def test_sample_from_dataset_positive02(self):
base_dir = os.path.join(os.path.dirname(__file__), 'testdata')
_, y = load_dataset_from_json(
os.path.join(base_dir, 'true_named_entities.json'))
subset_index = sample_from_dataset(y=y,
n=(len(y) + 1) * 2,
n_restarts=10)
self.assertIsInstance(subset_index, np.ndarray)
self.assertEqual(len(y), len(subset_index))
self.assertEqual(set(range(len(y))), set(subset_index.tolist()))
def test_calculate_prediction_quality(self):
base_dir = os.path.join(os.path.dirname(__file__), 'testdata')
X_true, y_true = load_dataset_from_json(os.path.join(base_dir, 'true_named_entities.json'))
X_pred, y_pred = load_dataset_from_json(os.path.join(base_dir, 'predicted_named_entities.json'))
self.assertEqual(X_true, X_pred)
f1, precision, recall, quality_by_entities = calculate_prediction_quality(
y_true, y_pred, ('LOCATION', 'PERSON', 'ORG')
)
self.assertIsInstance(f1, float)
self.assertIsInstance(precision, float)
self.assertIsInstance(recall, float)
self.assertAlmostEqual(f1, 0.842037, places=3)
self.assertAlmostEqual(precision, 0.908352, places=3)
self.assertAlmostEqual(recall, 0.784746, places=3)
self.assertIsInstance(quality_by_entities, dict)
self.assertEqual({'LOCATION', 'PERSON', 'ORG'}, set(quality_by_entities.keys()))
f1_macro = 0.0
precision_macro = 0.0
recall_macro = 0.0
for ne_type in quality_by_entities:
self.assertIsInstance(quality_by_entities[ne_type], tuple)
self.assertEqual(len(quality_by_entities[ne_type]), 3)
self.assertIsInstance(quality_by_entities[ne_type][0], float)
self.assertIsInstance(quality_by_entities[ne_type][1], float)
self.assertIsInstance(quality_by_entities[ne_type][2], float)
self.assertLess(quality_by_entities[ne_type][0], 1.0)
self.assertGreater(quality_by_entities[ne_type][0], 0.0)
self.assertLess(quality_by_entities[ne_type][1], 1.0)
self.assertGreater(quality_by_entities[ne_type][1], 0.0)
self.assertLess(quality_by_entities[ne_type][2], 1.0)
self.assertGreater(quality_by_entities[ne_type][2], 0.0)
f1_macro += quality_by_entities[ne_type][0]
precision_macro += quality_by_entities[ne_type][1]
recall_macro += quality_by_entities[ne_type][2]
f1_macro /= float(len(quality_by_entities))
precision_macro /= float(len(quality_by_entities))
recall_macro /= float(len(quality_by_entities))
for ne_type in quality_by_entities:
self.assertGreater(abs(quality_by_entities[ne_type][0] - f1_macro), 1e-4)
self.assertGreater(abs(quality_by_entities[ne_type][1] - precision_macro), 1e-4)
self.assertGreater(abs(quality_by_entities[ne_type][2] - recall_macro), 1e-4)
def test_sample_from_dataset_positive01(self):
base_dir = os.path.join(os.path.dirname(__file__), 'testdata')
_, y = load_dataset_from_json(
os.path.join(base_dir, 'true_named_entities.json'))
subset_index = sample_from_dataset(y=y, n=3, n_restarts=10)
self.assertIsInstance(subset_index, np.ndarray)
self.assertEqual(3, len(subset_index))
self.assertGreater(len(y), len(subset_index))
self.assertEqual(len(subset_index), len(set(subset_index.tolist())))
true_set_of_classes = {'ORG', 'PERSON', 'LOCATION'}
subset_of_classes = set()
for idx in subset_index:
subset_of_classes |= set(y[idx].keys())
self.assertEqual(true_set_of_classes, subset_of_classes)
def test_positive01(self):
base_dir = os.path.join(os.path.dirname(__file__), 'testdata')
_, y = load_dataset_from_json(os.path.join(base_dir, 'true_named_entities.json'))
train_index, test_index = split_dataset(y, 0.3, 10)
self.assertIsInstance(train_index, np.ndarray)
self.assertIsInstance(test_index, np.ndarray)
self.assertEqual(len(y), len(train_index) + len(test_index))
self.assertEqual(len(train_index), len(set(train_index.tolist())))
self.assertEqual(len(test_index), len(set(test_index.tolist())))
self.assertEqual(0, len(set(train_index.tolist()) & set(test_index.tolist())))
true_set_of_classes = {'ORG', 'PERSON', 'LOCATION'}
set_of_classes_for_training = set()
for idx in train_index:
set_of_classes_for_training |= set(y[idx].keys())
set_of_classes_for_testing = set()
for idx in test_index:
set_of_classes_for_testing |= set(y[idx].keys())
self.assertEqual(set_of_classes_for_training, set_of_classes_for_testing)
self.assertEqual(true_set_of_classes, set_of_classes_for_training)
self.assertEqual(true_set_of_classes, set_of_classes_for_testing)
def train(factrueval2016_devset_dir: str,
split_by_paragraphs: bool,
elmo_will_be_tuned: bool,
use_additional_features: bool,
max_epochs: int,
patience: int,
batch_size: int,
lr: float,
l2: float,
gpu_memory_frac: float,
model_name: str,
collection3_dir: Union[str, None] = None) -> ELMo_NER:
if os.path.isfile(model_name):
with open(model_name, 'rb') as fp:
recognizer = pickle.load(fp)
assert isinstance(recognizer, ELMo_NER)
print('The NER has been successfully loaded from the file `{0}`...'.
format(model_name))
print('')
else:
temp_json_name = tempfile.NamedTemporaryFile(mode='w').name
try:
factrueval2016_to_json(factrueval2016_devset_dir, temp_json_name,
split_by_paragraphs)
X, y = load_dataset_from_json(temp_json_name)
finally:
if os.path.isfile(temp_json_name):
os.remove(temp_json_name)
print('The FactRuEval-2016 data for training have been loaded...')
print('Number of samples is {0}.'.format(len(y)))
print('')
max_number_of_tokens = 0
pipeline = create_udpipe_pipeline('ru')
for cur in X:
spacy_doc = pipeline(cur)
n_tokens = 0
for _ in spacy_doc:
n_tokens += 1
del spacy_doc
if n_tokens > max_number_of_tokens:
max_number_of_tokens = n_tokens
del pipeline
print('Maximal number of tokens is {0}.'.format(max_number_of_tokens))
n_tokens = 2
while n_tokens < max_number_of_tokens:
n_tokens *= 2
elmo_hub_module_handle = 'http://files.deeppavlov.ai/deeppavlov_data/elmo_ru-news_wmt11-16_1.5M_steps.tar.gz'
recognizer = ELMo_NER(finetune_elmo=elmo_will_be_tuned,
batch_size=batch_size,
l2_reg=l2,
max_seq_length=n_tokens,
elmo_hub_module_handle=elmo_hub_module_handle,
validation_fraction=0.25,
max_epochs=max_epochs,
patience=patience,
gpu_memory_frac=gpu_memory_frac,
verbose=True,
random_seed=42,
lr=lr,
udpipe_lang='ru',
use_additional_features=use_additional_features)
if collection3_dir is None:
recognizer.fit(X, y)
else:
X_train, y_train = load_dataset_from_brat(collection3_dir,
split_by_paragraphs=True)
if not split_by_paragraphs:
X_train, y_train = divide_dataset_by_sentences(
X_train, y_train, sent_tokenize_func=ru_sent_tokenize)
for sample_idx in range(len(y_train)):
new_y_sample = dict()
for ne_type in sorted(list(y_train[sample_idx].keys())):
if ne_type == 'PER':
new_y_sample['PERSON'] = y_train[sample_idx][ne_type]
elif ne_type == 'LOC':
new_y_sample['LOCATION'] = y_train[sample_idx][ne_type]
else:
new_y_sample[ne_type] = y_train[sample_idx][ne_type]
y_train[sample_idx] = new_y_sample
del new_y_sample
print('The Collection3 data for training have been loaded...')
print('Number of samples is {0}.'.format(len(y_train)))
print('')
recognizer.fit(X_train, y_train, validation_data=(X, y))
with open(model_name, 'wb') as fp:
pickle.dump(recognizer, fp)
print('')
print(
'The NER has been successfully fitted and saved into the file `{0}`...'
.format(model_name))
print('')
return recognizer
def recognize(factrueval2016_testset_dir: str, split_by_paragraphs: bool,
recognizer: ELMo_NER, results_dir: str):
temp_json_name = tempfile.NamedTemporaryFile(mode='w').name
try:
factrueval2016_to_json(factrueval2016_testset_dir, temp_json_name,
split_by_paragraphs)
with codecs.open(temp_json_name,
mode='r',
encoding='utf-8',
errors='ignore') as fp:
data_for_testing = json.load(fp)
_, true_entities = load_dataset_from_json(temp_json_name)
finally:
if os.path.isfile(temp_json_name):
os.remove(temp_json_name)
texts = []
additional_info = []
for cur_document in data_for_testing:
base_name = os.path.join(results_dir,
cur_document['base_name'] + '.task1')
for cur_paragraph in cur_document['paragraph_bounds']:
texts.append(
cur_document['text'][cur_paragraph[0]:cur_paragraph[1]])
additional_info.append((base_name, cur_paragraph))
print('Data for final testing have been loaded...')
print('Number of samples is {0}.'.format(len(true_entities)))
print('')
predicted_entities = recognizer.predict(texts)
assert len(predicted_entities) == len(true_entities)
f1, precision, recall, quality_by_entities = calculate_prediction_quality(
true_entities, predicted_entities, recognizer.classes_list_)
print('All entities:')
print(' F1-score is {0:.2%}.'.format(f1))
print(' Precision is {0:.2%}.'.format(precision))
print(' Recall is {0:.2%}.'.format(recall))
for ne_type in sorted(list(quality_by_entities.keys())):
print(' {0}'.format(ne_type))
print(' F1-score is {0:.2%}.'.format(
quality_by_entities[ne_type][0]))
print(' Precision is {0:.2%}.'.format(
quality_by_entities[ne_type][1]))
print(' Recall is {0:.2%}.'.format(quality_by_entities[ne_type][2]))
results_for_factrueval_2016 = dict()
for sample_idx, cur_result in enumerate(predicted_entities):
base_name, paragraph_bounds = additional_info[sample_idx]
for entity_type in cur_result:
if entity_type == 'ORG':
prepared_entity_type = 'org'
elif entity_type == 'PERSON':
prepared_entity_type = 'per'
elif entity_type == 'LOCATION':
prepared_entity_type = 'loc'
else:
prepared_entity_type = None
if prepared_entity_type is None:
raise ValueError(
'`{0}` is unknown entity type!'.format(entity_type))
for entity_bounds in cur_result[entity_type]:
postprocessed_entity = (prepared_entity_type,
entity_bounds[0] + paragraph_bounds[0],
entity_bounds[1] - entity_bounds[0])
if base_name in results_for_factrueval_2016:
results_for_factrueval_2016[base_name].append(
postprocessed_entity)
else:
results_for_factrueval_2016[base_name] = [
postprocessed_entity
]
for base_name in results_for_factrueval_2016:
with codecs.open(base_name,
mode='w',
encoding='utf-8',
errors='ignore') as fp:
for cur_entity in sorted(results_for_factrueval_2016[base_name],
key=lambda it: (it[1], it[2], it[0])):
fp.write('{0} {1} {2}\n'.format(cur_entity[0], cur_entity[1],
cur_entity[2]))
def train(factrueval2016_devset_dir: str,
split_by_paragraphs: bool,
bert_will_be_tuned: bool,
use_lang_features: bool,
use_shapes: bool,
lstm_layer_size: Union[int, None],
l2: float,
max_epochs: int,
patience: int,
batch_size: int,
gpu_memory_frac: float,
model_name: str,
collection3_dir: Union[str, None] = None,
n_max_samples: int = 0) -> BERT_NER:
if os.path.isfile(model_name):
with open(model_name, 'rb') as fp:
recognizer = pickle.load(fp)
assert isinstance(recognizer, BERT_NER)
print('The NER has been successfully loaded from the file `{0}`...'.
format(model_name))
print('')
else:
temp_json_name = tempfile.NamedTemporaryFile(mode='w').name
try:
factrueval2016_to_json(factrueval2016_devset_dir, temp_json_name,
split_by_paragraphs)
X, y = load_dataset_from_json(temp_json_name)
finally:
if os.path.isfile(temp_json_name):
os.remove(temp_json_name)
print('The FactRuEval-2016 data for training have been loaded...')
print('Number of samples is {0}.'.format(len(y)))
print('')
if BERT_NER.PATH_TO_BERT is None:
bert_hub_module_handle = 'https://tfhub.dev/google/bert_multi_cased_L-12_H-768_A-12/1'
else:
bert_hub_module_handle = None
recognizer = BERT_NER(finetune_bert=bert_will_be_tuned,
batch_size=batch_size,
l2_reg=l2,
bert_hub_module_handle=bert_hub_module_handle,
lstm_units=lstm_layer_size,
validation_fraction=0.25,
max_epochs=max_epochs,
patience=patience,
gpu_memory_frac=gpu_memory_frac,
verbose=True,
random_seed=42,
lr=3e-6 if bert_will_be_tuned else 1e-4,
udpipe_lang='ru',
use_nlp_features=use_lang_features,
use_shapes=use_shapes)
if collection3_dir is None:
if n_max_samples > 0:
train_index, test_index = split_dataset(
y=y, test_part=recognizer.validation_fraction)
X_train = np.array(X, dtype=object)[train_index]
y_train = np.array(y, dtype=object)[train_index]
X_val = np.array(X, dtype=object)[test_index]
y_val = np.array(y, dtype=object)[test_index]
del train_index, test_index
index = sample_from_dataset(y=y_train, n=n_max_samples)
recognizer.fit(X_train[index],
y_train[index],
validation_data=(X_val, y_val))
else:
recognizer.fit(X, y)
else:
X_train, y_train = load_dataset_from_brat(collection3_dir,
split_by_paragraphs=True)
if not split_by_paragraphs:
X_train, y_train = divide_dataset_by_sentences(
X_train, y_train, sent_tokenize_func=ru_sent_tokenize)
for sample_idx in range(len(y_train)):
new_y_sample = dict()
for ne_type in sorted(list(y_train[sample_idx].keys())):
if ne_type == 'PER':
new_y_sample['PERSON'] = y_train[sample_idx][ne_type]
elif ne_type == 'LOC':
new_y_sample['LOCATION'] = y_train[sample_idx][ne_type]
else:
new_y_sample[ne_type] = y_train[sample_idx][ne_type]
y_train[sample_idx] = new_y_sample
del new_y_sample
print('The Collection3 data for training have been loaded...')
print('Number of samples is {0}.'.format(len(y_train)))
print('')
if n_max_samples > 0:
index = sample_from_dataset(y=y_train, n=n_max_samples)
X_train = np.array(X_train, dtype=object)[index]
y_train = np.array(y_train, dtype=object)[index]
del index
recognizer.fit(X_train, y_train, validation_data=(X, y))
with open(model_name, 'wb') as fp:
pickle.dump(recognizer, fp)
print('')
print(
'The NER has been successfully fitted and saved into the file `{0}`...'
.format(model_name))
print('')
return recognizer