def _counter_known(parsed, train, known, discount, tag2id, word2id, prior):
emission, transition = None, None
if known:
emission, _ = counter(parsed, tag2id, word2id, discount, prior)
_, transition = counter(train, tag2id, word2id, discount, prior)
else:
emission, transition = counter(train, tag2id, word2id, discount, prior)
return emission, transition
for j in xrange(1, len(seq)):
for i in xrange(len(transition)):
k_score = scores[:, j - 1] + np.log(transition[:, i]) + np.log(emission[i, seq[j]])
backpointer[i, j] = np.argmax(k_score)
scores[i, j] = k_score[backpointer[i, j]]
j = int(np.argmax(scores, axis=0)[-1])
sol = [j]
for i in xrange(len(seq) - 1, 0, -1):
j = backpointer[j, i]
sol.append(j)
sol.reverse()
return zip(seq[:-1], sol[:-1])
if __name__ == "__main__":
path = "../WSJ-2-12/*/*.POS"
docs = glob(path)
parsed = parse(docs)
np.random.shuffle(parsed)
parsed = trigramize(parsed)
tag2id, word2id = build_dict(parsed[:-10])
id2word = {v: k for k, v in word2id.iteritems()}
id2tag = {v: k for k, v in tag2id.iteritems()}
emission, transition = counter(parsed[:-10], tag2id, word2id)
print "test POS", parsed[-1][1:-1]
output = viterbi(parsed[-1], transition, emission, word2id, tag2id)
print "TAGGED", id_to_token(output, id2word, id2tag)
