wiki_split.py

# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" WIKI_SPLIT metric."""

import re
import string
from collections import Counter

import datasets
import sacrebleu
import sacremoses
from packaging import version

import evaluate


_CITATION = """
@inproceedings{xu-etal-2016-optimizing,
    title = {Optimizing Statistical Machine Translation for Text Simplification},
    authors={Xu, Wei and Napoles, Courtney and Pavlick, Ellie and Chen, Quanze and Callison-Burch, Chris},
    journal = {Transactions of the Association for Computational Linguistics},
    volume = {4},
    year={2016},
    url = {https://www.aclweb.org/anthology/Q16-1029},
    pages = {401--415
},
@inproceedings{post-2018-call,
    title = "A Call for Clarity in Reporting {BLEU} Scores",
    author = "Post, Matt",
    booktitle = "Proceedings of the Third Conference on Machine Translation: Research Papers",
    month = oct,
    year = "2018",
    address = "Belgium, Brussels",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/W18-6319",
    pages = "186--191",
}
"""

_DESCRIPTION = """\
WIKI_SPLIT is the combination of three metrics SARI, EXACT and SACREBLEU
It can be used to evaluate the quality of machine-generated texts.
"""


_KWARGS_DESCRIPTION = """
Calculates sari score (between 0 and 100) given a list of source and predicted
sentences, and a list of lists of reference sentences. It also computes the BLEU score as well as the exact match score.
Args:
    sources: list of source sentences where each sentence should be a string.
    predictions: list of predicted sentences where each sentence should be a string.
    references: list of lists of reference sentences where each sentence should be a string.
Returns:
    sari: sari score
    sacrebleu: sacrebleu score
    exact: exact score

Examples:
    >>> sources=["About 95 species are currently accepted ."]
    >>> predictions=["About 95 you now get in ."]
    >>> references=[["About 95 species are currently known ."]]
    >>> wiki_split = evaluate.load("wiki_split")
    >>> results = wiki_split.compute(sources=sources, predictions=predictions, references=references)
    >>> print(results)
    {'sari': 21.805555555555557, 'sacrebleu': 14.535768424205482, 'exact': 0.0}
"""


def normalize_answer(s):
    """Lower text and remove punctuation, articles and extra whitespace."""

    def remove_articles(text):
        regex = re.compile(r"\b(a|an|the)\b", re.UNICODE)
        return re.sub(regex, " ", text)

    def white_space_fix(text):
        return " ".join(text.split())

    def remove_punc(text):
        exclude = set(string.punctuation)
        return "".join(ch for ch in text if ch not in exclude)

    def lower(text):
        return text.lower()

    return white_space_fix(remove_articles(remove_punc(lower(s))))


def compute_exact(a_gold, a_pred):
    return int(normalize_answer(a_gold) == normalize_answer(a_pred))


def compute_em(predictions, references):
    scores = [any([compute_exact(ref, pred) for ref in refs]) for pred, refs in zip(predictions, references)]
    return (sum(scores) / len(scores)) * 100


def SARIngram(sgrams, cgrams, rgramslist, numref):
    rgramsall = [rgram for rgrams in rgramslist for rgram in rgrams]
    rgramcounter = Counter(rgramsall)

    sgramcounter = Counter(sgrams)
    sgramcounter_rep = Counter()
    for sgram, scount in sgramcounter.items():
        sgramcounter_rep[sgram] = scount * numref

    cgramcounter = Counter(cgrams)
    cgramcounter_rep = Counter()
    for cgram, ccount in cgramcounter.items():
        cgramcounter_rep[cgram] = ccount * numref

    # KEEP
    keepgramcounter_rep = sgramcounter_rep & cgramcounter_rep
    keepgramcountergood_rep = keepgramcounter_rep & rgramcounter
    keepgramcounterall_rep = sgramcounter_rep & rgramcounter

    keeptmpscore1 = 0
    keeptmpscore2 = 0
    for keepgram in keepgramcountergood_rep:
        keeptmpscore1 += keepgramcountergood_rep[keepgram] / keepgramcounter_rep[keepgram]
        # Fix an alleged bug [2] in the keep score computation.
        # keeptmpscore2 += keepgramcountergood_rep[keepgram] / keepgramcounterall_rep[keepgram]
        keeptmpscore2 += keepgramcountergood_rep[keepgram]
    # Define 0/0=1 instead of 0 to give higher scores for predictions that match
    #      a target exactly.
    keepscore_precision = 1
    keepscore_recall = 1
    if len(keepgramcounter_rep) > 0:
        keepscore_precision = keeptmpscore1 / len(keepgramcounter_rep)
    if len(keepgramcounterall_rep) > 0:
        # Fix an alleged bug [2] in the keep score computation.
        # keepscore_recall = keeptmpscore2 / len(keepgramcounterall_rep)
        keepscore_recall = keeptmpscore2 / sum(keepgramcounterall_rep.values())
    keepscore = 0
    if keepscore_precision > 0 or keepscore_recall > 0:
        keepscore = 2 * keepscore_precision * keepscore_recall / (keepscore_precision + keepscore_recall)

    # DELETION
    delgramcounter_rep = sgramcounter_rep - cgramcounter_rep
    delgramcountergood_rep = delgramcounter_rep - rgramcounter
    delgramcounterall_rep = sgramcounter_rep - rgramcounter
    deltmpscore1 = 0
    deltmpscore2 = 0
    for delgram in delgramcountergood_rep:
        deltmpscore1 += delgramcountergood_rep[delgram] / delgramcounter_rep[delgram]
        deltmpscore2 += delgramcountergood_rep[delgram] / delgramcounterall_rep[delgram]
    # Define 0/0=1 instead of 0 to give higher scores for predictions that match
    # a target exactly.
    delscore_precision = 1
    if len(delgramcounter_rep) > 0:
        delscore_precision = deltmpscore1 / len(delgramcounter_rep)

    # ADDITION
    addgramcounter = set(cgramcounter) - set(sgramcounter)
    addgramcountergood = set(addgramcounter) & set(rgramcounter)
    addgramcounterall = set(rgramcounter) - set(sgramcounter)

    addtmpscore = 0
    for addgram in addgramcountergood:
        addtmpscore += 1

    # Define 0/0=1 instead of 0 to give higher scores for predictions that match
    # a target exactly.
    addscore_precision = 1
    addscore_recall = 1
    if len(addgramcounter) > 0:
        addscore_precision = addtmpscore / len(addgramcounter)
    if len(addgramcounterall) > 0:
        addscore_recall = addtmpscore / len(addgramcounterall)
    addscore = 0
    if addscore_precision > 0 or addscore_recall > 0:
        addscore = 2 * addscore_precision * addscore_recall / (addscore_precision + addscore_recall)

    return (keepscore, delscore_precision, addscore)


def SARIsent(ssent, csent, rsents):
    numref = len(rsents)

    s1grams = ssent.split(" ")
    c1grams = csent.split(" ")
    s2grams = []
    c2grams = []
    s3grams = []
    c3grams = []
    s4grams = []
    c4grams = []

    r1gramslist = []
    r2gramslist = []
    r3gramslist = []
    r4gramslist = []
    for rsent in rsents:
        r1grams = rsent.split(" ")
        r2grams = []
        r3grams = []
        r4grams = []
        r1gramslist.append(r1grams)
        for i in range(0, len(r1grams) - 1):
            if i < len(r1grams) - 1:
                r2gram = r1grams[i] + " " + r1grams[i + 1]
                r2grams.append(r2gram)
            if i < len(r1grams) - 2:
                r3gram = r1grams[i] + " " + r1grams[i + 1] + " " + r1grams[i + 2]
                r3grams.append(r3gram)
            if i < len(r1grams) - 3:
                r4gram = r1grams[i] + " " + r1grams[i + 1] + " " + r1grams[i + 2] + " " + r1grams[i + 3]
                r4grams.append(r4gram)
        r2gramslist.append(r2grams)
        r3gramslist.append(r3grams)
        r4gramslist.append(r4grams)

    for i in range(0, len(s1grams) - 1):
        if i < len(s1grams) - 1:
            s2gram = s1grams[i] + " " + s1grams[i + 1]
            s2grams.append(s2gram)
        if i < len(s1grams) - 2:
            s3gram = s1grams[i] + " " + s1grams[i + 1] + " " + s1grams[i + 2]
            s3grams.append(s3gram)
        if i < len(s1grams) - 3:
            s4gram = s1grams[i] + " " + s1grams[i + 1] + " " + s1grams[i + 2] + " " + s1grams[i + 3]
            s4grams.append(s4gram)

    for i in range(0, len(c1grams) - 1):
        if i < len(c1grams) - 1:
            c2gram = c1grams[i] + " " + c1grams[i + 1]
            c2grams.append(c2gram)
        if i < len(c1grams) - 2:
            c3gram = c1grams[i] + " " + c1grams[i + 1] + " " + c1grams[i + 2]
            c3grams.append(c3gram)
        if i < len(c1grams) - 3:
            c4gram = c1grams[i] + " " + c1grams[i + 1] + " " + c1grams[i + 2] + " " + c1grams[i + 3]
            c4grams.append(c4gram)

    (keep1score, del1score, add1score) = SARIngram(s1grams, c1grams, r1gramslist, numref)
    (keep2score, del2score, add2score) = SARIngram(s2grams, c2grams, r2gramslist, numref)
    (keep3score, del3score, add3score) = SARIngram(s3grams, c3grams, r3gramslist, numref)
    (keep4score, del4score, add4score) = SARIngram(s4grams, c4grams, r4gramslist, numref)
    avgkeepscore = sum([keep1score, keep2score, keep3score, keep4score]) / 4
    avgdelscore = sum([del1score, del2score, del3score, del4score]) / 4
    avgaddscore = sum([add1score, add2score, add3score, add4score]) / 4
    finalscore = (avgkeepscore + avgdelscore + avgaddscore) / 3
    return finalscore


def normalize(sentence, lowercase: bool = True, tokenizer: str = "13a", return_str: bool = True):

    # Normalization is requried for the ASSET dataset (one of the primary
    # datasets in sentence simplification) to allow using space
    # to split the sentence. Even though Wiki-Auto and TURK datasets,
    # do not require normalization, we do it for consistency.
    # Code adapted from the EASSE library [1] written by the authors of the ASSET dataset.
    # [1] https://github.com/feralvam/easse/blob/580bba7e1378fc8289c663f864e0487188fe8067/easse/utils/preprocessing.py#L7

    if lowercase:
        sentence = sentence.lower()

    if tokenizer in ["13a", "intl"]:
        if version.parse(sacrebleu.__version__).major >= 2:
            normalized_sent = sacrebleu.metrics.bleu._get_tokenizer(tokenizer)()(sentence)
        else:
            normalized_sent = sacrebleu.TOKENIZERS[tokenizer]()(sentence)
    elif tokenizer == "moses":
        normalized_sent = sacremoses.MosesTokenizer().tokenize(sentence, return_str=True, escape=False)
    elif tokenizer == "penn":
        normalized_sent = sacremoses.MosesTokenizer().penn_tokenize(sentence, return_str=True)
    else:
        normalized_sent = sentence

    if not return_str:
        normalized_sent = normalized_sent.split()

    return normalized_sent


def compute_sari(sources, predictions, references):

    if not (len(sources) == len(predictions) == len(references)):
        raise ValueError("Sources length must match predictions and references lengths.")
    sari_score = 0
    for src, pred, refs in zip(sources, predictions, references):
        sari_score += SARIsent(normalize(src), normalize(pred), [normalize(sent) for sent in refs])
    sari_score = sari_score / len(predictions)
    return 100 * sari_score


def compute_sacrebleu(
    predictions,
    references,
    smooth_method="exp",
    smooth_value=None,
    force=False,
    lowercase=False,
    use_effective_order=False,
):
    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)]
    output = sacrebleu.corpus_bleu(
        predictions,
        transformed_references,
        smooth_method=smooth_method,
        smooth_value=smooth_value,
        force=force,
        lowercase=lowercase,
        use_effective_order=use_effective_order,
    )
    return output.score


@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class WikiSplit(evaluate.Metric):
    def _info(self):
        return evaluate.MetricInfo(
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            features=[
                datasets.Features(
                    {
                        "predictions": datasets.Value("string", id="sequence"),
                        "references": datasets.Sequence(datasets.Value("string", id="sequence"), id="references"),
                    }
                ),
                datasets.Features(
                    {
                        "predictions": datasets.Value("string", id="sequence"),
                        "references": datasets.Value("string", id="sequence"),
                    }
                ),
            ],
            codebase_urls=[
                "https://github.com/huggingface/transformers/blob/master/src/transformers/data/metrics/squad_metrics.py",
                "https://github.com/cocoxu/simplification/blob/master/SARI.py",
                "https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/utils/sari_hook.py",
                "https://github.com/mjpost/sacreBLEU",
            ],
            reference_urls=[
                "https://www.aclweb.org/anthology/Q16-1029.pdf",
                "https://github.com/mjpost/sacreBLEU",
                "https://en.wikipedia.org/wiki/BLEU",
                "https://towardsdatascience.com/evaluating-text-output-in-nlp-bleu-at-your-own-risk-e8609665a213",
            ],
        )

    def _compute(self, sources, predictions, references):
        # 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]
        result = {}
        result.update({"sari": compute_sari(sources=sources, predictions=predictions, references=references)})
        result.update({"sacrebleu": compute_sacrebleu(predictions=predictions, references=references)})
        result.update({"exact": compute_em(predictions=predictions, references=references)})
        return result