def get_aer(self, dataset, epoch):
"""Compute the Alignment Error Rate of the model using the best alignments
Parameters:
dataset: The training dataset
epoch: The current epoch
Returns:
AER score
"""
print("Computing AER on validation dataset")
gold_sets = aer.read_naacl_alignments(
"datasets/validation/dev.wa.nonullalign")
if self.opt.mode == 'test':
gold_sets = aer.read_naacl_alignments(
"datasets/testing/answers/test.wa.nonullalign")
metric = aer.AERSufficientStatistics()
predictions = self.get_best_alignments(dataset.val_data, epoch)
for gold, pred in zip(gold_sets, predictions):
prediction = set([(alignment[1], alignment[2])
for alignment in pred])
metric.update(sure=gold[0], probable=gold[1], predicted=prediction)
return metric.aer()
def compute_aer(predictions, file_path):
"""
Computes the Alignment Error Rate.
"""
gold_sets = aer.read_naacl_alignments(file_path)
metric = aer.AERSufficientStatistics()
for gold, prediction in zip(gold_sets, predictions):
prediction = set([(alignment[1], alignment[2]) for alignment in prediction])
metric.update(sure=gold[0], probable=gold[1], predicted=prediction)
print(metric.aer())
return metric.aer()
def aer(self):
gold_sets = aer.read_naacl_alignments(self.path_true)
validation_corpus = read_data(self.english_val, self.french_val)
predictions = []
for E, F in validation_corpus.corpus:
values, _ = self.viterbi_alignment(E.s, F.s, split=False)
links = set()
for j in range(1, values.shape[0]):
winner = np.argwhere(values[j] == np.max(values[j])).flatten()
score = np.abs(winner - j)
best = winner[np.argmin(score)]
links.add((j, best + 1))
predictions.append(links)
metric = aer.AERSufficientStatistics()
# then we iterate over the corpus
for gold, pred in zip(gold_sets, predictions):
metric.update(sure=gold[0], probable=gold[1], predicted=pred)
return metric.aer()