def main(bert_model, conll_path, simple_connectives, sense_lvl):
logger = init_logger()
logger.info('Load data')
docs_val = load_bert_conll_dataset(os.path.join(conll_path, 'en.dev'),
simple_connectives=simple_connectives,
cache_dir=os.path.join(
conll_path,
f'en.dev.{bert_model}.joblib'),
bert_model=bert_model,
sense_level=sense_lvl)
docs_test = load_bert_conll_dataset(os.path.join(conll_path, 'en.test'),
simple_connectives=simple_connectives,
cache_dir=os.path.join(
conll_path,
f'en.test.{bert_model}.joblib'),
bert_model=bert_model,
sense_level=sense_lvl)
docs_blind = load_bert_conll_dataset(
os.path.join(conll_path, 'en.blind-test'),
simple_connectives=simple_connectives,
cache_dir=os.path.join(conll_path,
f'en.blind-test.{bert_model}.joblib'),
bert_model=bert_model,
sense_level=sense_lvl)
docs_train = load_bert_conll_dataset(os.path.join(conll_path, 'en.train'),
simple_connectives=simple_connectives,
cache_dir=os.path.join(
conll_path,
f'en.train.{bert_model}.joblib'),
bert_model=bert_model,
sense_level=sense_lvl)
logger.info('Init model')
clf = ImplicitArgumentExtractor(window_length=100,
input_dim=docs_val[0].get_embedding_dim(),
hidden_dim=128,
rnn_dim=256)
clf.batch_size = 256
logger.info('Train model')
clf.fit(docs_train, docs_val)
clf.save('models/nn')
logger.info('Evaluation on TEST')
clf.score(docs_val)
for title, docs_eval in [('TEST', docs_test), ('BLIND', docs_blind)]:
logger.info(f'Evaluate parser {title}')
preds = [d.with_relations(clf.parse(d)) for d in docs_eval]
for threshold in [0.7, 0.95]:
res = evaluate_docs([
d.with_relations([
r for r in d.relations if r.type in ['Implicit', 'EntRel']
]) for d in docs_eval
],
preds,
threshold=threshold)
print_results(res, title=f'{title}-{threshold}')
def main(conll_path):
logger = init_logger()
docs_val = load_bert_conll_dataset(os.path.join(conll_path, 'en.dev'),
cache_dir=os.path.join(
conll_path,
'en.dev.bert-base-cased.joblib'))
docs_train = load_bert_conll_dataset(
os.path.join(conll_path, 'en.train'),
cache_dir=os.path.join(conll_path, 'en.train.bert-base-cased.joblib'))
clf = ConnectiveSenseClassifier(input_dim=docs_val[0].get_embedding_dim(),
used_context=2)
logger.info('Train model')
clf.fit(docs_train, docs_val)
logger.info('Evaluation on TRAIN')
clf.score(docs_train)
logger.info('Evaluation on TEST')
clf.score(docs_val)
# logger.info('Parse one document')
# print(docs_val[0].to_json())
print(clf.parse(docs_val[0], []))
preds = [d.with_relations(clf.parse(d)) for d in docs_val]
print_results(evaluate_docs(docs_val, preds))
def main(conll_path, pred_path, threshold, simple_connectives, sense_level):
gold_docs = load_parsed_conll_dataset(conll_path, simple_connectives=simple_connectives, sense_level=sense_level)
pred_docs = load_documents(get_fh(pred_path))
pred_doc_ids = {doc.doc_id for doc in pred_docs}
gold_docs = [doc for doc in gold_docs if doc.doc_id in pred_doc_ids]
if not pred_docs or not gold_docs:
logger.warning('No documents found')
return
print_results(evaluate_docs(
[doc.with_relations([r for r in doc.relations if r.is_explicit()]) for doc in gold_docs],
[doc.with_relations([r for r in doc.relations if r.is_explicit()]) for doc in pred_docs],
threshold=threshold), title='EXPLICIT')
print_results(evaluate_docs(
[doc.with_relations([r for r in doc.relations if not r.is_explicit()]) for doc in gold_docs],
[doc.with_relations([r for r in doc.relations if not r.is_explicit()]) for doc in pred_docs],
threshold=threshold), title='NON-EXPLICIT')
print_results(evaluate_docs(gold_docs, pred_docs, threshold=threshold), title='ALL')
def main(bert_model, conll_path, save_path, simple_connectives, sense_lvl=2):
logger = init_logger()
logger.info('Load train data')
docs_train = load_bert_conll_dataset(os.path.join(conll_path, 'en.train'),
simple_connectives=simple_connectives,
cache_dir=os.path.join(
conll_path,
f'en.train.{bert_model}.joblib'),
bert_model=bert_model,
sense_level=sense_lvl)
logger.info('Load dev data')
docs_val = load_bert_conll_dataset(os.path.join(conll_path, 'en.dev'),
simple_connectives=simple_connectives,
cache_dir=os.path.join(
conll_path,
f'en.dev.{bert_model}.joblib'),
bert_model=bert_model,
sense_level=sense_lvl)
docs_test = load_bert_conll_dataset(os.path.join(conll_path, 'en.test'),
simple_connectives=simple_connectives,
cache_dir=os.path.join(
conll_path,
f'en.test.{bert_model}.joblib'),
bert_model=bert_model,
sense_level=sense_lvl)
docs_blind = load_bert_conll_dataset(
os.path.join(conll_path, 'en.blind-test'),
simple_connectives=simple_connectives,
cache_dir=os.path.join(conll_path,
f'en.blind-test.{bert_model}.joblib'),
bert_model=bert_model,
sense_level=sense_lvl)
logger.info('Init model')
parser = ParserPipeline([
ConnectiveSenseClassifier(input_dim=docs_val[0].get_embedding_dim(),
used_context=1),
ConnectiveArgumentExtractor(window_length=100,
input_dim=docs_val[0].get_embedding_dim(),
hidden_dim=256,
rnn_dim=512,
ckpt_path=save_path),
NonExplicitRelationClassifier(
input_dim=docs_val[0].get_embedding_dim(), arg_length=50),
])
logger.info('Train model')
parser.fit(docs_train, docs_val)
parser.save(save_path)
logger.info('LOAD model')
parser.load(save_path)
for title, docs_eval in [('TEST', docs_test), ('BLIND', docs_blind)]:
logger.info(f'Evaluate parser {title}')
test_preds = parser.parse(docs_eval)
for threshold in [0.7, 0.9]:
print_results(evaluate_docs([
d.with_relations(d.get_explicit_relations()) for d in docs_eval
], [
d.with_relations(d.get_explicit_relations())
for d in test_preds
],
threshold=threshold),
title=f'{title}-EXPLICIT-{threshold}')
print_results(evaluate_docs([
d.with_relations(
[r for r in d.relations if not r.is_explicit()])
for d in docs_eval
], [
d.with_relations(
[r for r in d.relations if not r.is_explicit()])
for d in test_preds
],
threshold=threshold),
title=f'{title}-NON-EXPLICIT-{threshold}')