README.md

metadata

license: cc-by-4.0
language:
  - de
tags:
  - legal
  - statutory-interpretation
  - german-law
pretty_name: Interpretive Canons (decision-level)
configs:
  - config_name: default
    data_files:
      - split: train
        path: decisions.jsonl

Interpretive Canons — decision-level

Companion benchmark to the paper Classifying Interpretive Canons at the Sentence Level: A Benchmark from the German Federal Constitutional Court. This is the decision-level release: one record per fully (exhaustively) annotated decision, intended for end-to-end evaluation that runs the full pipeline over a whole decision. Only the 15 exhaustively annotated decisions are included; the selectively annotated decisions are not, because their unannotated sentences are not negatives. The companion instance-level benchmark (eight subtask files with splits) is at felix453/interpretive-canons-instances.

Record format (decisions.jsonl, one decision per line)

• decision_name: BVerfGE / file identifier.
• plain_text: the full text of the reasons (Entscheidungsgründe).
• document_structure: paragraph/sentence structure inherited from the L.L.Con corpus (ebene1, absatz, satz, ...), with character offsets into plain_text.
• readings: every annotated reading (Deutung) as a character span (start, end, satz_id) with its criteria — abstrakt, selbst_aufgestellt, konkretes_gesetz (list of provisions), bestimmt, vermutlich_keine_behauptung — and the free-text reasonings.
• arguments: every annotated (reading, candidate) pair as a character span with deutung_id, the general_argument gate (present + reasoning), and the four canon labels wortlaut/systematik/geschichte/zweck (null when the gate is negative).

Loading

from datasets import load_dataset
ds = load_dataset("felix453/interpretive-canons-decisions", split="train")
print(ds[0]["decision_name"], len(ds[0]["readings"]))

Raw annotations (provenance)

The raw, pre-merge Label Studio export these data are built from is available at felix453/interpretive-canons-raw-export. It retains two fields the postprocessing merge does not preserve: the per-reading determinate-content judgment bestimmt_moeglicheDeutung (dropped by the merge) and the un-flattened list of concrete provisions (the merge joins the entries into one unsplittable string). Join by Label Studio region id.