Method _create_examples

tasks/superglue/dataset.py:432–517

(self, path: str, set_type: str, cloze_eval=True)

        return text_a, text_b

    def _create_examples(self, path: str, set_type: str, cloze_eval=True) -> 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 = punctuation_standardization(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']
                }
                if 'candidates' in example_json:
                    candidates = [cand['text'] for cand in example_json['candidates']]
                    # candidates = list(set(candidates))
                    filtered = []
                    for i, cand in enumerate(candidates):
                        if not cand in candidates[:i]:
                            filtered.append(cand)
                    candidates = filtered

                # 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:
                #     print_rank_0(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