docs: refresh data card with FAIR principles + Datasheet for Datasets (Gebru et al. 2018)

README.md

@@ -1,112 +1,135 @@
 ---
-license: cc-by-4.0
+license:
+  - cc-by-4.0
+  - cc0-1.0
 language:
   - en
   - ja
 task_categories:
   - text-retrieval
-pretty_name: ReCaRe
+pretty_name: ReCaRe — A Bilingual Legal Benchmark for Revision Candidate Retrieval
+size_categories:
+  - 100K<n<1M
 tags:
   - legal
-  - revision-candidate-retrieval
+  - law
+  - multilingual
   - bilingual
+  - retrieval
+  - benchmark
+  - bm25
+  - dense-retrieval
+  - cross-lingual
+  - eu-law
+  - japanese-law
 configs:
   - config_name: corpus-en
-    data_files:
-      - split: corpus
-        path: corpus-en/corpus.jsonl
   - config_name: corpus-ja
-    data_files:
-      - split: corpus
-        path: corpus-ja/corpus.jsonl
   - config_name: queries-rat2rev-en
-    data_files:
-      - split: queries
-        path: queries-rat2rev-en/queries.jsonl
   - config_name: queries-rat2rev-ja
-    data_files:
-      - split: queries
-        path: queries-rat2rev-ja/queries.jsonl
   - config_name: queries-rev2rev-en
-    data_files:
-      - split: queries
-        path: queries-rev2rev-en/queries.jsonl
   - config_name: queries-rev2rev-ja
-    data_files:
-      - split: queries
-        path: queries-rev2rev-ja/queries.jsonl
   - config_name: qrels-rat2rev-en
-    data_files:
-      - split: train
-        path: qrels-rat2rev-en/train.jsonl
-      - split: validation
-        path: qrels-rat2rev-en/validation.jsonl
-      - split: test
-        path: qrels-rat2rev-en/test.jsonl
   - config_name: qrels-rat2rev-ja
-    data_files:
-      - split: train
-        path: qrels-rat2rev-ja/train.jsonl
-      - split: validation
-        path: qrels-rat2rev-ja/validation.jsonl
-      - split: test
-        path: qrels-rat2rev-ja/test.jsonl
   - config_name: qrels-rev2rev-en
-    data_files:
-      - split: train
-        path: qrels-rev2rev-en/train.jsonl
-      - split: validation
-        path: qrels-rev2rev-en/validation.jsonl
-      - split: test
-        path: qrels-rev2rev-en/test.jsonl
   - config_name: qrels-rev2rev-ja
-    data_files:
-      - split: train
-        path: qrels-rev2rev-ja/train.jsonl
-      - split: validation
-        path: qrels-rev2rev-ja/validation.jsonl
-      - split: test
-        path: qrels-rev2rev-ja/test.jsonl
   - config_name: metadata-en
-    data_files:
-      - split: metadata
-        path: metadata-en/dataset.jsonl
   - config_name: metadata-ja
-    data_files:
-      - split: metadata
-        path: metadata-ja/dataset.jsonl
 ---
 
 # 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:
+> ReCaRe (pronounced "re-care") supports retrieval research on **document
+> maintenance**: locating which provisions of a legal corpus must change when
+> a new amendment is proposed, and which other provisions must change with
+> them. Built from European Union law (EUR-Lex) and Japanese law (e-Gov)
+> across 703 amendment events and ~181k articles, ReCaRe defines two
+> Revision Candidate Retrieval (RCR) tasks — Rat2Rev and Rev2Rev — over a
+> shared bilingual 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.
+- **Repository**: <https://huggingface.co/datasets/kasys/ReCaRe>
+- **Owner / Org**: `kasys` (HuggingFace) — corresponds to `kasys-lab` on GitHub
+- **License**: CC BY 4.0 (EU subset) / CC BY 4.0 + CC0 1.0 (Japanese subset)
+- **DOI**: _to be filled in once HF Settings → Generate DOI is run on `v0.1.0`_
+- **Version**: `v0.1.0` (initial public release for CIKM 2026 Resource paper)
+- **Code**: analysis & baseline code at `kasys-lab/ReCaRe` (GitHub, see #15)
 
-## Configurations
+---
+
+## Dataset summary
+
+Document corpora in regulated settings are not static. Statutes are amended,
+internal policies are revised, software specifications are updated. Yet most
+information retrieval (IR) research has framed retrieval as a one-shot
+question-answering problem over a frozen corpus, leaving the IR aspects of
+*document maintenance* — finding which documents need to change, and which
+other documents must change with them — comparatively underexplored.
+
+ReCaRe formalises two complementary retrieval problems that arise during
+document maintenance:
+
+- **Rat2Rev** (Rationale-to-Revision Retrieval): given the *textual rationale*
+  of a proposed amendment (long, abstract), retrieve the concrete articles
+  that must be modified to implement the amendment.
+- **Rev2Rev** (Revision-to-Revision Retrieval): given one already-revised
+  article, retrieve the other articles revised in the same legislative event
+  (co-revised articles).
+
+Both tasks share a single bilingual corpus (~91k EU articles in English +
+~90k Japanese articles) and are released under CC BY 4.0 / CC0.
+
+The dataset is the resource artifact of the CIKM 2026 paper
+*"ReCaRe: A Bilingual Legal Benchmark for Revision Candidate Retrieval."*
+
+## Configs (12)
 
-| 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 |
+| Config | Splits | Schema | Records |
+| --- | --- | --- | --- |
+| `corpus-en` | `corpus` | `{_id, text}` | 91,361 |
+| `corpus-ja` | `corpus` | `{_id, text}` | 90,170 |
+| `queries-rat2rev-en` | `queries` | `{_id, text}` | 340 |
+| `queries-rat2rev-ja` | `queries` | `{_id, text}` | 363 |
+| `queries-rev2rev-en` | `queries` | `{_id, text}` | 1,509 |
+| `queries-rev2rev-ja` | `queries` | `{_id, text}` | 1,653 |
+| `qrels-rat2rev-en` | `train` / `validation` / `test` | `{query-id, corpus-id, score}` | 2,063 / 1,948 / 2,080 |
+| `qrels-rat2rev-ja` | same | same | 3,228 / 2,501 / 3,395 |
+| `qrels-rev2rev-en` | same | same | 12,088 / 8,189 / 8,156 |
+| `qrels-rev2rev-ja` | same | same | 15,054 / 13,591 / 14,853 |
+| `metadata-en` | `metadata` | 16-field amendment metadata (see below) | 91,361 |
+| `metadata-ja` | `metadata` | same | 90,170 |
 
+> All configs are JSONL. Use a 4-line snippet (below) to convert qrels to
+> TREC format if your stack expects that.
 
+### `metadata-*` schema (verbatim from construction pipeline)
+
+`amendment_law_id`, `law_id`, `type_of_change`, `egov_compare_url`,
+`law_title_before`, `revision_id_before`, `article_id_before`,
+`article_number_before`, `caption_before`, `text_before`,
+`law_title_after`, `revision_id_after`, `article_id_after`,
+`article_number_after`, `caption_after`, `text_after`.
+
+Records with `amendment_law_id == "None"` are unchanged articles
+(~85–93%); records with non-`None` amendment IDs are revisions traceable to
+a specific amending act. The metadata configs are intended for downstream
+analyses that need full provenance and are **not** required for running the
+retrieval tasks themselves (use the `corpus`, `queries`, `qrels` configs).
 
 ## Quick start
 
+### Python (`datasets`)
+
 ```python
 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
-```
 
+print(corpus[0])           # {'_id': '...', 'text': '...'}
+print(queries[0])          # {'_id': '...', 'text': '...'}
+print(qrels["test"][0])    # {'query-id': '...', 'corpus-id': '...', 'score': 1}
+```
 
 ### Convert qrels to `trec_eval` format
 
@@ -119,29 +142,251 @@ df[["query-id", "iter", "corpus-id", "score"]].to_csv(
     "qrels.trec", sep=" ", header=False, index=False)
 ```
 
-## Splits
+### BEIR loader (after rename)
 
-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.
+```python
+from datasets import load_dataset
+import pandas as pd, os
+os.makedirs("recare-en/qrels", exist_ok=True)
+load_dataset("kasys/ReCaRe", "corpus-en", split="corpus")\
+    .to_json("recare-en/corpus.jsonl", lines=True, force_ascii=False)
+load_dataset("kasys/ReCaRe", "queries-rat2rev-en", split="queries")\
+    .to_json("recare-en/queries.jsonl", lines=True, force_ascii=False)
+pd.DataFrame(load_dataset("kasys/ReCaRe", "qrels-rat2rev-en", split="test"))\
+    .to_csv("recare-en/qrels/test.tsv", sep="\t", index=False)
+# then:
+from beir.datasets.data_loader import GenericDataLoader
+corpus, queries, qrels = GenericDataLoader(data_folder="recare-en").load(split="test")
+```
 
-## 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.
+## FAIR principles
 
-```python
-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)
-```
+ReCaRe is published with explicit attention to the FAIR principles
+(Findable, Accessible, Interoperable, Reusable) for research data
+[Wilkinson et al. 2016, *Sci. Data*].
+
+### F — Findable
+
+The dataset has a globally unique persistent identifier — a HuggingFace-issued
+DOI (DataCite) on this page (Settings → Generate DOI on tag `v0.1.0`), and is
+discoverable via HuggingFace Hub search, the `datasets` library, the project
+homepage, and the CIKM 2026 paper. A Zenodo mirror with an additional
+permanent DOI is published in the companion repository. Each record carries
+a stable `_id` (article identifier) traceable back to its source act
+(CELEX number for EU, e-Gov / 法令索引 / 衆議院議案 ID for JP) so that
+individual articles, queries, and amendments can be cited unambiguously.
+
+### A — Accessible
+
+The full dataset is openly accessible without authentication or registration
+via the HuggingFace `datasets` library and via direct HTTPS download from
+this page. The repository is `public` and licenses are
+machine-readable (SPDX) in the YAML front-matter. We commit to keeping the
+artifact accessible for the lifetime of HuggingFace; should the platform
+become unavailable, the Zenodo mirror remains canonical and self-describing
+JSONL means consumers can ingest the data with any standard parser.
+
+### I — Interoperable
+
+The data is stored in **JSONL with BEIR-canonical schemas** so it works out
+of the box with the existing IR ecosystem (BEIR loaders, `ir_measures`,
+`pytrec_eval`, Pyserini, sentence-transformers). Field names follow BEIR
+conventions (`_id`, `text`, `query-id`, `corpus-id`, `score`); a one-paste
+TREC qrels conversion snippet is provided above for `trec_eval`. The
+metadata configs use a flat 16-field schema documented in this README, so
+relational joins back to amendment-level provenance are straightforward.
+
+### R — Reusable
+
+Each record is released under a clearly identified open license (CC BY 4.0
+for EU sources; CC BY 4.0 / CC0 1.0 for Japanese sources, both inherited
+from the upstream public bodies). Provenance is preserved at the record
+level via the `metadata-*` configs, which carry the source amendment ID,
+the canonical comparison URL, and the before/after article text. Construction
+methodology (alignment via Dice/Simpson with minor-edit filtering, query
+typing, train/validation/test split inheritance) is documented in the
+companion paper and summarised in the *Datasheet* section below. The
+versioning scheme is semver-aligned (`v0.1.0` for this release); future
+patches that reissue qrels or correct alignment errors will land as
+`v0.1.x`.
+
+---
+
+## Datasheet for Datasets
+
+The following follows the *Datasheets for Datasets* template
+[Gebru et al. 2018, arXiv:1803.09010] and answers each section concisely.
+For details, see the CIKM 2026 paper.
+
+### Motivation
+
+ReCaRe was created because existing legal-IR benchmarks (COLIEE, BSARD,
+LeCaRD, STARD, LegalBench-RAG) cover statutory or case-law question
+answering but not *document maintenance* — finding which provisions must
+change when a corpus evolves. Two retrieval tasks of high practical
+relevance to legal drafting and policy revision (Rat2Rev, Rev2Rev) had no
+shared benchmark. The dataset was constructed by Hiroyoshi Itō, Yūma
+Kurokawa, Makoto P. Kato (University of Tsukuba) and Sumio Fujita
+(LY Corporation) for academic research, with no commissioned funding.
+
+### Composition
+
+Each instance is either an **article** (a single legal provision; ~181k
+total across two languages), a **query** (an amendment rationale text in
+Rat2Rev, or a revised article in Rev2Rev; 703 + 3,162 queries respectively),
+or a **qrel** (a relevance judgment of the form `(query, article)`).
+Coverage spans 340 EU amendment events (2010–2025) and 363 Japanese
+amendment events (2019–2025). Articles contain only public legal text;
+metadata records carry IDs and timestamps but no personal information.
+There are no missing values in the released JSONL — articles or amendments
+without complete provenance were dropped during construction.
+
+### Collection process
+
+Source documents were retrieved from public legal portals: EUR-Lex
+(CC BY 4.0 / CC0; CELEX-numbered consolidated acts) for EU law, and
+e-Gov 法令検索 + 日本法令索引 + 衆議院議案 (CC BY 4.0 / CC0) for Japanese
+law. We downloaded full consolidated text per act, extracted articles by
+the official numbering scheme, and collected explicit
+amendment events (acts that explicitly amend prior acts). Acquisition spanned
+2025-Q3–2026-Q1. No human subjects were involved. The two-language design
+follows from the choice to ground the benchmark in two jurisdictions whose
+legal-revision practices differ (EU consolidation tradition vs. Japanese
+amending-act tradition).
+
+### Preprocessing / cleaning / labeling
+
+Articles are paired with revisions via **Dice / Simpson coefficient
+matching** between before/after article text, with **minor-edit removal**
+filtering out punctuation-only and renumbering-only changes. Queries for
+Rat2Rev are the official rationale text of each amending act; queries for
+Rev2Rev are individual revised articles (one query per revised article, up
+to 5 queries per amendment). Qrels are derived deterministically from the
+amendment alignment: a `(rationale, article)` pair is positive iff the
+article is among those revised by that amendment; a `(revised-article,
+article)` pair is positive iff both are revised by the same amendment
+event. Train / validation / test splits are inherited from
+the prior DEIM 2026 papers (Itō et al. 2F-01, Kurokawa et al. 4F-04) for
+direct comparability with their reported numbers. Construction code is
+**not** released (it depends on internal scrapers); the relationship is
+documented in the paper, and label validity is independently verified by
+two-rater blind annotation on a stratified 200-pair sample (see *Limitations*).
+
+### Uses
+
+The dataset is intended for academic IR research on document maintenance,
+revision retrieval, multilingual / cross-jurisdictional retrieval, and as
+a difficulty contrast point for general-domain BEIR benchmarks (low
+query-document lexical overlap, multi-target qrels, implicit dependency).
+It is **not** intended to support automated legal drafting, automated
+revision recommendation in production legal workflows, or legal advice;
+retrieved articles are revision **candidates** for expert review, not
+authoritative outputs. Users should not infer that a model performing well
+on ReCaRe is suitable for unsupervised legal-corpus maintenance. Other
+appropriate uses include: training and evaluation of multilingual dense
+retrievers, ablation of long-context vs. short-context retrieval models,
+cross-task generalisation studies between Rat2Rev and Rev2Rev, and as
+held-out legal benchmark in foundation-model evaluations.
+
+### Distribution
+
+ReCaRe is distributed publicly via HuggingFace Datasets (`kasys/ReCaRe`,
+this page) under CC BY 4.0 / CC0 (per source-language license), with a
+HuggingFace-issued DOI (DataCite) on tag `v0.1.0`. A Zenodo mirror with an
+additional permanent DOI provides redundancy. The dataset is freely
+downloadable without registration. Source code for analysis / baselines is
+distributed separately at `kasys-lab/ReCaRe` (GitHub).
 
-This does not affect Rat2Rev, where queries are amendment rationales rather than corpus articles.
+### Maintenance
 
-## Licenses
+The dataset is maintained by the authors. Versioning follows semver:
+`v0.1.0` is the initial public release; corrective releases (e.g. label
+fixes, new amendment events) will land as `v0.1.x` (patch) or `v0.x.0`
+(minor). Each release is git-tagged on this HuggingFace repo and mirrored
+to Zenodo. Issues (errata, license clarifications, schema questions) are
+tracked at the project's GitHub repository
+(<https://github.com/mpkato/CIKM2026-ito-kurokawa>). There is no
+guaranteed support timeline, but we intend to keep the dataset accessible
+and answer well-formed issues for at least the duration of the CIKM 2026
+review cycle and the subsequent academic year.
+
+---
+
+## Statistics summary
+
+| Subset | Amendments | Articles | Rat2Rev queries | Rev2Rev queries | Period | License |
+| --- | --- | --- | --- | --- | --- | --- |
+| ReCaRe-EN (EU) | 340 | 91,361 | 340 | 1,509 | 2010–2025 | CC BY 4.0 |
+| ReCaRe-JA (JP) | 363 | 90,170 | 363 | 1,653 | 2019–2025 | CC BY 4.0 / CC0 |
+| **Total** | **703** | **181,531** | **703** | **3,162** | — | — |
+
+## Limitations
+
+- **Annotator-construction overlap.** The relevance-label validity check
+  (sampled 200 q-d pairs, 2 raters, blind) is performed by authors of the
+  construction pipeline. Future work: third-party annotation.
+- **Coverage windows differ.** EU period is 2010–2025; Japanese period is
+  2019–2025. This is a function of upstream availability and is documented
+  in the paper.
+- **Construction code not released.** The alignment pipeline depends on
+  internal scrapers (EUR-Lex / e-Gov) that are not part of the release.
+  The dataset itself, however, is fully public so that retrieval-side
+  experiments are completely reproducible.
+- **Implicit dependency in Rev2Rev.** Some co-revised article pairs share
+  no surface vocabulary; this is the intended difficulty of the task but
+  means lexical retrievers underperform.
+- **Not a substitute for legal advice.** The dataset is a research
+  resource. Retrieved articles are revision *candidates* for expert review.
+
+## Ethical considerations
+
+ReCaRe contains only **public legal text** (statutes and amending acts
+published by the European Union and the Japanese government). No personal
+data, no human-subjects data, and no proprietary content is included.
+References to officials or named parties in legal text appear because they
+are part of the public record and are not subject to research-purpose
+privacy obligations. The annotation step (label validity verification) is
+performed by the paper's co-authors on already-public legal articles, with
+no external participants, so the work is **not subject to IRB review**.
+See the CIKM 2026 paper §3.3 (Ethical Considerations) for the full
+treatment.
+
+## License & attribution
+
+- **EU subset**: data sourced from EUR-Lex under CC BY 4.0 (EUR-Lex
+  re-use notice). Attribution: European Union, EUR-Lex.
+- **Japanese subset**: data sourced from e-Gov 法令検索, 日本法令索引,
+  衆議院議案 — under CC BY 4.0 / CC0 1.0 as published. Attribution: 日本国
+  政府.
+- **ReCaRe construction & curation**: © the authors (Itō, Kurokawa, Kato,
+  Fujita), released under CC BY 4.0 to match the upstream EU subset.
+
+When using ReCaRe, please cite both the dataset DOI and the CIKM 2026
+paper.
+
+## Citation
+
+```bibtex
+@inproceedings{ito2026recare,
+  title     = {{ReCaRe}: A Bilingual Legal Benchmark for Revision Candidate Retrieval},
+  author    = {It\={o}, Hiroyoshi and Kurokawa, Y\={u}ma and Kato, Makoto P. and Fujita, Sumio},
+  booktitle = {Proceedings of the 35th ACM International Conference on Information and Knowledge Management},
+  series    = {CIKM '26},
+  year      = {2026},
+  publisher = {Association for Computing Machinery},
+  note      = {Resource Track. Dataset DOI: \url{https://doi.org/<HF DOI>} ; \url{https://huggingface.co/datasets/kasys/ReCaRe}},
+}
+```
 
-- **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)
+> The DOI placeholder will be replaced once HF Settings → Generate DOI is
+> run on tag `v0.1.0`. Camera-ready citation will list both the HF DOI and
+> the Zenodo mirror DOI.
 
-Source identifiers (CELEX numbers, 法令履歴 IDs) are preserved in `_id` fields so each record is traceable to its origin.
+## Contact
 
+- Hiroyoshi Itō (first author, Rev2Rev) — University of Tsukuba
+- Yūma Kurokawa (first author, Rat2Rev) — University of Tsukuba
+- Makoto P. Kato (corresponding author) — `mpkato@slis.tsukuba.ac.jp`
+- Sumio Fujita — LY Corporation