def to_scored_spans(self, probs, score_mask):
probs = probs.data.cpu()
score_mask = score_mask.data.cpu()
batch_size, num_spans = probs.size()
spans = []
for b in range(batch_size):
batch_spans = []
for start, end, i in self.start_end_range(num_spans):
if score_mask[b, i] == 1 and probs[b, i] > 0:
batch_spans.append((Span(start, end), probs[b, i]))
spans.append(batch_spans)
return spans
def _read(self, file_list: str):
instances = []
for file_path in file_list.split(","):
if file_path.strip() == "":
continue
logger.info("Reading QASRL instances from dataset file at: %s",
file_path)
data = []
if file_path.endswith('.gz'):
with gzip.open(cached_path(file_path), 'r') as f:
for line in f:
data.append(json.loads(line))
elif file_path.endswith(".jsonl"):
with codecs.open(cached_path(file_path), 'r',
encoding='utf8') as f:
for line in f:
data.append(json.loads(line))
for item in data:
sent_id = item["sentenceId"]
sentence_tokens = item["sentenceTokens"]
annotations = []
for _, verb_entry in item["verbEntries"].items():
verb_index = verb_entry["verbIndex"]
self._num_verbs += 1
annotations = []
for _, question_label in verb_entry[
"questionLabels"].items():
answers = len(question_label["answerJudgments"])
valid_answers = len([
ans for ans in question_label["answerJudgments"]
if ans["isValid"]
])
if self._question_sources is not None:
if not any([
source.startswith(prefix) for source in
question_label["questionSources"]
for prefix in self._question_sources
]):
continue
if answers < self._min_answers:
self._not_enough_answers += 1
continue
if valid_answers < self._min_valid_answers:
self._not_enough_valid_answers += 1
continue
slots = []
for l in self._slot_labels:
slots.append(question_label["questionSlots"][l])
provinence = list(question_label["questionSources"])[0]
spans = []
for ans in question_label["answerJudgments"]:
if ans["isValid"]:
for s in ans["spans"]:
spans.append(Span(s[0], s[1] - 1))
self._qa_pairs += 1
annotations.append(
AnnotatedSpan(slots=slots,
all_spans=spans,
provinence=provinence))
if annotations:
self._instances += 1
yield self._make_instance_from_text(
sentence_tokens,
verb_index,
annotations=annotations,
sent_id=sent_id)
else:
self._no_ann += 1
logger.info("Produced %d instances" % self._instances)
logger.info("\t%d Verbs" % self._num_verbs)
logger.info("\t%d QA pairs" % self._qa_pairs)
logger.info("\t%d no annotation" % self._no_ann)
logger.info("\t%d not enough answers" % self._not_enough_answers)
logger.info("\t%d not enough valid answers" %
self._not_enough_valid_answers)