ReCaRe — A Bilingual Legal Benchmark for Revision Candidate Retrieval

ReCaRe (pronounced "re-care") is a bilingual legal IR benchmark grounded in EU and Japanese statutory amendments. It supports two Revision Candidate Retrieval (RCR) tasks over a shared per-language corpus:

• Rat2Rev — given an amendment's textual rationale, retrieve the articles modified by the amendment.
• Rev2Rev — given one already-revised article, retrieve other articles revised in the same legislative event.

Configurations

Config name	What it is	Records
corpus-en / corpus-ja	Per-language article corpus (BEIR canonical: _id, text)	91,361 / 90,170
queries-rat2rev-{en,ja}	Rat2Rev queries (rationale text)	340 / 363
queries-rev2rev-{en,ja}	Rev2Rev queries (pre-update article text)	1,509 / 1,653
qrels-rat2rev-{en,ja}	Rat2Rev relevance judgments (JSONL: query-id, corpus-id, score)	6,091 / 9,124 (train+val+test)
qrels-rev2rev-{en,ja}	Rev2Rev relevance judgments	28,433 / 43,498
metadata-{en,ja}	Verbatim construction-pipeline output (article-level provenance with before/after text and type_of_change)	91,361 / 90,170

Quick start

from datasets import load_dataset

corpus  = load_dataset("kasys/ReCaRe", "corpus-en", split="corpus")
queries = load_dataset("kasys/ReCaRe", "queries-rat2rev-en", split="queries")
qrels   = load_dataset("kasys/ReCaRe", "qrels-rat2rev-en")  # train/validation/test

Convert qrels to trec_eval format

from datasets import load_dataset
import pandas as pd
ds = load_dataset("kasys/ReCaRe", "qrels-rat2rev-en", split="test")
df = pd.DataFrame(ds).assign(iter=0)
df[["query-id", "iter", "corpus-id", "score"]].to_csv(
    "qrels.trec", sep=" ", header=False, index=False)

Splits

Train / validation / test split at amendment-event granularity (~1/3 each, inherited from the underlying construction pipeline). Same-event queries from both tasks are guaranteed to be in the same split.

Evaluation note: query–document overlap (Rev2Rev)

In Rev2Rev, queries are articles drawn from the corpus being searched, so the query document itself appears as a candidate at retrieval time and is trivially matched. Rebuilding a per-query corpus (with the query document excluded) is expensive — a separate index per query — so we recommend the simpler workaround: retrieve against the full corpus, then post-filter the TREC-format run file by dropping rows where qid == docid and promoting the subsequent ranks to fill the gap.

import pandas as pd
cols = ["qid", "Q0", "docid", "rank", "score", "tag"]
run = pd.read_csv("run.trec", sep=r"\s+", names=cols, dtype={"qid": str, "docid": str})
run = run[run["qid"] != run["docid"]].copy()
run["rank"] = run.groupby("qid").cumcount() + 1
run.to_csv("run.filtered.trec", sep=" ", header=False, index=False)

This does not affect Rat2Rev, where queries are amendment rationales rather than corpus articles.

Licenses

• EU subset: CC BY 4.0 (EUR-Lex)
• JP subset: CC BY 4.0 / CC0 (e-Gov 法令検索, 日本法令索引, 衆議院議案ページ)

Source identifiers (CELEX numbers, 法令履歴 IDs) are preserved in _id fields so each record is traceable to its origin.