence(self, result):
sequence = []
num_classes = len(self.classes)
bp = len(self.classes)-1
for i in reversed(range(self.length)):
sequence.append(bp)
bp = self.trellis[bp][i][1]
for class_index in reversed(sequence):
result.append(self.classes[class_index])
