aieng-lab/wortschatz-leipzig-de-grammar-neutral

Wortschatz Leipzig DE Grammar Neutral

Dataset Summary

aieng-lab/wortschatz-leipzig-de-grammar-neutral is a collection of German news sentences filtered to be as grammar-neutral as possible with respect to grammatical gender and case cues. It serves as a reference set of sentence contexts that (largely) avoid explicit morphosyntactic gender–case markers.

The dataset consists of 9,570 well-formed sentences and contains a single split (train) with a single column: text.

It is used in the codebase for the paper “Understanding or Memorizing? A Case Study of German Definite Articles in Language Models”.

• Source corpus: Wortschatz Leipzig German news corpus deu_news_2024_300K
• Code to (re)generate the dataset:
  https://github.com/aieng-lab/gradiend-german-articles/blob/main/gradiend/data/german_neutral.py
• spaCy model: de_core_news_sm==3.5.0

---

Dataset Creation

Goal

The goal is to create sentence contexts that do not provide obvious morphosyntactic cues for grammatical gender/case, to act as a grammar-neutral reference in experiments.

Linguistic Filtering (spaCy-based; exact implementation)

Filtering is implemented in german_neutral.py and uses spaCy (de_core_news_sm) to remove sentences that may contain explicit or implicit grammatical gender/case cues.

Given an input sentence sentence, we apply the following rules in this order:

1. Minimum length (characters)
   Skip sentences with fewer than 50 characters:

   • if len(sentence) < 50: continue

2. Named entity threshold (NER; conservative)
   To avoid sentences dominated by proper names, we run NER and discard sentences with more than 2 named entities.
   Implementation detail: line breaks are normalized by joining stripped lines with ", " before running NER:

   • ent_search = ", ".join(line.strip() for line in sentence.splitlines())
   • if len(nlp(ent_search).ents) > 2: continue

3. Exclude token das (homonym avoidance)
   After spaCy tokenization of the original sentence, discard any sentence that contains a token whose lowercased surface form is exactly das:

   • if any(token.text.lower() == "das" for token in doc): continue

4. Exclude determiners by POS
   Discard any sentence containing a token with spaCy POS tag DET:

   • if any(token.pos_ == "DET" for token in doc): continue

5. Exclude 3rd-person pronouns (by morphological Person feature)
   For tokens with POS PRON, read token.morph.get("Person").
   If Person is present and not in {1, 2}, discard the sentence. This removes (most) third-person pronouns:

   • if token.pos_ == "PRON":
     • person = token.morph.get("Person")
     • if person and person[0] not in ("1", "2"): skip

6. Exclude explicit article strings (regex; case-insensitive)
   As an additional surface-form guardrail, discard any sentence that matches the following word-boundary pattern (case-insensitive):

   \b(der|die|das|den|dem|des|ein|eine|einer|einem|einen|eines|kein|keine|keiner|keinem|keinen|keines)\b

   (Implemented via re.search(..., re.IGNORECASE).)

7. Deduplication
   After filtering, duplicates are removed by exact string match on the sentence string:

   • drop_duplicates(subset=["text"])

Only the final sentence text is uploaded to Hugging Face (single column: text).

---

Intended Use

• Grammar-neutral reference contexts for controlled experiments on German morphosyntax
• Baseline / control set when studying agreement behavior or debiasing methods
• Qualitative prompting contexts where explicit article/pronoun cues are undesirable

Not intended as general-purpose German language modeling data; it is heavily filtered and intentionally excludes common function words.

---

Limitations

• Automatic annotation noise: filters depend on spaCy tokenization, POS/morphology, and NER quality.
• Residual cues may remain (e.g., gendered occupational nouns, honorifics, titles, or culturally gendered names that slip past NER heuristics).
• The das exclusion is conservative and may remove otherwise valid neutral sentences.

---

Personal and Sensitive Information

The source is a news corpus; sentences may contain references to real people and organizations. While entity-heavy sentences are filtered out, this does not guarantee removal of personal data.

---

How to Use

from datasets import load_dataset

ds = load_dataset("aieng-lab/wortschatz-leipzig-de-grammar-neutral")
print(ds)
print(ds["train"][0]["text"])

Citation

@misc{drechsel2026understandingmemorizingcasestudy,
      title={Understanding or Memorizing? A Case Study of German Definite Articles in Language Models}, 
      author={Jonathan Drechsel and Erisa Bytyqi and Steffen Herbold},
      year={2026},
      eprint={2601.09313},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.09313}, 
}