harim_plus.py

import datasets
import evaluate

from harim_scorer import Harimplus_Scorer



logger = evaluate.logging.get_logger(__name__)

CODEBASE_URL='https://huggingface.co/spaces/NCSOFT/harim_plus'
PAPER_URL='https://arxiv.org/abs/2211.12118'

_CITATION = """\
@inproceedings{son-etal-2022-harim,
    title = "{H}a{R}i{M}$^+$: Evaluating Summary Quality with Hallucination Risk",
    author = "Son, Seonil (Simon)  and
      Park, Junsoo  and
      Hwang, Jeong-in  and
      Lee, Junghwa  and
      Noh, Hyungjong  and
      Lee, Yeonsoo",
    booktitle = "Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing",
    month = nov,
    year = "2022",
    address = "Online only",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.aacl-main.66",
    pages = "895--924",
    abstract = "One of the challenges of developing a summarization model arises from the difficulty in measuring the factual inconsistency of the generated text. In this study, we reinterpret the decoder overconfidence-regularizing objective suggested in (Miao et al., 2021) as a hallucination risk measurement to better estimate the quality of generated summaries. We propose a reference-free metric, HaRiM+, which only requires an off-the-shelf summarization model to compute the hallucination risk based on token likelihoods. Deploying it requires no additional training of models or ad-hoc modules, which usually need alignment to human judgments. For summary-quality estimation, HaRiM+ records state-of-the-art correlation to human judgment on three summary-quality annotation sets: FRANK, QAGS, and SummEval. We hope that our work, which merits the use of summarization models, facilitates the progress of both automated evaluation and generation of summary.",
}
"""

_DESCRIPTION = """\
HaRiM+ is a reference-less (i.e. scoring summary quality only requires an article) evaluation metric score for summarization task which hurls the power of summarization model.
It will work great ranking the summary-article pairs according to its quality.
Note that the score range is unbound.

Summarization model inside the HaRiM+ will read and evaluate how good the quality of a summary given the paired source article.

HaRiM+ is proved effective for benchmarking summarization systems (system-level performance) as well as ranking the article-summary pairs (segment-level performance) in comprehensive aspect such as factuality, consistency, coherency, fluency, and relevance. For details, refer to our paper published in AACL2022.
"""

_KWARGS_DESCRIPTION = """
HaRiM+ score.
Args:
    For scorer = evaluate.load():
    `pretrained_name` (str or pathlib.Path): summarization model checkpoint or path, loaded by transformers.AutoModelForSeq2SeqLM.from_pretrained(). Defaults to Yale-LILY/brio-cnndm-uncased.
    `tokenizer`: (use when your tokenizer cannot be loaded by from_pretrained)Tokenizer function compatible with transformers.PreTrainedTokenizer. It requires tokenizer.pad_token|eos_token|bos_token and tokenizer.__call__() method for HaRiM+ score computation.

    For scorer.compute():
    `predictions` (list of str): generated summaries
    `references` (list of str): source articles to be summarized
    `use_aggregator` (bool): if True, average of the scores are returned

Returns:
    'results' (dict): {
        'harim+' (List[float] or float): HaRiM+ score to use,
        'harim' (List[float] or float): HaRiM term for computing the score above,
        'log_ppl' (List[float] or float): Log perplexity term. Same as (Yuan et al., NeurIPS 2021),
        'lambda' (float): (recommend not to modify this) Balancing coeff. for computing harim+ from harim and log_ppl.
    }

Examples:
    >>> summaries = ["hello there", "hello there"]
    >>> articles = ["hello, this is the article to be summarized", "hello, this is the article to be summarized"]
    >>> scorer = evaluate.load("NCSOFT/harim_plus") #, pretrained_name='PRETRAINEDNAME', tokenizer=TOKENIZER # optional
    >>> results = scorer.compute(predictions=summaries, references=articles) # use_aggregator=True # optional
    >>> print([round(v, 2) for v in results["harim+"]])
    [float, float]
"""



@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class Harimplus(evaluate.Metric):
    def __init__(self,
                    pretrained_name='facebook/bart-large-cnn',
                    tokenizer=None,
                    device='cuda',
                    **kwargs
                    ):
        super().__init__(**kwargs)
        self.myconfig = dict(
                            pretrained_name=pretrained_name,
                            tokenizer=tokenizer,
                            device=device,
                            )

    def _info(self):
        return evaluate.MetricInfo(
            description=_DESCRIPTION,
            citation=_CITATION,
            homepage=CODEBASE_URL,
            inputs_description=_KWARGS_DESCRIPTION,
            features=datasets.Features(
                {
                    "predictions": datasets.Value("string", id="sequence"),
                    "references": datasets.Value("string", id="sequence"),
                }
            ),
            codebase_urls=[CODEBASE_URL],
            reference_urls=[CODEBASE_URL, PAPER_URL],
        )

    def _download_and_prepare(self, dl_manager):
        pretrained_name = self.myconfig['pretrained_name']
        is_custom_tokenizer = self.myconfig['tokenizer'] is not None
        logger.warning(
            "Loading HaRiM+ score"
            f"\tpretrained_name = {pretrained_name}"
        )
        if is_custom_tokenizer:
            logger.warning(
                f"tokenizer is overriden by \n\tself.myconfig['tokenizer']"
            )
        logger.warning(
            "You can change checkpoints with `pretrained_name` kwarg in evaluate.load. Strongly recommend to use *-large or larger ones."
            "Refrain from using checkpoints trained on noisy corpus such as bbc-XSUM.")

        # download the model checkpoint specified by self.myconfig_name and set up the scorer
        self.scorer = score.Harimplus_Scorer(**self.myconfig)

    def _compute(self, predictions=None,
                        references=None,
                        use_aggregator=False,
                        bsz=32,
                        tokenwise_score=False):
        summaries = predictions
        articles = references
        scores = self.scorer.compute(predictions=summaries, references=articles,  use_aggregator=use_aggregator, bsz=bsz, tokenwise_score=tokenwise_score)
        return scores