def load_from_splits(paths, original_test_filename, model_predicted_filename):
sentence_potential_mistake_count = defaultdict(int)
for path in paths:
original_test = os.path.join(path, original_test_filename)
model_predicted = os.path.join(path, model_predicted_filename)
assert os.path.exists(original_test)
assert os.path.exists(model_predicted)
original_test = load_dataset_from_column(original_test)
model_predicted = load_dataset_from_column(
model_predicted,
schema="none") # since there may be invalid label sequences.
for (original_sentence,
original_labels), (model_sentence,
model_labels) in zip(original_test,
model_predicted):
assert ' '.join(original_sentence) == ' '.join(model_sentence)
if ' '.join(original_labels) != ' '.join(model_labels):
sentence_potential_mistake_count[' '.join(
original_sentence)] += 1
return sentence_potential_mistake_count
def form_weighted_train_set(train_files, train_file_schema, eps,
mistake_count):
for train_file in train_files:
assert os.path.exists(train_file)
train_set = []
for train_file in train_files:
train_set.extend(
load_dataset_from_column(train_file, schema="train_file_schema"))
weighted_train_set = []
for sentence, labels in train_set:
mistakes = mistake_count.get(' '.join(sentence), 0)
weight = eps**mistakes
weighted_train_set.append([sentence, labels, [weight] * len(labels)])
return weighted_train_set