def test_not_semeval_file(self, conflict: bool):
# Test that it will raise a SyntaxError as the file does not follow
# SemEval format.
unreadable_fp = Path(self.DATA_PATH_DIR,
'not_semeval_14_example.xml')
with pytest.raises(SyntaxError):
semeval_2014(unreadable_fp, conflict)
def test_unreadable_file(self, conflict: bool):
# Test that it will raise a ParseError is not formatted correctly e.g.
# Contains mismatched tags
unreadable_fp = Path(self.DATA_PATH_DIR,
'unpassable_semeval_14_example.xml')
with pytest.raises(ParseError):
semeval_2014(unreadable_fp, conflict)
def test_read_from_file(self, conflict: bool):
data_fp = Path(self.DATA_PATH_DIR, 'semeval_14_example.xml')
target_text_collection = semeval_2014(data_fp, conflict)
assert len(target_text_collection) == 5
_ids = ['3121', '2777', '2534', '2634', '1793']
assert list(target_text_collection.keys()) == _ids
true_answers = self._target_answer(conflict=conflict)
for answer_key, answer in true_answers.items():
test_answer = target_text_collection[answer_key]
for key in answer:
assert answer[key] == test_answer[key]
help='File Path to the output predictions')
args = parser.parse_args()
dataset_name = args.dataset_name
model_name = f'{dataset_name} model'
model = AllenNLPModel(model_name, args.model_config, 'target-tagger',
args.model_save_dir)
if dataset_name == 'semeval_2014':
if not args.train_fp or not args.test_fp:
raise ValueError('If training and predicting for the SemEval '
'datasets the training and test file paths must '
'be given')
# As we are performing target extraction we use the conflict polarity
# targets like prior work
train_data = semeval_2014(args.train_fp, conflict=True)
test_data = semeval_2014(args.test_fp, conflict=True)
else:
temp_election_directory = Path('.', 'data', 'twitter_election_dataset')
train_data = wang_2017_election_twitter_train(temp_election_directory)
test_data = wang_2017_election_twitter_test(temp_election_directory)
if not args.model_save_dir.is_dir():
# Use the same size validation as the test data
test_size = len(test_data)
# Create the train and validation splits
train_data = list(train_data.values())
train_data, val_data = train_test_split(train_data,
test_size=test_size)
train_data = TargetTextCollection(train_data)
val_data = TargetTextCollection(val_data)
for dataset_name in dataset_names:
for split_name in split_names:
dataset_fp = config.neural_dataset_dir / f'{dataset_name} {split_name}.json'
if dataset_fp.exists():
continue
dataset = None
if dataset_name == 'Election':
if split_name == 'train':
election_train = parsers.election_train(config.ELECTION,
name='Election Train')
print(f'Number of targets before {len(election_train)}')
election_train = get_targets_from_spans(election_train)
temp_fp = config.neural_dataset_dir / 'Temp Election train.xml'
write_data.semeval_14(temp_fp, election_train)
dataset = semeval_2014(temp_fp, conflict=False)
else:
election_test = parsers.election_test(config.ELECTION,
name='Election Train')
print(f'Number of targets before {len(election_test)}')
election_test = get_targets_from_spans(election_test)
temp_fp = config.neural_dataset_dir / 'Temp Election test.xml'
write_data.semeval_14(temp_fp, election_test)
dataset = semeval_2014(temp_fp, conflict=False)
elif dataset_name == 'Dong':
dong_dataset = parsers.dong(
dataset_fp_mapper[f'{dataset_name} {split_name}'])
print(f'Number of targets before {len(dong_dataset)}')
new_dong_dataset = []
for value in dong_dataset.data_dict():
target_spans = value['spans']
args = parser.parse_args()
dataset_name = args.dataset_name
model_name = f'{dataset_name} model'
model = AllenNLPModel(model_name, args.model_config, 'target-tagger',
args.model_save_dir)
if not args.model_save_dir.is_dir():
if dataset_name == 'semeval_2014':
if not args.train_fp or not args.test_fp:
raise ValueError(
'If training and predicting for the SemEval '
'datasets the training and test file paths must '
'be given')
train_data = semeval_2014(args.train_fp, False)
test_data = semeval_2014(args.test_fp, False)
else:
temp_election_directory = Path('/tmp/election_dataset_dir')
train_data = wang_2017_election_twitter_train(
temp_election_directory)
test_data = wang_2017_election_twitter_test(
temp_election_directory)
# Use the same size validation as the test data
test_size = len(test_data)
# Create the train and validation splits
train_data = list(train_data.values())
train_data, val_data = train_test_split(train_data,
test_size=test_size)
train_data = TargetTextCollection(train_data)
val_data = TargetTextCollection(val_data)
from target_extraction.data_types import TargetTextCollection
from target_extraction.dataset_parsers import semeval_2014
from target_extraction.tokenizers import spacy_tokenizer
from target_extraction.allen import AllenNLPModel
semeval_2014_dir = Path(
'..',
'original_target_datasets',
'semeval_2014',
).resolve()
train_fp = Path(semeval_2014_dir,
"SemEval'14-ABSA-TrainData_v2 & AnnotationGuidelines",
"Laptop_Train_v2.xml")
test_fp = Path(semeval_2014_dir, "ABSA_Gold_TestData", 'Laptops_Test_Gold.xml')
train_data = semeval_2014(train_fp, False)
test_data = semeval_2014(test_fp, False)
test_size = len(test_data)
print(f'Size of train {len(train_data)}, size of test {test_size}')
train_data = list(train_data.values())
train_data, val_data = train_test_split(train_data, test_size=test_size)
train_data = TargetTextCollection(train_data)
val_data = TargetTextCollection(val_data)
datasets = [train_data, val_data, test_data]
tokenizer = spacy_tokenizer()
sizes = []
for dataset in datasets:
dataset.tokenize(tokenizer)
value['target'] = value['text'][target_spans[0][0]: target_spans[0][1]]
new_dong_dataset.append(bella.data_types.Target(**value))
dataset = bella.data_types.TargetCollection(new_dong_dataset)
else:
dataset = parsers.semeval_14(dataset_fp_mapper[f'{dataset_name} train'])
assert dataset is not None
dataset_size = len(dataset)
test_split_size = size_of_small / dataset_size
_, small_dataset = bella.data_types.TargetCollection.split_dataset(dataset, test_split_size,
random=False)
assert len(small_dataset) == size_of_small
write_data.semeval_14(dataset_fp, small_dataset)
print(f'Number of targets in {dataset_name} non-neural training dataset {len(small_dataset)}')
neural_small_dataset = semeval_2014(dataset_fp, conflict=False)
# Just making sure each sentence contains at least one target.
assert neural_small_dataset.one_sample_per_span(remove_empty=True).number_targets() == neural_small_dataset.number_targets()
assert len(neural_small_dataset) == len(neural_small_dataset.samples_with_targets())
print(f'Number of targets with new format {neural_small_dataset.number_targets()}')
# For reproducibility reasons
random_state = 42
# Validation size is 20% of the training set size based on sentences not
# targets
train_size = len(neural_small_dataset)
val_size = int(train_size * 0.2)
train_dataset = list(neural_small_dataset.values())
train, val = train_test_split(train_dataset, test_size=val_size,
random_state=random_state)
train_dataset = target_extraction.data_types.TargetTextCollection(train)
val_dataset = target_extraction.data_types.TargetTextCollection(val)