testbed/embeddings-benchmark__mteb/mteb/tasks/Retrieval/multilingual/MLQARetrieval.py

from typing import Dict, List

import datasets

from mteb.abstasks import AbsTaskRetrieval, CrosslingualTask, TaskMetadata

_LANGUAGES = {
    "mlqa.ar.ar": ["ara-Arab", "ara-Arab"],
    "mlqa.ar.de": ["ara-Arab", "deu-Latn"],
    "mlqa.ar.en": ["ara-Arab", "eng-Latn"],
    "mlqa.ar.es": ["ara-Arab", "spa-Latn"],
    "mlqa.ar.hi": ["ara-Arab", "hin-Deva"],
    "mlqa.ar.vi": ["ara-Arab", "vie-Latn"],
    "mlqa.ar.zh": ["ara-Arab", "zho-Hans"],
    "mlqa.de.ar": ["deu-Latn", "ara-Arab"],
    "mlqa.de.de": ["deu-Latn", "deu-Latn"],
    "mlqa.de.en": ["deu-Latn", "eng-Latn"],
    "mlqa.de.es": ["deu-Latn", "spa-Latn"],
    "mlqa.de.hi": ["deu-Latn", "hin-Deva"],
    "mlqa.de.vi": ["deu-Latn", "vie-Latn"],
    "mlqa.de.zh": ["deu-Latn", "zho-Hans"],
    "mlqa.en.ar": ["eng-Latn", "ara-Arab"],
    "mlqa.en.de": ["eng-Latn", "deu-Latn"],
    "mlqa.en.en": ["eng-Latn", "eng-Latn"],
    "mlqa.en.es": ["eng-Latn", "spa-Latn"],
    "mlqa.en.hi": ["eng-Latn", "hin-Deva"],
    "mlqa.en.vi": ["eng-Latn", "vie-Latn"],
    "mlqa.en.zh": ["eng-Latn", "zho-Hans"],
    "mlqa.es.ar": ["spa-Latn", "ara-Arab"],
    "mlqa.es.de": ["spa-Latn", "deu-Latn"],
    "mlqa.es.en": ["spa-Latn", "eng-Latn"],
    "mlqa.es.es": ["spa-Latn", "spa-Latn"],
    "mlqa.es.hi": ["spa-Latn", "hin-Deva"],
    "mlqa.es.vi": ["spa-Latn", "vie-Latn"],
    "mlqa.es.zh": ["spa-Latn", "zho-Hans"],
    "mlqa.hi.ar": ["hin-Deva", "ara-Arab"],
    "mlqa.hi.de": ["hin-Deva", "deu-Latn"],
    "mlqa.hi.en": ["hin-Deva", "eng-Latn"],
    "mlqa.hi.es": ["hin-Deva", "spa-Latn"],
    "mlqa.hi.hi": ["hin-Deva", "hin-Deva"],
    "mlqa.hi.vi": ["hin-Deva", "vie-Latn"],
    "mlqa.hi.zh": ["hin-Deva", "zho-Hans"],
    "mlqa.vi.ar": ["vie-Latn", "ara-Arab"],
    "mlqa.vi.de": ["vie-Latn", "deu-Latn"],
    "mlqa.vi.en": ["vie-Latn", "eng-Latn"],
    "mlqa.vi.es": ["vie-Latn", "spa-Latn"],
    "mlqa.vi.hi": ["vie-Latn", "hin-Deva"],
    "mlqa.vi.vi": ["vie-Latn", "vie-Latn"],
    "mlqa.vi.zh": ["vie-Latn", "zho-Hans"],
    "mlqa.zh.ar": ["zho-Hans", "ara-Arab"],
    "mlqa.zh.de": ["zho-Hans", "deu-Latn"],
    "mlqa.zh.en": ["zho-Hans", "eng-Latn"],
    "mlqa.zh.es": ["zho-Hans", "spa-Latn"],
    "mlqa.zh.hi": ["zho-Hans", "hin-Deva"],
    "mlqa.zh.vi": ["zho-Hans", "vie-Latn"],
    "mlqa.zh.zh": ["zho-Hans", "zho-Hans"],
}


def _build_lang_pair(langs: List[str]) -> str:
    """Builds a language pair separated by a dash.
    e.g., ['eng-Latn', 'deu-Latn'] -> 'eng-deu'.
    """
    return langs[0].split("-")[0] + "-" + langs[1].split("-")[0]


def extend_lang_pairs() -> Dict[str, List[str]]:
    eval_langs = {}
    for langs in _LANGUAGES.values():
        lang_pair = _build_lang_pair(langs)
        eval_langs[lang_pair] = langs
    return eval_langs


_EVAL_LANGS = extend_lang_pairs()


class MLQARetrieval(AbsTaskRetrieval, CrosslingualTask):
    metadata = TaskMetadata(
        name="MLQARetrieval",
        description="""MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance.
        MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic,
        German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between
        4 different languages on average.""",
        reference="https://huggingface.co/datasets/mlqa",
        dataset={
            "path": "facebook/mlqa",
            "revision": "397ed406c1a7902140303e7faf60fff35b58d285",
        },
        type="Retrieval",
        category="s2p",
        eval_splits=["validation", "test"],
        eval_langs=_EVAL_LANGS,
        main_score="ndcg_at_10",
        date=("2019-01-01", "2020-12-31"),
        form=["written"],
        domains=["Encyclopaedic"],
        task_subtypes=["Question answering"],
        license="cc-by-sa-3.0",
        socioeconomic_status="mixed",
        annotations_creators="human-annotated",
        dialect=[],
        text_creation="found",
        bibtex_citation="""@article{lewis2019mlqa,
        title = {MLQA: Evaluating Cross-lingual Extractive Question Answering},
        author = {Lewis, Patrick and Oguz, Barlas and Rinott, Ruty and Riedel, Sebastian and Schwenk, Holger},
        journal = {arXiv preprint arXiv:1910.07475},
        year = 2019,
        eid = {arXiv: 1910.07475}
        }""",
        n_samples={"test": 158083, "validation": 15747},
        avg_character_length={
            "test": 37352.28,
            "validation": 36952.7,
        },  # avergae context lengths
    )

    def load_data(self, **kwargs):
        """In this retrieval datasets, corpus is in lang XX and queries in lang YY."""
        if self.data_loaded:
            return

        _dataset_raw = {}
        self.queries, self.corpus, self.relevant_docs = {}, {}, {}

        for hf_subset, langs in _LANGUAGES.items():
            # Builds a language pair separated by an underscore. e.g., "ara-Arab_eng-Latn".
            # Corpus is in ara-Arab and queries in eng-Latn
            lang_pair = _build_lang_pair(langs)

            _dataset_raw[lang_pair] = datasets.load_dataset(
                name=hf_subset,
                **self.metadata_dict["dataset"],
            )
            _dataset_raw[lang_pair] = _dataset_raw[lang_pair].rename_column(
                "context", "text"
            )

            self.queries[lang_pair] = {
                eval_split: {
                    str(i): q["question"]
                    for i, q in enumerate(_dataset_raw[lang_pair][eval_split])
                }
                for eval_split in self.metadata_dict["eval_splits"]
            }

            self.corpus[lang_pair] = {
                eval_split: {
                    str(row["id"]): row for row in _dataset_raw[lang_pair][eval_split]
                }
                for eval_split in self.metadata_dict["eval_splits"]
            }

            self.relevant_docs[lang_pair] = {
                eval_split: {
                    str(i): {str(q["id"]): 1}
                    for i, q in enumerate(_dataset_raw[lang_pair][eval_split])
                }
                for eval_split in self.metadata_dict["eval_splits"]
            }

        self.data_loaded = True