import csv from pathlib import Path from typing import Dict, List, Tuple import datasets from datasets.download.download_manager import DownloadManager from seacrowd.utils import schemas from seacrowd.utils.configs import SEACrowdConfig from seacrowd.utils.constants import Licenses, Tasks _CITATION = r""" @inproceedings{kabra-etal-2023-multi, title = "Multi-lingual and Multi-cultural Figurative Language Understanding", author = "Kabra, Anubha and Liu, Emmy and Khanuja, Simran and Aji, Alham Fikri and Winata, Genta and Cahyawijaya, Samuel and Aremu, Anuoluwapo and Ogayo, Perez and Neubig, Graham", editor = "Rogers, Anna and Boyd-Graber, Jordan and Okazaki, Naoaki", booktitle = "Findings of the Association for Computational Linguistics: ACL 2023", month = jul, year = "2023", address = "Toronto, Canada", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2023.findings-acl.525", doi = "10.18653/v1/2023.findings-acl.525", pages = "8269--8284", } """ _LOCAL = False _LANGUAGES = ["ind", "jav", "sun"] _DATASETNAME = "mabl" _DESCRIPTION = r"""\ The MABL (Metaphors Across Borders and Languages) dataset is a collection of 6,366 figurative language expressions from seven languages, crafted to improve multilingual models' understanding of figurative speech and its linguistic variations. It was built by crowdsourcing native speakers to generate paired metaphors that began with the same words but had different meanings, as well as the literal interpretations of both phrases. Each expression was checked by fluent speakers to ensure they were clear, appropriate, and followed the format, discarding any that didn't meet these standards. """ _HOMEPAGE = "https://github.com/simran-khanuja/Multilingual-Fig-QA" _LICENSE = Licenses.MIT.value _URL = "https://raw.githubusercontent.com/simran-khanuja/Multilingual-Fig-QA/main/langdata/" _SUPPORTED_TASKS = [Tasks.QUESTION_ANSWERING] _SOURCE_VERSION = "1.0.0" _SEACROWD_VERSION = "2024.06.20" def iso3to2(lang: str) -> str: """Convert 3-letter ISO code to its 2-letter equivalent""" iso_map = {"ind": "id", "jav": "jv", "sun": "su"} return iso_map[lang] class MABLDataset(datasets.GeneratorBasedBuilder): """MABL dataset by Liu et al (2023)""" SOURCE_VERSION = datasets.Version(_SOURCE_VERSION) SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION) SEACROWD_SCHEMA_NAME = "qa" dataset_names = sorted([f"{_DATASETNAME}_{lang}" for lang in _LANGUAGES]) BUILDER_CONFIGS = [] for name in dataset_names: source_config = SEACrowdConfig( name=f"{name}_source", version=SOURCE_VERSION, description=f"{_DATASETNAME} source schema", schema="source", subset_id=name, ) BUILDER_CONFIGS.append(source_config) seacrowd_config = SEACrowdConfig( name=f"{name}_seacrowd_{SEACROWD_SCHEMA_NAME}", version=SEACROWD_VERSION, description=f"{_DATASETNAME} SEACrowd schema", schema=f"seacrowd_{SEACROWD_SCHEMA_NAME}", subset_id=name, ) BUILDER_CONFIGS.append(seacrowd_config) # Add configuration that allows loading all languages at once. BUILDER_CONFIGS.extend( [ # mabl_source SEACrowdConfig( name=f"{_DATASETNAME}_source", version=SOURCE_VERSION, description=f"{_DATASETNAME} source schema (all)", schema="source", subset_id=_DATASETNAME, ), # mabl_seacrowd_qa SEACrowdConfig( name=f"{_DATASETNAME}_seacrowd_{SEACROWD_SCHEMA_NAME}", version=SEACROWD_VERSION, description=f"{_DATASETNAME} SEACrowd schema (all)", schema=f"seacrowd_{SEACROWD_SCHEMA_NAME}", subset_id=_DATASETNAME, ), ] ) DEFAULT_CONFIG_NAME = f"{_DATASETNAME}_source" def _info(self) -> datasets.DatasetInfo: if self.config.schema == "source": features = datasets.Features( { "id": datasets.Value("string"), "startphrase": datasets.Value("string"), "ending1": datasets.Value("string"), "ending2": datasets.Value("string"), "labels": datasets.Value("string"), } ) elif self.config.schema == f"seacrowd_{self.SEACROWD_SCHEMA_NAME}": features = schemas.qa_features return datasets.DatasetInfo( description=_DESCRIPTION, features=features, homepage=_HOMEPAGE, license=_LICENSE, citation=_CITATION, ) def _split_generators(self, dl_manager: DownloadManager) -> List[datasets.SplitGenerator]: """Return SplitGenerators.""" mabl_source_data = [] languages = [] lang = self.config.name.split("_")[1] if lang in _LANGUAGES: # Load data per language mabl_source_data.append(dl_manager.download_and_extract(_URL + f"{iso3to2(lang)}.csv")) languages.append(lang) else: # Load examples for all languages at once. # We run this block when mabl_source / mabl_seacrowd_qa was chosen. for lang in _LANGUAGES: mabl_source_data.append(dl_manager.download_and_extract(_URL + f"{iso3to2(lang)}.csv")) languages.append(lang) return [ datasets.SplitGenerator( # The MABL paper mentions that due to the size of each subset, # they consider each split as a test set. name=datasets.Split.TEST, gen_kwargs={ "filepaths": mabl_source_data, "split": "test", "languages": languages, }, ) ] def _generate_examples(self, filepaths: List[Path], split: str, languages: List[str]) -> Tuple[int, Dict]: """Yield examples as (key, example) tuples""" startphrases = [] endings1 = [] endings2 = [] labels = [] for lang, filepath in zip(languages, filepaths): with open(filepath, encoding="utf-8") as f: csv_reader = csv.reader(f, delimiter=",") next(csv_reader, None) # skip the headers for row in csv_reader: # Unfortunately, the columns in the subfiles of the MABL # dataset are inconsistent. For 'ind', it is [ending1, # ending2, labels, startphrase]. But for 'jav' and 'sun', # the labels and startphrase columns were switched. Here, # I'm just hard-coding the column names if lang == "ind": end1, end2, label, start = row if lang == "jav" or lang == "sun": end1, end2, start, label = row startphrases.append(start) endings1.append(end1) endings2.append(end2) labels.append(label) for idx, (start, end1, end2, label) in enumerate(zip(startphrases, endings1, endings2, labels)): if self.config.schema == "source": example = { "id": str(idx), "startphrase": start, "ending1": end1, "ending2": end2, "labels": label, } elif self.config.schema == f"seacrowd_{self.SEACROWD_SCHEMA_NAME}": # Create QA-specific items choices = [end1, end2] answer = choices[int(label)] # MABL doesn't differentiate between question and context. # It only contains a startphrase. Given that, I put the # startphrase in question and kept the context blank. example = { "id": str(idx), "question_id": idx, "document_id": idx, "question": start, "type": "multiple_choice", "choices": choices, "context": "", "answer": [answer], "meta": {}, } yield idx, example