def extract_pairs(sess, source_sentences, target_sentences,
source_sentences_ids, target_sentences_ids, probs_op,
placeholders):
"""Extract sentence pairs from a pair of articles in source and target languages.
Returns a list of (source sentence, target sentence, probability score) tuples.
"""
pairs = [(i, j) for i, j in product(range(len(source_sentences)),
range(len(target_sentences)))]
data = [(source_sentences_ids[i], target_sentences_ids[j], 1.0)
for i, j in product(range(len(source_sentences)),
range(len(target_sentences)))]
data_iterator = utils.TestingIterator(np.array(data, dtype=object))
y_score = inference(sess, data_iterator, probs_op, placeholders)
y_score = [(score, k) for k, score in enumerate(y_score)]
y_score.sort(reverse=True)
i_aligned = set()
j_aligned = set()
sentence_pairs = []
for score, k in y_score:
i, j = pairs[k]
if score < FLAGS.decision_threshold or i in i_aligned or j in j_aligned:
continue
if FLAGS.use_greedy:
i_aligned.add(i)
j_aligned.add(j)
sentence_pairs.append(
(source_sentences[i], target_sentences[j], score))
return sentence_pairs
def extract_pairs(sess, source_sentences, target_sentences,
source_sentences_ids, target_sentences_ids, probs_op,
placeholders):
"""Extract sentence pairs from a pair of articles in source and target languages.
Only keep sentence pairs that occur in a block of at least MINBLOCKSIZE sentence pairs with high scores.
Returns a list of (source sentence, target sentence, probability score) tuples.
"""
pairs = [(i, j) for i, j in product(range(len(source_sentences)),
range(len(target_sentences)))]
data = [(source_sentences_ids[i], target_sentences_ids[j], 1.0)
for i, j in product(range(len(source_sentences)),
range(len(target_sentences)))]
data_iterator = utils.TestingIterator(np.array(data, dtype=object))
y_score = inference(sess, data_iterator, probs_op, placeholders)
y_score = [(score, k) for k, score in enumerate(y_score)]
sentence_pairs = []
Pairs = set()
minsizeblock = 2
in_block = []
for score, k in y_score:
if score < FLAGS.decision_threshold:
continue
i, j = pairs[k]
print("---", i, j)
# scoresum = 0
if (i, j) in in_block:
print("in block")
continue
block = []
for score2, k2 in y_score:
if score2 < FLAGS.decision_threshold:
continue
i2, j2 = pairs[k2]
if i2 == i and j2 == j:
print(i2, j2)
# scoresum += score2
block.append((i2, j2, score2))
i += 1
j += 1
if (len(block) > 1):
in_block.append((i2, j2))
if len(block) >= minsizeblock:
print("--------------------------")
for i, j, score in block:
print(i, j)
print(source_sentences[i])
print(target_sentences[j])
print(score)
print("\n")
Pairs.add((source_sentences[i], target_sentences[j], score))
for srcsent, trgsent, score in Pairs:
sentence_pairs.append((srcsent, trgsent, score))
return sentence_pairs
def evaluate(sess, source_sentences, target_sentences, references,
source_sentences_ids, target_sentences_ids, probs_op,
placeholders):
""""Evalute BiRNN at decision threshold value maximizing the area
under the precison-recall curve.
"""
pairs = [(i, j) for i, j in product(range(len(source_sentences)),
range(len(target_sentences)))]
data = [(source_sentences_ids[i], target_sentences_ids[j], 1.0) if
(i, j) in references else
(source_sentences_ids[i], target_sentences_ids[j], 0.0)
for i, j in product(range(len(source_sentences)),
range(len(target_sentences)))]
data_iterator = utils.TestingIterator(np.array(data, dtype=object))
y_score = inference(sess, data_iterator, probs_op, placeholders)
y_true = data_iterator.data[:, 2].astype(int)
p, r, t = precision_recall_curve(y_true, y_score, pos_label=1)
f1 = utils.f1_score(p, r)
i = np.argmax(f1)
print("Evaluation metrics at decision threshold = {:.4f}\n"
"Precision = {:.2f}, Recall = {:.2f}, F1 = {:.2f}\n"
"-------------------------------------------------".format(
p[i], 100 * r[i], 100 * f1[i], 100 * t[i]))