def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path, encoding='utf8') as f:
for line in f:
example_json = json.loads(line)
label = str(example_json['label']) if 'label' in example_json else None
idx = example_json['idx']
guid = "%s-%s" % (set_type, idx)
text_a = example_json['premise']
meta = {
'choice1': example_json['choice1'],
'choice2': example_json['choice2'],
'question': example_json['question']
}
example = InputExample(guid=guid, text_a=text_a, label=label, meta=meta, idx=idx)
examples.append(example)
if set_type == 'train' or set_type == 'unlabeled':
mirror_examples = []
for ex in examples:
label = "1" if ex.label == "0" else "0"
meta = {
'choice1': ex.meta['choice2'],
'choice2': ex.meta['choice1'],
'question': ex.meta['question']
}
mirror_example = InputExample(guid=ex.guid + 'm', text_a=ex.text_a, label=label, meta=meta)
mirror_examples.append(mirror_example)
examples += mirror_examples
logger.info(f"Added {len(mirror_examples)} mirror examples, total size is {len(examples)}...")
return examples
def _create_examples(self,
path: str,
set_type: str,
hypothesis_name: str = "hypothesis",
premise_name: str = "premise") -> List[InputExample]:
examples = []
with open(path, encoding='utf8') as f:
for line_idx, line in enumerate(f):
example_json = json.loads(line)
idx = example_json['idx']
if isinstance(idx, str):
try:
idx = int(idx)
except ValueError:
idx = line_idx
label = example_json.get('label')
guid = "%s-%s" % (set_type, idx)
text_a = example_json[premise_name]
text_b = example_json[hypothesis_name]
example = InputExample(guid=guid,
text_a=text_a,
text_b=text_b,
label=label,
idx=idx)
examples.append(example)
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path, encoding='utf8') as f:
for line in f:
example_json = json.loads(line)
passage_idx = example_json['idx']
text = example_json['passage']['text']
questions = example_json['passage']['questions']
for question_json in questions:
question = question_json["question"]
question_idx = question_json['idx']
answers = question_json["answers"]
for answer_json in answers:
label = str(answer_json["label"]) if 'label' in answer_json else None
answer_idx = answer_json["idx"]
guid = f'{set_type}-p{passage_idx}-q{question_idx}-a{answer_idx}'
meta = {
'passage_idx': passage_idx,
'question_idx': question_idx,
'answer_idx': answer_idx,
'answer': answer_json["text"]
}
idx = [passage_idx, question_idx, answer_idx]
example = InputExample(guid=guid, text_a=text, text_b=question, label=label, meta=meta, idx=idx)
examples.append(example)
question_indices = list(set(example.meta['question_idx'] for example in examples))
label_distribution = Counter(example.label for example in examples)
logger.info(f"Returning {len(examples)} examples corresponding to {len(question_indices)} questions with label "
f"distribution {list(label_distribution.items())}")
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path, encoding='utf8') as f:
for line in f:
example_json = json.loads(line)
idx = example_json['idx']
label = str(example_json['label']) if 'label' in example_json else None
guid = "%s-%s" % (set_type, idx)
text_a = example_json['text']
meta = {
'span1_text': example_json['target']['span1_text'],
'span2_text': example_json['target']['span2_text'],
'span1_index': example_json['target']['span1_index'],
'span2_index': example_json['target']['span2_index']
}
# the indices in the dataset are wrong for some examples, so we manually fix them
span1_index, span1_text = meta['span1_index'], meta['span1_text']
span2_index, span2_text = meta['span2_index'], meta['span2_text']
words_a = text_a.split()
words_a_lower = text_a.lower().split()
words_span1_text = span1_text.lower().split()
span1_len = len(words_span1_text)
if words_a_lower[span1_index:span1_index + span1_len] != words_span1_text:
for offset in [-1, +1]:
if words_a_lower[span1_index + offset:span1_index + span1_len + offset] == words_span1_text:
span1_index += offset
if words_a_lower[span1_index:span1_index + span1_len] != words_span1_text:
logger.warning(f"Got '{words_a_lower[span1_index:span1_index + span1_len]}' but expected "
f"'{words_span1_text}' at index {span1_index} for '{words_a}'")
if words_a[span2_index] != span2_text:
for offset in [-1, +1]:
if words_a[span2_index + offset] == span2_text:
span2_index += offset
if words_a[span2_index] != span2_text and words_a[span2_index].startswith(span2_text):
words_a = words_a[:span2_index] \
+ [words_a[span2_index][:len(span2_text)], words_a[span2_index][len(span2_text):]] \
+ words_a[span2_index + 1:]
assert words_a[span2_index] == span2_text, \
f"Got '{words_a[span2_index]}' but expected '{span2_text}' at index {span2_index} for '{words_a}'"
text_a = ' '.join(words_a)
meta['span1_index'], meta['span2_index'] = span1_index, span2_index
example = InputExample(guid=guid, text_a=text_a, label=label, meta=meta, idx=idx)
if set_type == 'train' and label != 'True':
continue
examples.append(example)
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path, encoding="utf8") as f:
for line in f:
example_json = json.loads(line)
label = str(
example_json["label"]) if "label" in example_json else None
idx = example_json["idx"]
guid = "%s-%s" % (set_type, idx)
text_a = example_json["premise"]
meta = {
"choice1": example_json["choice1"],
"choice2": example_json["choice2"],
"question": example_json["question"],
}
example = InputExample(guid=guid,
text_a=text_a,
label=label,
meta=meta,
idx=idx)
examples.append(example)
if set_type == "train" or set_type == "unlabeled":
mirror_examples = []
for ex in examples:
label = "1" if ex.label == "0" else "0"
meta = {
"choice1": ex.meta["choice2"],
"choice2": ex.meta["choice1"],
"question": ex.meta["question"]
}
mirror_example = InputExample(guid=ex.guid + "m",
text_a=ex.text_a,
label=label,
meta=meta)
mirror_examples.append(mirror_example)
examples += mirror_examples
logger.info(
f"Added {len(mirror_examples)} mirror examples, total size is {len(examples)}..."
)
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path) as f:
reader = csv.reader(f, delimiter=',')
for idx, row in enumerate(reader):
label, body = row
guid = "%s-%s" % (set_type, idx)
text_a = body.replace('\\n', ' ').replace('\\', ' ')
example = InputExample(guid=guid, text_a=text_a, label=label)
examples.append(example)
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path, encoding='utf8') as f:
for line in f:
example_json = json.loads(line)
idx = example_json['idx']
label = str(example_json['label']) if 'label' in example_json else None
guid = "%s-%s" % (set_type, idx)
text_a = example_json['passage']
text_b = example_json['question']
example = InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label, idx=idx)
examples.append(example)
return examples
def _create_examples(lines: List[List[str]], set_type: str) -> List[InputExample]:
examples = []
for (i, line) in enumerate(lines):
if i == 0:
continue
guid = "%s-%s" % (set_type, line[0])
text_a = line[8]
text_b = line[9]
label = line[-1]
example = InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label)
examples.append(example)
return examples
def _create_examples(self, path, set_type, max_examples=-1, skip_first=0):
"""Creates examples for the training and dev sets."""
examples = []
with open(path) as f:
reader = csv.reader(f, delimiter=':->')
for idx, row in enumerate(reader):
guid = "%s-%s" % (set_type, idx)
label = row[MyTaskDataProcessor.LABEL_COLUMN]
text_a = row[MyTaskDataProcessor.TEXT_A_COLUMN]
text_b = row[MyTaskDataProcessor.TEXT_B_COLUMN] if MyTaskDataProcessor.TEXT_B_COLUMN >= 0 else None
example = InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label)
examples.append(example)
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path, encoding='utf8') as f:
reader = csv.reader(f, delimiter=',')
for idx, row in enumerate(reader):
label, question_title, question_body, answer = row
guid = "%s-%s" % (set_type, idx)
text_a = ' '.join([question_title.replace('\\n', ' ').replace('\\', ' '),
question_body.replace('\\n', ' ').replace('\\', ' ')])
text_b = answer.replace('\\n', ' ').replace('\\', ' ')
example = InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label)
examples.append(example)
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path, encoding='utf8') as f:
for line in f:
example_json = json.loads(line)
idx = example_json['idx']
if isinstance(idx, str):
idx = int(idx)
label = "T" if example_json.get('label') else "F"
guid = "%s-%s" % (set_type, idx)
text_a = example_json['sentence1']
text_b = example_json['sentence2']
meta = {'word': example_json['word']}
example = InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label, idx=idx, meta=meta)
examples.append(example)
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
df = pd.read_table(path)
for idx, row in df.iterrows():
label = str(row['prefix'])
guid = "%s-%s" % (set_type, idx)
text_a = str(row['input_text'])
text_b = str(row['target_text'])
example = InputExample(guid=guid,
text_a=text_a,
text_b=text_b,
label=label,
idx=idx)
examples.append(example)
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path) as f:
reader = csv.reader(f, delimiter=",")
for idx, row in enumerate(reader):
label, headline, body = row
guid = "%s-%s" % (set_type, idx)
text_a = headline.replace("\\", " ")
text_b = body.replace("\\", " ")
example = InputExample(guid=guid,
text_a=text_a,
text_b=text_b,
label=label)
examples.append(example)
return examples
def _create_examples(self, path: str) -> List[InputExample]:
examples = []
with open(path, encoding='utf8') as f:
for line in f:
example_json = json.loads(line)
label = example_json['label']
id_ = example_json['id']
text_a = example_json['question']
text_b = example_json['comment']
language = example_json['language']
if self.language is not None and language != self.language:
continue
example = InputExample(guid=id_, text_a=text_a, text_b=text_b, label=label)
examples.append(example)
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path, encoding="utf8") as f:
for line in f:
example_json = json.loads(line)
idx = example_json["idx"]
label = str(
example_json["label"]) if "label" in example_json else None
guid = "%s-%s" % (set_type, idx)
text_a = example_json["passage"]
text_b = example_json["question"]
example = InputExample(guid=guid,
text_a=text_a,
text_b=text_b,
label=label,
idx=idx)
examples.append(example)
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path, encoding="utf8") as f:
for line in f:
example_json = json.loads(line)
idx = example_json["idx"]
if isinstance(idx, str):
idx = int(idx)
label = "T" if example_json.get("label") else "F"
guid = "%s-%s" % (set_type, idx)
text_a = example_json["sentence1"]
text_b = example_json["sentence2"]
meta = {"word": example_json["word"]}
example = InputExample(guid=guid,
text_a=text_a,
text_b=text_b,
label=label,
idx=idx,
meta=meta)
examples.append(example)
return examples
def _create_examples_unlabelled(self,
path,
set_type,
max_examples=1,
skip_first=0):
"""Creates examples for the unlabelled set."""
examples = []
with open(path, encoding="utf8") as f:
reader = csv.reader(f, delimiter=":")
for idx, row in enumerate(reader):
guid = "%s-%s" % (set_type, idx)
#label = row[MyTaskDataProcessor.LABEL_COLUMN]
text_a = row[0]
text_b = row[
MyTaskDataProcessor.
TEXT_B_COLUMN] if MyTaskDataProcessor.TEXT_B_COLUMN >= 0 else None
example = InputExample(guid=guid, text_a=text_a, text_b=text_b)
examples.append(example)
return examples
def _create_examples(self, lines: List[List[str]],
set_type: str) -> List[InputExample]:
examples = []
id_to_lables = self.get_labels()
for (i, line) in enumerate(lines):
if i == 0:
continue
guid = f"{set_type}-{line['idx']}"
text_a = line['premise']
text_b = line['hypothesis']
label = id_to_lables[
line['label']] # need to return string, hf datasets uses int
example = InputExample(guid=guid,
text_a=text_a,
text_b=text_b,
label=label)
examples.append(example)
return examples
def _create_examples(
path,
set_type,
seed=42,
max_train_candidates_per_question: int = 10) -> List[InputExample]:
examples = []
entity_shuffler = random.Random(seed)
with open(path, encoding='utf8') as f:
for idx, line in enumerate(f):
example_json = json.loads(line)
idx = example_json['idx']
text = example_json['passage']['text']
entities = set()
for entity_json in example_json['passage']['entities']:
start = entity_json['start']
end = entity_json['end']
entity = text[start:end + 1]
entities.add(entity)
entities = list(entities)
text = text.replace(
"@highlight\n", "- "
) # we follow the GPT-3 paper wrt @highlight annotations
questions = example_json['qas']
for question_json in questions:
question = question_json['query']
question_idx = question_json['idx']
answers = set()
for answer_json in question_json.get('answers', []):
answer = answer_json['text']
answers.add(answer)
answers = list(answers)
if set_type == 'train':
# create a single example per *correct* answer
for answer_idx, answer in enumerate(answers):
candidates = [
ent for ent in entities if ent not in answers
]
if len(candidates
) > max_train_candidates_per_question - 1:
entity_shuffler.shuffle(candidates)
candidates = candidates[:
max_train_candidates_per_question
- 1]
guid = f'{set_type}-p{idx}-q{question_idx}-a{answer_idx}'
meta = {
'passage_idx': idx,
'question_idx': question_idx,
'candidates': [answer] + candidates,
'answers': [answer]
}
ex_idx = [idx, question_idx, answer_idx]
example = InputExample(guid=guid,
text_a=text,
text_b=question,
label="1",
meta=meta,
idx=ex_idx)
examples.append(example)
else:
# create just one example with *all* correct answers and *all* answer candidates
guid = f'{set_type}-p{idx}-q{question_idx}'
meta = {
'passage_idx': idx,
'question_idx': question_idx,
'candidates': entities,
'answers': answers
}
example = InputExample(guid=guid,
text_a=text,
text_b=question,
label="1",
meta=meta)
examples.append(example)
question_indices = list(
set(example.meta['question_idx'] for example in examples))
label_distribution = Counter(example.label for example in examples)
logger.info(
f"Returning {len(examples)} examples corresponding to {len(question_indices)} questions with label "
f"distribution {list(label_distribution.items())}")
return examples
def _create_examples(path: str, set_type: str) -> List[InputExample]:
examples = []
with open(path, encoding="utf8") as f:
for line in f:
example_json = json.loads(line)
idx = example_json["idx"]
label = str(
example_json["label"]) if "label" in example_json else None
guid = "%s-%s" % (set_type, idx)
text_a = example_json["text"]
meta = {
"span1_text": example_json["target"]["span1_text"],
"span2_text": example_json["target"]["span2_text"],
"span1_index": example_json["target"]["span1_index"],
"span2_index": example_json["target"]["span2_index"],
}
# the indices in the dataset are wrong for some examples, so we manually fix them
span1_index, span1_text = meta["span1_index"], meta[
"span1_text"]
span2_index, span2_text = meta["span2_index"], meta[
"span2_text"]
words_a = text_a.split()
words_a_lower = text_a.lower().split()
words_span1_text = span1_text.lower().split()
span1_len = len(words_span1_text)
if words_a_lower[span1_index:span1_index +
span1_len] != words_span1_text:
for offset in [-1, +1]:
if words_a_lower[span1_index + offset:span1_index +
span1_len +
offset] == words_span1_text:
span1_index += offset
if words_a_lower[span1_index:span1_index +
span1_len] != words_span1_text:
logger.warning(
f"Got '{words_a_lower[span1_index:span1_index + span1_len]}' but expected "
f"'{words_span1_text}' at index {span1_index} for '{words_a}'"
)
if words_a[span2_index] != span2_text:
for offset in [-1, +1]:
if words_a[span2_index + offset] == span2_text:
span2_index += offset
if words_a[span2_index] != span2_text and words_a[
span2_index].startswith(span2_text):
words_a = (words_a[:span2_index] + [
words_a[span2_index][:len(span2_text)],
words_a[span2_index][len(span2_text):]
] + words_a[span2_index + 1:])
assert (
words_a[span2_index] == span2_text
), f"Got '{words_a[span2_index]}' but expected '{span2_text}' at index {span2_index} for '{words_a}'"
text_a = " ".join(words_a)
meta["span1_index"], meta[
"span2_index"] = span1_index, span2_index
example = InputExample(guid=guid,
text_a=text_a,
label=label,
meta=meta,
idx=idx)
if set_type == "train" and label != "True":
continue
examples.append(example)
return examples