morphbench-de
收藏资源简介:
MorphBench (German) 是一个用于评估德语形态感知分词器和语言模型的综合性基准测试套件。该数据集包含七个核心任务,旨在系统性地探究德语的形态学能力,包括词形屈折变化、派生词分割、派生词构建、复合词分割、词缀功能分类、派生词定义生成以及派生词定义多选。每个任务作为一个独立的配置,数据被组织为多个分片,如训练集、开发集以及多个具有不同泛化难度的测试集。这些测试分片根据词基和目标任务形式在预训练语料中的暴露频率进行划分,核心评估目标是模型在词基可见而目标形式不可见情况下的组合泛化能力。数据来源于德语维基词典和UniMorph项目,数据集规模可观,例如,任务1(屈折变化)包含约1.5万条样本,任务6a(定义生成)包含约1.2万条样本。该数据集适用于自然语言处理中形态学分析、分词、语言模型理解和生成等任务的评估与研究。
MorphBench (German) is a comprehensive benchmark suite for evaluating German morphology-aware tokenizers and language models. The dataset consists of seven core tasks designed to systematically investigate the morphological capabilities of German, including inflection, derivation segmentation, derivation construction, compound segmentation, affix function classification, derivation definition generation, and derivation definition multiple-choice questions. Each task is organized as an independent configuration, with data split into multiple subsets such as train, dev, and various test sets with different generalization difficulties. These test subsets are divided based on the exposure frequency of lemmas/bases and target forms in pre-training corpora, with the core evaluation goal being the models compositional generalization ability in scenarios where the lemma is seen but the target form is unseen. The data is sourced from the German Wiktionary and the UniMorph project, and the dataset is substantial in scale; for example, Task 1 (inflection) contains approximately 15,000 samples, and Task 6a (definition generation) contains about 12,000 samples. This dataset is suitable for evaluation and research in natural language processing tasks such as morphological analysis, tokenization, and language model understanding and generation.
数据集总览
MorphBench (German) 是一个德语词法评估基准,旨在测试词法感知的分词器和语言模型。数据集遵循 CC BY-SA 4.0 许可,语言为德语,涵盖7个子任务(共8个配置),每个子任务对应一个独立的配置(config)。
子任务详情
| 配置名称 | 任务描述 | 输入字段 | 输出字段 | 示例 |
|---|---|---|---|---|
task1_inflection |
根据词元+特征生成屈折形式 | lemma, feats, target, status |
- | interagieren + v;ind;pl;1;pst → interagierten |
task2_deriv_segmentation |
将派生词切分为词基+词缀 | word, segmentation, status |
- | Dreher → `drehen |
task3_derivation |
根据词基+词缀构建派生词 | base, affix, target, status |
- | Apotheke + pflichtig → apothekenpflichtig |
task4_compound |
将复合词切分为组成部分 | word, segmentation, difficulty, source |
- | Freiheitsheld → `Freiheit |
task5_affix_function |
分类派生词中词缀的语义功能(13类) | word, function, freq, status |
- | stimmlos → without |
task6a_definition |
生成派生词的词典释义 | word, gloss, base, affix, function, derived_freq, base_exposure, status |
- | Malocher → Arbeitnehmer, der überwiegend körperlich hart … arbeitet |
task6b_definition_mcq |
多选题版本:从干扰项中选择正确释义 | word, base, affix, function, gold_gloss, distractors, pretrain_status, derived_freq, base_exposure |
- | - |
数据划分与难度层级
每个子任务包含多个 split,通过预训练暴露度(在BabyLM风格德语语料库中的频率)定义,核心思想是测试组合泛化能力(目标词未见,但其组成部分已见):
| 配置 | 划分集合 | 划分条件 | 测试目标 |
|---|---|---|---|
task1_inflection |
train, dev, test, test_rare, test_memorization |
test: 词元已见,屈折形式未见;test_rare: 词元未见;test_memorization: 词元已见且形式已见 | 组合泛化 vs. 记忆 |
task2_deriv_segmentation |
train, dev, test_main, test_memorization, test_oov |
test_main: 词基已见,派生词未见;test_oov: 词基未见;test_memorization: 两者均已见 | 组合泛化 vs. 记忆 vs. OOV |
task3_derivation |
train, dev, test_main, test_memorization, test_oov |
同上 | 同上 |
task4_compound |
train, dev, test_main, test_rare, test_memorization, test_hardest |
test_main: 复合词未见,所有成分已见;test_hardest: 复合词未见,某些成分未见;test_memorization: 复合词已见;test_rare: 混合 | 组合泛化 vs. 最难 vs. 记忆 |
task5_affix_function |
train, dev, test |
无难度分层 | 标准监督 |
task6a_definition |
train, dev, test_main, test_main_oov, test_rare, test_memorization |
main: 派生词频率=0且词基暴露≥300;main_oov: 派生词频率=0且词基暴露<300;rare: 频率1-5;memorization: 频率≥20 | 泛化 vs. 罕见 vs. 记忆 |
task6b_definition_mcq |
dev, main, main_oov, rare, memorization |
同上 | 同上 |
数据规模与文件
- 数据集总大小:通过各配置的
download_size和dataset_size数据,所有配置的总下载大小约为 3,332,690 字节,总数据集大小约为 7,519,799 字节。 - 数据文件格式:每个配置的数据文件路径为
{config_name}/{split}-*,例如task1_inflection/train-*。
来源与许可证
- 来源:构建自德语维基词典(通过 kaikki.org / wiktextract)和 UniMorph。
- 同级资源:英语版对应数据集为 yuanxin112/morphbench-en,辅助词汇资源为 yuanxin112/wiktionary-morph。
- 许可证:CC BY-SA 4.0(归因于维基词典及其贡献者)。




