gsarti/mt_geneval
收藏资源简介:
MT-GenEval基准测试用于评估从英语到阿拉伯语、法语、德语、印地语、意大利语、葡萄牙语、俄语和西班牙语的性别翻译准确性。数据集包含带有性别化目标词注释的单个句子,以及带有额外上下文的对比性原始-反转翻译。数据集分为两种配置类型:`sentences`和`context`,分别包含单个句子和带有上下文的句子。数据集的大小在1K到10K之间,且由专家生成。
MT-GenEval is a benchmark for evaluating gender translation accuracy from English into Arabic, French, German, Hindi, Italian, Portuguese, Russian, and Spanish. The dataset comprises individual sentences annotated with gendered target words, as well as contrasting original-inverted translations with additional context. It is divided into two configuration types: `sentences` and `context`, which contain single sentences and context-augmented sentences respectively. The dataset size ranges from 1K to 10K, and it is expert-generated.
数据集卡片 for MT-GenEval
数据集描述
数据集摘要
MT-GenEval 基准评估了从英语到 {阿拉伯语, 法语, 德语, 印地语, 意大利语, 葡萄牙语, 俄语, 西班牙语} 的性别翻译准确性。该数据集包含带有性别目标词注释的单个句子,以及带有额外前置上下文的对比原始-反转翻译。
支持的任务和排行榜
机器翻译
有关使用 MT-GenEval 进行性别准确性评估的更多详细信息,请参阅原始论文。
语言
该数据集包含从维基百科提取的源英语句子,翻译成以下语言:阿拉伯语 (ar)、法语 (fr)、德语 (de)、印地语 (hi)、意大利语 (it)、葡萄牙语 (pt)、俄语 (ru) 和西班牙语 (es)。
数据集结构
数据实例
数据集包含两种配置类型,sentences 和 context,反映了原始仓库结构,源语言和目标语言在配置名称中指定(例如 sentences_en_ar,context_en_it)。sentences 配置包含带有性别词注释的单个句子的男性化和女性化版本。以下是 sentences_en_it 分割的一个示例条目(所有 sentences_en_XX 分割具有相同的结构):
json { "orig_id": 0, "source_feminine": "Pagratidis quickly recanted her confession, claiming she was psychologically pressured and beaten, and until the moment of her execution, she remained firm in her innocence.", "reference_feminine": "Pagratidis subito ritrattò la sua confessione, affermando che era aveva subito pressioni psicologiche e era stata picchiata, e fino al momento della sua esecuzione, rimase ferma sulla sua innocenza.", "source_masculine": "Pagratidis quickly recanted his confession, claiming he was psychologically pressured and beaten, and until the moment of his execution, he remained firm in his innocence.", "reference_masculine": "Pagratidis subito ritrattò la sua confessione, affermando che era aveva subito pressioni psicologiche e era stato picchiato, e fino al momento della sua esecuzione, rimase fermo sulla sua innocenza.", "source_feminine_annotated": "Pagratidis quickly recanted <F>her</F> confession, claiming <F>she</F> was psychologically pressured and beaten, and until the moment of <F>her</F> execution, <F>she</F> remained firm in <F>her</F> innocence.", "reference_feminine_annotated": "Pagratidis subito ritrattò la sua confessione, affermando che era aveva subito pressioni psicologiche e era <F>stata picchiata</F>, e fino al momento della sua esecuzione, rimase <F>ferma</F> sulla sua innocenza.", "source_masculine_annotated": "Pagratidis quickly recanted <M>his</M> confession, claiming <M>he</M> was psychologically pressured and beaten, and until the moment of <M>his</M> execution, <M>he</M> remained firm in <M>his</M> innocence.", "reference_masculine_annotated": "Pagratidis subito ritrattò la sua confessione, affermando che era aveva subito pressioni psicologiche e era <M>stato picchiato</M>, e fino al momento della sua esecuzione, rimase <M>fermo</M> sulla sua innocenza.", "source_feminine_keywords": "her;she;her;she;her", "reference_feminine_keywords": "stata picchiata;ferma", "source_masculine_keywords": "his;he;his;he;his", "reference_masculine_keywords": "stato picchiato;fermo" }
context 配置包含与刻板职业角色相关的不同英语源,带有额外的前置上下文和对比原始-反转翻译。以下是 context_en_it 分割的一个示例条目(所有 context_en_XX 分割具有相同的结构):
json { "orig_id": 0, "context": "Pierpont told of entering and holding up the bank and then fleeing to Fort Wayne, where the loot was divided between him and three others.", "source": "However, Pierpont stated that Skeer was the planner of the robbery.", "reference_original": "Comunque, Pierpont disse che Skeer era il pianificatore della rapina.", "reference_flipped": "Comunque, Pierpont disse che Skeer era la pianificatrice della rapina." }
数据分割
所有 sentences_en_XX 配置在 train 分割中有 1200 个示例,在 test 分割中有 300 个示例。对于 context_en_XX 配置,示例数量取决于语言对:
| 配置 | # 训练 | # 测试 |
|---|---|---|
context_en_ar |
792 | 1100 |
context_en_fr |
477 | 1099 |
context_en_de |
598 | 1100 |
context_en_hi |
397 | 1098 |
context_en_it |
465 | 1904 |
context_en_pt |
574 | 1089 |
context_en_ru |
583 | 1100 |
context_en_es |
534 | 1096 |
数据集创建
从原始论文中:
在开发 MT-GenEval 时,我们的目标是创建一个现实、性别平衡的数据集,自然地包含多样化的性别现象。为此,我们从维基百科中提取英语源句子作为我们数据集的基础。我们使用 EN 性别指代词基于 Zhao et al. (2018) 提供的列表自动预选相关句子。
有关数据集创建的更多信息,请参阅原始文章 MT-GenEval: A Counterfactual and Contextual Dataset for Evaluating Gender Accuracy in Machine Translation。
附加信息
数据集策展人
MT-GenEval 的原始作者是该数据集的策展人。对于此 🤗 Datasets 版本的任何问题或更新,请联系 gabriele.sarti996@gmail.com。
许可信息
该数据集根据 Creative Commons Attribution-ShareAlike 3.0 International License 进行许可。
引用信息
如果您在工作中使用这些语料库,请引用作者。
bibtex @inproceedings{currey-etal-2022-mtgeneval, title = "{MT-GenEval}: {A} Counterfactual and Contextual Dataset for Evaluating Gender Accuracy in Machine Translation", author = "Currey, Anna and Nadejde, Maria and Pappagari, Raghavendra and Mayer, Mia and Lauly, Stanislas, and Niu, Xing and Hsu, Benjamin and Dinu, Georgiana", booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing", month = dec, year = "2022", publisher = "Association for Computational Linguistics", url = "https://arxiv.org/abs/2211.01355", }




