Dataset Card for SLDS (Swiss Landmark Decisions Summarization)

Dataset Summary

The Swiss Landmark Decisions Summarization (SLDS) dataset is a large-scale, multilingual benchmark for judicial summarization. It contains over 20,000 landmark decisions issued by the Swiss Federal Supreme Court (SFSC) between 1954 and 2024, written in German, French, or Italian. Each decision is accompanied by headnotes authored in all three official languages, resulting in approximately 60,000 decision–headnote pairs.

Headnotes in Swiss law are concise, domain-specific digests written by clerks and judges, summarizing the key legal reasoning, cited laws, and case significance. Unlike typical abstractive summaries, they follow strict stylistic and legal conventions, making the summarization task highly challenging.

The dataset enables monolingual and cross-lingual summarization, supporting research in multilingual legal NLP, judicial reasoning, and evaluation of LLMs in specialized domains.

---

Supported Tasks

• Monolingual Summarization: Generate headnotes in the same language as the source decision.
• Cross-lingual Summarization: Generate headnotes in a different target language (e.g., German decision → French headnote).

The dataset has been used for benchmarking proprietary and open-source models (e.g., GPT-4o, Claude 3.5, DeepSeek, Qwen, Llama, Phi) across summarization tasks with traditional metrics (ROUGE, BERTScore) and a domain-specific LLM-as-a-Judge framework.

---

Languages

SLDS covers three official Swiss languages: German, French, and Italian

---

Dataset Structure

Data Fields

• sample_id: Unique identifier for a sample.
• decision_id: Identifier for a specific decision (appears three times, once per headnote language).
• decision: Full text of the decision (German, French, or Italian).
• decision_language: ISO code of decision language (de, fr, it).
• headnote: Official headnote text (legal citations, keywords, and free-form summary).
• headnote_language: ISO code of headnote language (de, fr, it).
• law_area: Legal domain of the case.
• year: Year of publication.
• volume: Official publication volume.
• url: Official link to the SFSC case.

---

Dataset Instances

Each decision appears once per headnote language, yielding three samples per case.
For example:

{
  "sample_id": "21646",
  "decision_id": "106 Ib 307",
  "decision": "<Full decision text in German>",
  "decision_language": "de",
  "headnote": "<Official headnote in French>",
  "headnote_language": "fr",
  "law_area": "administrative law and public international law",
  "year": 1980,
  "volume": "I",
  "url": "<Link to the decision on the SFSC repository>"   
}

---

Data Splits

The dataset is chronologically split to prevent leakage of stylistic or temporal trends:

Split	Years	# Decisions	# Samples	Language Distribution
Train	1954–2021	~20K	~60K	DE: 67.9%, FR: 27.4%, IT: 4.7%
Val	2022	200	600	DE: 68.5%, FR: 27.5%, IT: 4.0%
Test	2023–2024	326	978	DE: 63.5%, FR: 32.8%, IT: 3.7%

Dataset Configurations

The dataset provides multiple configurations:

• default contains all decision–headnote pairs combined (≈ 60k samples).
• Pairwise configs (e.g., de_fr, fr_it, it_de) restrict to a specific decision language (xx) and a specific headnote language (yy). For example, de_fr contains German decisions with French headnotes.
• Each decision appears three times across configs, once per headnote language.

One-shot Examples

Each config includes a small one_shot_examples split. These are predefined samples selected from validation to serve as prompting examples in few-shot settings, as described in the paper.

---

Dataset Creation

Curation Rationale

The dataset was created to provide a real-world multilingual legal benchmark for abstractive summarization. Unlike legislative corpora (e.g., EUR-Lex), SLDS focuses on case law, emphasizing concise and legally authoritative headnotes.

Source Data

• Collection: Decisions were scraped from the official SFSC archive (bger.ch).
• Coverage: 70 years (1954–2024), covering all five legal volumes (I–V).
• Processing:
  • Extracted full decision text and multilingual headnotes.
  • Normalized metadata (year, volume, law area).
  • Applied language detection and formatting.
  • Structured into decision–headnote pairs for training and evaluation.

Who are the source language producers?

Decisions and headnotes are written by judges and clerks of the Swiss Federal Supreme Court, the highest judicial body in Switzerland.

---

Annotations

• Annotation Process: Headnotes are official summaries, authored by clerks and judges, not crowdsourced.
• Annotators: Legal experts at the SFSC.
• Metadata: Derived from official publication metadata.

---

Personal and Sensitive Information

The dataset consists of publicly available legal documents. The SFSC applies strict anonymization guidelines before publication to protect personal data: Anonymisierungsregeln.

---

Considerations for Using the Data

Social Impact

SLDS supports multilingual access to Swiss case law, enabling legal professionals, researchers, and NLP systems to work across language barriers. It may assist in legal information retrieval, case comparison, and legal education.

Discussion of Biases

• Language imbalance: German dominates the dataset, reflecting its prevalence in Swiss court proceedings.
• Legal domain distribution: Some law areas (e.g., criminal law and criminal procedure) are more frequent, potentially biasing models.
• Stylistic rigidity: Headnotes follow legal drafting conventions that may not generalize to other summarization domains.

Other Known Limitations

• Headnotes are highly formulaic, which can lead to overfitting.
• Cross-lingual evaluation may be skewed by differences in legal phrasing traditions across languages.
• Evaluation metrics such as ROUGE may not fully capture legal correctness.

---

Licensing Information

Released under CC BY 4.0.

---

Citation Information

If you use SLDS in your work, please cite:

@inproceedings{rolshoven-etal-2025-unlocking,
    title = "Unlocking Legal Knowledge: A Multilingual Dataset for Judicial Summarization in {S}witzerland",
    author = {Rolshoven, Luca  and
      Rasiah, Vishvaksenan  and
      Bose, Srinanda Br{\"u}gger  and
      Hostettler, Sarah  and
      Burkhalter, Lara  and
      St{\"u}rmer, Matthias  and
      Niklaus, Joel},
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2025",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.findings-emnlp.832/",
    pages = "15382--15411",
    ISBN = "979-8-89176-335-7",
}