ibleu.py

"""iBleu metric."""

import datasets
import sacrebleu as scb
from packaging import version

import evaluate


_CITATION = """\
@inproceedings{sun-zhou-2012-joint,
    title = "Joint Learning of a Dual {SMT} System for Paraphrase Generation",
    author = "Sun, Hong  and
      Zhou, Ming",
    booktitle = "Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)",
    month = jul,
    year = "2012",
    address = "Jeju Island, Korea",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/P12-2008",
    pages = "38--42",
}

"""

_DESCRIPTION = """\
iBLEU measures the adequacy and dissimilarity of generated paraphrases.
"""

_KWARGS_DESCRIPTION = """
Produces iBLEU score from an input and a prediction against one or more references.
Args:
    inputs (`list` of `str`): list of model inputs. Each input should be tokenized into a list of tokens.
    predictions (`list` of `str`): list of translations to score. Each translation should be tokenized into a list of tokens.
    references (`list` of `list` of `str`): A list of lists of references. The contents of the first sub-list are the references for the first prediction, the contents of the second sub-list are for the second prediction, etc. Note that there must be the same number of references for each prediction (i.e. all sub-lists must be of the same length).
    alpha (`float`): parameter for balancing between adequacy and dissimilarity; smaller α value indicates larger punishment on self-paraphrase.
    smooth_method (`str`): The smoothing method to use, defaults to `'exp'`. Possible values are:
        - `'none'`: no smoothing
        - `'floor'`: increment zero counts
        - `'add-k'`: increment num/denom by k for n>1
        - `'exp'`: exponential decay
    smooth_value (`float`): The smoothing value. Only valid when `smooth_method='floor'` (in which case `smooth_value` defaults to `0.1`) or `smooth_method='add-k'` (in which case `smooth_value` defaults to `1`).
    tokenize (`str`): Tokenization method to use for iBLEU. If not provided, defaults to `'zh'` for Chinese, `'ja-mecab'` for Japanese and `'13a'` (mteval) otherwise. Possible values are:
        - `'none'`: No tokenization.
        - `'zh'`: Chinese tokenization.
        - `'13a'`: mimics the `mteval-v13a` script from Moses.
        - `'intl'`: International tokenization, mimics the `mteval-v14` script from Moses
        - `'char'`: Language-agnostic character-level tokenization.
        - `'ja-mecab'`: Japanese tokenization. Uses the [MeCab tokenizer](https://pypi.org/project/mecab-python3).
    lowercase (`bool`): If `True`, lowercases the input, enabling case-insensitivity. Defaults to `False`.
    force (`bool`): If `True`, insists that your tokenized input is actually detokenized. Defaults to `False`.
    use_effective_order (`bool`): If `True`, stops including n-gram orders for which precision is 0. This should be `True`, if sentence-level BLEU will be computed. Defaults to `False`.
Returns:
    'score': iBLEU score,
Example:
    >>> inputs = ["greetings general kenobi", "foo  foo bar bar"]
    >>> predictions = ["hello there general kenobi", "foo bar foobar"]
    >>> references = [["hello there general kenobi", "hello there !"], ["foo bar foobar", "foo bar foobar"]]
    >>> ibleu = evaluate.load("rahular/ibleu")
    >>> results = ibleu.compute(inputs=inputs, predictions=predictions, references=references)
    >>> print(results)
    {'score': 60.41585343630594}
"""


@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class ibleu(evaluate.Metric):
    def _info(self):
        if version.parse(scb.__version__) < version.parse("1.4.12"):
            raise ImportWarning(
                "To use `sacrebleu`, the module `sacrebleu>=1.4.12` is required, and the current version of `sacrebleu` doesn't match this condition.\n"
                'You can install it with `pip install "sacrebleu>=1.4.12"`.'
            )
        return evaluate.MetricInfo(
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            features=[
                datasets.Features(
                    {
                        "inputs": datasets.Value("string", id="sequence"),
                        "predictions": datasets.Value("string", id="sequence"),
                        "references": datasets.Sequence(
                            datasets.Value("string", id="sequence"), id="references"
                        ),
                    }
                ),
                datasets.Features(
                    {
                        "inputs": datasets.Value("string", id="sequence"),
                        "predictions": datasets.Value("string", id="sequence"),
                        "references": datasets.Value("string", id="sequence"),
                    }
                ),
            ],
            reference_urls=[
                "https://scikit-learn.org/stable/modules/generated/sklearn.metrics.accuracy_score.html"
            ],
        )

    def _compute(
        self,
        inputs,
        predictions,
        references,
        alpha=0.7,
        smooth_method="exp",
        smooth_value=None,
        force=False,
        lowercase=False,
        tokenize=None,
        use_effective_order=False,
    ):
        # if only one reference is provided make sure we still use list of lists
        if isinstance(references[0], str):
            references = [[ref] for ref in references]
        # we need to do the same for inputs
        if isinstance(inputs[0], str):
            inputs = [[inp] for inp in inputs]
        else:
            raise ValueError("There can be only one input string")
        
        references_per_prediction = len(references[0])
        if any(len(refs) != references_per_prediction for refs in references):
            raise ValueError("Sacrebleu requires the same number of references for each prediction")
        transformed_references = [[refs[i] for refs in references] for i in range(references_per_prediction)]

        tgt_bleu = scb.corpus_bleu(
            predictions,
            transformed_references,
            smooth_method=smooth_method,
            smooth_value=smooth_value,
            force=force,
            lowercase=lowercase,
            use_effective_order=use_effective_order,
            **(dict(tokenize=tokenize) if tokenize else {}),
        ).score
        self_bleu = scb.corpus_bleu(
            predictions,
            inputs,
            smooth_method=smooth_method,
            smooth_value=smooth_value,
            force=force,
            lowercase=lowercase,
            use_effective_order=use_effective_order,
            **(dict(tokenize=tokenize) if tokenize else {}),
        ).score
        output_dict = {
            "score": alpha * tgt_bleu - (1 - alpha) * self_bleu
        }
        return output_dict