DiaBLa.py

# coding=utf-8
'''DiaBLA: Dialogue Bilingue datset'''

import json
import datasets
from datasets.features import ClassLabel


logger = datasets.logging.get_logger(__name__)

_CITATION = '''\
@article{bawden_DiaBLa:-A-Corpus-of_2021,
  author = {Bawden, Rachel and Bilinski, Eric and Lavergne, Thomas and Rosset, Sophie},
  doi = {10.1007/s10579-020-09514-4},
  title = {DiaBLa: A Corpus of Bilingual Spontaneous Written Dialogues for Machine Translation},
  year = {2021},
  journal = {Language Resources and Evaluation},
  publisher = {Springer Verlag},
  volume = {55},
  pages = {635--660},
  url = {https://hal.inria.fr/hal-03021633},
  pdf = {https://hal.inria.fr/hal-03021633/file/diabla-lre-personal-formatting.pdf},
}
'''

_DESCRIPTION = '''\
English-French parallel dataset for the evaluation of \
Machine Translation (MT) for informal, written bilingual dialogue.
'''
 
_URLS = {
    'test': 'DiaBLa.json',
}


class DiablaConfig(datasets.BuilderConfig):
    '''BuilderConfig for DiaBLa.'''

    def __init__(self, **kwargs):
        """BuilderConfig for DiaBLa.

        Args:
          **kwargs: keyword arguments forwarded to super.
        """
        super(DiablaConfig, self).__init__(**kwargs)


class Diabla(datasets.GeneratorBasedBuilder):
    '''DiaBLa: English-French parallel dataset of bilingual dialogue'''

    BUILDER_CONFIGS = [
        DiablaConfig(
            name='plain_text',
            version=datasets.Version('1.0.0', ''),
            description='Plain text',
        ),
    ]

    #TODO
    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    'id': datasets.Value('string'),
                    'orig': datasets.Value('string'),
                    'norm': datasets.Value('string'),
                    'mt': datasets.Value('string'),
                    'ref': datasets.Value('string'),
                    'utterance_meta': datasets.features.Sequence(
                        {
                            'eval-judgment': ClassLabel(num_classes=3, names=['poor', 'medium', 'perfect']),
                            'eval-verbatim': datasets.Value('string'),
                            'eval-problems': datasets.features.Sequence(
                                [
                                  ClassLabel(num_classes=6, names=['coherence', 'grammar', 'meaning', 'word choice', 'style', 'other'])
                                ]
                              ),
                             'lang': ClassLabel(num_classes=2, names=['english', 'french']),
                        }
                    ),
                }
            ),
            # TODO?
            supervised_keys=None,
            homepage='https://github.com/rbawden/DiaBLa-dataset',
            citation=_CITATION,
            task_templates=[
                # TODO
            ],
        )

    def _split_generators(self, dl_manager):
        downloaded_files = dl_manager.download_and_extract(_URLS)

        return [datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={'filepath': downloaded_files['test']})]

    def _generate_examples(self, filepath):
        '''This function returns the examples in the raw (text) form.'''
        logger.info("generating examples from = %s", filepath)
        key = 0
        with open(filepath, encoding="utf-8") as f:
            diabla = json.load(f)
            for dialogue_name in sorted(diabla['dialogues']):
                dialogue_history = [] # to store past utterances
                dialogue = diabla['dialogues'][dialogue_name]
                # Meta-information attached to the dialogue
                dialogue_info_keys = ['start_time', 'end_time', 'scenario',
                                      'user1', 'user2', 'translation_model',
                                      'final_evaluation_user1', 'final_evaluation_user2']
                for user in 'user1', 'user2':
                    for info_to_remove in ['eval-stage', 'useragent']:
                        if info_to_remove in dialogue[user]:
                            del dialogue[user][info_to_remove]
                dialogue_info = {k: dialogue[k] for k in dialogue_info_keys}
                if dialogue_info['end_time'] is None:
                    dialogue_info['end_time'] = ''
                for info_to_remove in ['interface','verbatim_quality', 
                                       'particular_problems', 'tech', 
                                       'would_use', 'timestamp', 'technical_issue']:
                    for final_eval in 'final_evaluation_user1', 'final_evaluation_user2':
                        if info_to_remove in dialogue_info[final_eval]:
                            del dialogue_info[final_eval][info_to_remove]
                    
                # Main data: the utterances
                for utterance_id in dialogue['utterances']:
                    utterance = dialogue['utterances'][utterance_id]
                    # Meta-information attached to the utterance
                    utterance_info_keys = ['judgment', 'verbatim', 'problems']
                    utterance_info = {'eval-' + k: utterance['eval'][k] for k in utterance_info_keys}
                    if utterance_info['eval-judgment'] is None:
                        utterance_info['eval-judgment'] = ''
                    utterance_info['lang'] = utterance['language']
                    # Utterance text
                    original_text = utterance['original_text']
                    mt_text = utterance['postprocessed_text']
                    reference_text = utterance['reference_translation']
                    normalised_text = utterance['normalised_version']
                    id_ = dialogue_name + '_' + utterance_id
                    utterance_instance = {
                        'orig': original_text,
                        'norm': normalised_text,
                        'mt': mt_text,
                        'id': id_,
                        'ref': reference_text,
                        'utterance_meta': utterance_info
                    }
                    
                    # add to history (without dialogue info and history)
                    dialogue_history.append(utterance_instance.copy()) 
                    #utterance_instance['dialogue_meta'] = dialogue_info
                    #utterance_instance['dialogue_history'] = dialogue_history
                    yield id_, utterance_instance