def load_train_val_examples(args):
lines = []
for guid, text, _, entities in train_data_generator(args.train_file, args.eval_file):
sl = LabeledText(guid, text, entities)
lines.append({'guid': guid, 'text': text, 'entities': entities})
allow_overlap = args.allow_overlap
if args.num_augements > 0:
allow_overlap = False
train_base_examples = load_ner_labeled_examples(
lines,
ner_labels,
seg_len=args.seg_len,
seg_backoff=args.seg_backoff,
num_augements=args.num_augements,
allow_overlap=allow_overlap)
train_examples, val_examples = split_train_eval_examples(
train_base_examples,
train_rate=args.train_rate,
fold=args.fold,
shuffle=True)
logger.info(f"Loaded {len(train_examples)} train examples, "
f"{len(val_examples)} val examples.")
return train_examples, val_examples
def train_data_generator(train_file):
lines = load_json_file(train_file)
for i, x in enumerate(tqdm(lines)):
guid = str(i)
text = clean_text(x['text'])
sl = LabeledText(guid, text)
# -------------------- 训练数据json格式 --------------------
# {
# "text": "万通地产设计总监刘克峰；",
# "label": {
# "name": {
# "刘克峰": [[8, 10]]
# },
# "company": {
# "万通地产": [[0, 3]]
# },
# "position": {
# "设计总监": [[4, 7]]
# }
# }
# }
entities = []
classes = x['label'].keys()
for c in classes:
c_labels = x['label'][c]
# logger.debug(f"c_labels:{c_labels}")
for label, span in c_labels.items():
x0, x1 = span[0]
sl.add_entity(c, x0, x1)
yield str(i), text, None, sl.entities
def train_data_from_last_generator(train_file):
train_files = [
'./data/event_element_sampling_0713.json',
]
for train_file in train_files:
tagged_train_json_data = json.load(open(train_file, 'r'))
all_labels = tagged_train_json_data['labelCategories']
id2label = {x['id']: x['text'] for x in all_labels}
all_entities = tagged_train_json_data['labels']
content = tagged_train_json_data['content']
# re_b = '(\\n[-]+ yanbao\\d\\d\\d\\.txt Begin [-]+\\n\\n)'
# re_e = '(\\n[-]+ yanbao\\d\\d\\d\\.txt End [-]+\\n\\n)'
re_b = '(\\n[-]+ [\d]+ Begin [-]+\\n\\n)'
re_e = '(\\n[-]+ [\d]+ End [-]+\\n\\n)'
b_list = []
for x in re.finditer(re_b, content):
b_list.append((x.start(), x.end()))
e_list = []
for x in re.finditer(re_e, content):
e_list.append((x.start(), x.end()))
pages = [(x_b[0], x_b[1], x_e[0], x_e[1])
for x_b, x_e in zip(b_list, e_list)]
logger.warning(f"pages: {pages}")
for i, page in enumerate(pages):
head_x0, head_x1, tail_x0, tail_x1 = page
guid = f"{i}"
text = content[head_x1:tail_x0]
sl = LabeledText(guid, text)
for entity in all_entities:
s = entity['startIndex']
e = entity['endIndex'] - 1
assert e >= s
if s >= head_x1 and e < tail_x0:
sl.add_entity(id2label[entity['categoryId']], s - head_x1,
e - head_x1)
yield guid, text, None, sl.entities
def train_data_generator(train_file):
lines = load_json_file(train_file)
for i, x in enumerate(tqdm(lines)):
guid = str(i)
text = clean_text(x['originalText'])
sl = LabeledText(guid, text)
entities = x['entities']
for entity in entities:
start_pos = entity['start_pos']
end_pos = entity['end_pos'] - 1
category = entity['label_type']
sl.add_entity(category, start_pos, end_pos)
yield str(i), text, None, sl.entities
def train_data_generator(train_file):
lines = load_json_file(train_file)
for i, x in enumerate(tqdm(lines)):
guid = str(i)
text = clean_text(x['text'])
sl = LabeledText(guid, text)
entities = []
classes = x['label'].keys()
for c in classes:
c_labels = x['label'][c]
# logger.debug(f"c_labels:{c_labels}")
for label, span in c_labels.items():
x0, x1 = span[0]
sl.add_entity(c, x0, x1)
print("index: ", str(i), ", text: ", text, ", entities: ", sl.entities)
break
def load_eval_examples(eval_text_file, eval_bio_file):
lines = []
for guid, text, _, entities in train_data_generator(
eval_text_file, eval_bio_file):
sl = LabeledText(guid, text, entities)
lines.append({'guid': guid, 'text': text, 'entities': entities})
train_base_examples = load_ner_labeled_examples(
lines,
ner_labels,
seg_len=args.seg_len,
seg_backoff=args.seg_backoff,
num_augments=0,
allow_overlap=False)
eval_examples = train_base_examples
logger.info(f"Loaded {len(eval_examples)} eval examples")
return eval_examples
def train_data_generator(train_file):
data = load_ner_train_data(train_file)
for i, x in enumerate(tqdm(data)):
guid = x[0]
text = clean_text(x[1])
sl = LabeledText(guid, text)
entities = x[2]
for entity in entities:
c = entity[0]
x0 = int(entity[1])
x1 = int(entity[2])
sl.add_entity(c, x0, x1)
#print("index: ", str(i), ", text: ", text, ", entities: ", sl.entities)
#if i > 10:
#break
yield guid, text, None, sl.entities
def train_data_generator_0(train_file):
with open(train_file, 'r') as fr:
lines = fr.readlines()
for line in tqdm(lines, desc=f"train & eval"):
d = json.loads(line)
guid = d['doc_id']
text = clean_text(d['content'])
seg_text = text
seg_labels = []
for e in d['events']:
event_type = e['event_type']
# if event_type not in ['破产清算']: # ['股东减持', '股东增持']:
# continue
for k, v in e.items():
if not v:
continue
if k not in ['event_id', 'event_type']:
label = '_'.join((event_type, k))
# if label not in ner_labels:
# ner_labels.append(label)
i0 = seg_text.find(v)
while i0 >= 0:
# if i0 >= 0:
if len(v) == 1:
# if labels[i0] == 'O':
# labels[i0] = f"S-{label}"
pass
else:
seg_labels.append((label, i0, i0 + len(v) - 1))
# break
i0 = seg_text.find(v, i0 + len(v))
sl = LabeledText(guid, text)
for category, start_char, end_char in seg_labels:
sl.add_entity(category, start_char, end_char)
yield guid, text, None, sl.entities
def train_data_generator(train_text_file, train_bio_file):
texts = load_texts(train_text_file)
cond, labels = load_bioattr_labels(train_bio_file)
for i, x in enumerate(tqdm(texts)):
guid = str(i)
text = clean_text(x)
sl = LabeledText(guid, text)
entities = labels[i]
for entity in entities:
c = entity[0]
x0 = int(entity[1])
x1 = int(entity[2])
sl.add_entity(c, x0, x1)
#print("index: ", str(i), ", text: ", text, ", entities: ", sl.entities)
#if i > 10:
#break
yield str(i), text, None, sl.entities
def train_data_generator(train_file):
data, _ = load_data(train_file)
for i, x in enumerate(tqdm(data)):
guid = x[1]
text = clean_text(x[0])
level1 = x[2]
level2 = x[3]
level3 = x[4]
sl = LabeledText(guid, text)
entities = x[5]
for entity in entities:
c = level1 + "_" + level2 + "_" + level3 + "_" + entity[2]
x0 = int(entity[0])
x1 = int(entity[1])
sl.add_entity(c, x0, x1)
#print("index: ", str(i), ", text: ", text, ", entities: ", sl.entities)
#if i > 10:
#break
yield str(i), text, None, sl.entities
def train_data_generator(train_file):
data = load_data(train_file)
for i, x in enumerate(tqdm(data)):
guid = str(i)
text = clean_text(x[0])
arguments = x[1]
sl = LabeledText(guid, text)
entities = []
for key, value in arguments.items():
argument = key
event_type = value[0]
role = value[1]
start_index = int(value[2])
c = event_type + "_" + role
x0 = start_index
x1 = start_index + len(argument) - 1
sl.add_entity(c, x0, x1)
#print("index: ", str(i), ", text: ", text, ", entities: ", sl.entities)
#if i > 10:
#break
yield str(i), text, None, sl.entities