aieng-lab/de-gender-case-articles

German Gender Case Articles

Dataset Summary

aieng-lab/de-gender-case-articles is a collection of German sentences containing definite singular articles in controlled gender × case settings. Each example provides an original sentence (text) and a version where all occurrences of the targeted article form (e.g., nominative-male der ) are replaced by a placeholder token ([MASK]). The gold label (label) is the grammatically licensed article form in uppercase (e.g., die).

The dataset is designed for fill-mask style evaluation and controlled experiments on German article agreement, and was used in the codebase for the paper “Understanding or Memorizing? A Case Study of German Definite Articles in Language Models”.

• Code to (re)generate the dataset: https://github.com/aieng-lab/gradiend-german-articles/blob/main/gradiend/data/german_articles.py
• spaCy model: de_core_news_sm==3.5.0 with this wheel

---

Dataset Structure

Configs (subsets)

The dataset is uploaded with one config per gender–case subset:

• Nominative: masc_nom, fem_nom, neut_nom
• Accusative: masc_acc, fem_acc, neut_acc
• Dative: masc_dat, fem_dat, neut_dat
• Genitive: masc_gen, fem_gen, neut_gen

Each config contains the HF splits train, validation, test.

Sizes

Dataset	ID	Total	Train	Val	Test
Nominative Masculine	masc_nom	34,350	27,829	3,084	3,437
Nominative Feminine	fem_nom	61,328	49,399	5,796	6,133
Nominative Neuter	neut_nom	33,350	26,680	3,335	3,335
Accusative Masculine	masc_acc	30,538	24,781	2,705	3,052
Accusative Feminine	fem_acc	34,801	28,155	3,166	3,480
Accusative Neuter	neut_acc	19,012	15,209	1,901	1,902
Dative Masculine	masc_dat	23,437	18,918	2,176	2,343
Dative Feminine	fem_dat	46,601	37,458	4,482	4,661
Dative Neuter	neut_dat	16,075	13,020	1,447	1,608
Genitive Masculine	masc_gen	34,087	27,417	3,254	3,416
Genitive Feminine	fem_gen	38,811	31,219	3,711	3,881
Genitive Neuter	neut_gen	25,351	20,436	2,387	2,528

---

Data Fields

Each example contains:

• text (string): original sentence
• masked (string): same sentence where every occurrence of the targeted article wrt. the selected gender-case pair is replaced by a placeholder [MASK]. Notice that for Nominative Male, for instance, not all occurences of der are masked, but only those corresponding to the usage as Nominative Male definite singular article, excluding its usage as, e.g., Genitive Female definite singular article.
• label (string): grammatically licensed article in lowercase, e.g. die
• token_count (int): number of masked occurrences (most often 1)
• dataset_label (string): short dataset identifier CASE+GENDER with CASE ∈ {N,A,D,G} and GENDER ∈ {M,F,N}, e.g. NM.
• split (string): train, validation, or test (duplicates HF split info)

---

Source Data

Sentences are extracted from the German Wikipedia dump (snapshot 20220301.de) and used as naturally occurring contexts for morphosyntactic filtering.

---

Annotation and Filtering

Sentences are processed with spaCy to obtain token-level POS tags and morphological features. Candidate determiner tokens are restricted to POS=DET. spaCy morphology is used to identify Gender, Case, and Number for definite singular articles.

For each gender–case combination, sentences are retained only if:

1. Article presence: the sentence contains at least one occurrence of the target article surface form.
2. Morphological agreement: all occurrences of the target article are annotated with Gender=g, Case=c, Number=Sing (plural uses excluded).
3. Limited ambiguity: sentences containing more than four occurrences of the target article are discarded.
4. Length constraints: keep sentences with 50–500 characters.
5. Named entity control: discard sentences containing more than three named entities.
6. Duplicate removal: remove duplicate sentences.

---

Intended Use

• Fill-mask style evaluation of article agreement in German
• Controlled probes of memorization vs. generalization for morphosyntactic patterns

Not intended as general-purpose German pretraining data: the dataset is heavily filtered and not a representative sample of Wikipedia.

---

Limitations

• Automatic annotation noise: spaCy POS/morphology and NER can be wrong, especially for ambiguous surface forms.
• Filtering (NER count, article-count limit, length limits) changes topical distribution and style.
• Focused scope: only definite singular articles in selected gender–case settings.

---

Personal and Sensitive Information

Wikipedia text can contain personal information about real individuals. This dataset reduces entity-heavy contexts via an NER-count filter (>3 named entities removed), but it does not guarantee removal of all personal data.

---

Licensing

This dataset is derived from Wikipedia content and is released under the same licensing regime used for Wikipedia dumps on Hugging Face:

• CC BY-SA 3.0
• GFDL

Users must comply with attribution and share-alike requirements.

---

How to Use

from datasets import load_dataset

# load one subset/config
ds = load_dataset("aieng-lab/de-gender-case-articles", "fem_acc")

example = ds["train"][0]
print(example["masked"])
print(example["label"])

Citation

@inproceedings{drechsel-etal-2026-understanding,
    title = "Understanding or Memorizing? A Case Study of {G}erman Definite Articles in Language Models",
    author = "Drechsel, Jonathan  and
      Bytyqi, Erisa  and
      Herbold, Steffen",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.436/",
    doi = "10.18653/v1/2026.acl-long.436",
    pages = "9626--9652",
    ISBN = "979-8-89176-390-6",
    abstract = "Language models perform well on grammatical agreement, but it is unclear whether this reflects rule-based generalization or memorization. We study this question for German definite singular articles, whose forms depend on gender and case. Using GRADIEND, a gradient-based interpretability method, we learn parameter update directions for gender-case specific article transitions. We find that updates learned for a specific gender-case article transition frequently affect unrelated gender-case settings, with substantial overlap among the most affected neurons across settings. These results argue against a strictly rule-based encoding of German definite articles, indicating that models at least partly rely on memorized associations rather than abstract grammatical rules."
}

Authors/ Maintainers

• Erisa Bytyqi
• Jonathan Drechsel (jdrechsel)