mpkato commited on
Commit
4e04886
·
verified ·
1 Parent(s): 3b801e8

Restore HF dataset card structure + reconcile with paper §3

Browse files

Restores: configs YAML (HF viewer needs the 12 configs), full HF dataset card template sections (Direct/Out-of-Scope Use, Source Data Producers, Annotations, Bias/Risks, Glossary), authors list. Adds from paper §3: formal Rat2Rev/Rev2Rev definitions (D, A_e, D+_q), per-cell statistics table (split sizes + |D+_q| stats), construction methodology (Dice/Simpson thresholds, article-level corpus rule, deterministic qrels derivation), validation study results (200 stratified pairs, Cohen's κ = 0.368 overall, 0.461 Rat2Rev, 0.270 Rev2Rev), per-source license breakdown, BibTeX citation with DOI. Reverts the previous over-deletion of the configs YAML that broke the dataset viewer.

Files changed (1) hide show
  1. README.md +435 -165
README.md CHANGED
@@ -1,191 +1,277 @@
1
  ---
2
- license:
3
- - cc-by-4.0
4
- - cc0-1.0
5
  language:
6
- - en
7
- - ja
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8
  size_categories:
9
- - 100K<n<1M
 
 
10
  task_categories:
11
- - text-retrieval
12
- tags:
13
- - legal
14
- - multilingual
15
- - amendment
16
- - retrieval
17
- pretty_name: ReCaRe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
18
  ---
19
 
20
  # ReCaRe: A Bilingual Legal Benchmark for Revision Candidate Retrieval
21
 
22
- ReCaRe is a bilingual legal information-retrieval benchmark for **Revision
 
 
23
  Candidate Retrieval (RCR)**: given information about a change, retrieve and
24
  rank documents that are plausible candidates for expert inspection and
25
- possible revision.
26
-
27
- The benchmark is built from European Union law (EUR-Lex) and Japanese law
28
- (e-Gov), and contains **91,361 EU articles** and **90,170 Japanese
29
- articles** covering **703 amendment events** (340 EU + 363 JA) from
30
- 2010-2025 (EU) and 2019-2025 (JA).
31
 
32
  - **DOI**: [10.57967/hf/8642](https://doi.org/10.57967/hf/8642)
33
  - **Code**: [github.com/kasys-lab/ReCaRe](https://github.com/kasys-lab/ReCaRe)
34
 
35
- ## Tasks
36
-
37
- ReCaRe defines two complementary retrieval tasks. Let `D` be the
38
- article-level corpus, `q` a query, and `D⁺_q ⊂ D` its relevant articles.
39
-
40
- **Rat2Rev (Rationale-to-Revision)** — Given the **amendment rationale**
41
- `r_e` of an amendment event `e` as the query, retrieve the set of target
42
- articles `A_e` modified by that amendment: `D⁺_q = A_e`.
43
-
44
- **Rev2Rev (Revision-to-Revision)** — Given an **observed revision** of one
45
- target article `a ∈ A_e` (its pre- and post-revision versions) as the
46
- query, retrieve the **other** co-revised articles from the same amendment
47
- event: `D⁺_q = A_e \ {a}`.
48
-
49
- ## Statistics
50
-
51
- | | Rat2Rev EU | Rat2Rev JA | Rev2Rev EU | Rev2Rev JA |
52
- |---------------------------|-----------:|-----------:|-----------:|-----------:|
53
- | # queries (train) | 113 | 121 | 506 | 563 |
54
- | # queries (validation) | 114 | 121 | 503 | 555 |
55
- | # queries (test) | 113 | 121 | 503 | 535 |
56
- | # queries (total) | 340 | 363 | 1,512 | 1,653 |
57
- | # amendment events (total)| 340 | 363 | 340 | 363 |
58
- | queries per event (mean) | 1.00 | 1.00 | 4.45 | 4.55 |
59
- | \|D⁺_q\| mean (test) | 18.41 | 28.06 | 16.21 | 27.76 |
60
- | \|D⁺_q\| median (test) | 8 | 14 | 8 | 14 |
61
- | \|D⁺_q\| min / max (test) | 2 / 177| 2 / 219| 1 / 176| 1 / 222|
62
-
63
- | Corpus | # articles | Mean chars | Median chars |
64
- |------------|-----------:|-----------:|-------------:|
65
- | `corpus-en`| 91,361 | 1,485 | 757 |
66
- | `corpus-ja`| 90,170 | 380 | 219 |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
67
 
68
  Each amendment event with `|A_e| ≥ 2` provides up to 5 Rev2Rev queries by
69
  random target-article selection. Splits are constructed at amendment-event
70
  granularity into approximately equal thirds; all queries from the same
71
  event are assigned to the same split to prevent event-level leakage. EU
72
- and JA corpora are independently sourced and contain no overlapping
73
- amendment events.
74
-
75
- ## File structure
76
-
77
- ```
78
- ReCaRe/
79
- ├── corpus-{en,ja}/corpus.jsonl # article-level corpus
80
- ├── queries-{rat2rev,rev2rev}-{en,ja}/queries.jsonl # queries per task × lang
81
- ├── qrels-{rat2rev,rev2rev}-{en,ja}/{train,validation,test}.jsonl
82
- └── metadata-{en,ja}/dataset.jsonl # full revision metadata
83
- ```
84
 
85
- Schemas:
86
-
87
- | File | Per-line fields |
88
- |---|---|
89
- | `corpus.jsonl` | `_id` (article ID), `text` |
90
- | `queries.jsonl` | `_id` (query ID), `text` |
91
- | `qrels/*.jsonl` | `query-id`, `corpus-id`, `score` (binary, score ≥ 1 = positive) |
92
- | `metadata-{lang}/dataset.jsonl` | one row per pre-/post-revision article pair: `amendment_law_id`, `amendment_reason`, `law_id`, `type_of_change`, `law_title_before/after`, `revision_id_before/after`, `article_id_before/after`, `article_number_before/after`, `caption_before/after`, `text_before/after` |
93
-
94
- ## Construction
95
-
96
- ReCaRe is constructed from **amendment events**, each comprising (1) an
97
- **amending act**, (2) an **amendment rationale** stating the motivation, (3)
98
- the **amended acts** modified by the amending act, and (4) the **target
99
- articles** (pre-revision versions of articles modified by the amending act).
100
-
101
- - **EU** (EUR-Lex): English Legislative Acts published 2010-2025. The
102
- rationale is taken from the corresponding COM document sections
103
- describing the proposal motivation (e.g. "Reasons for and objectives of
104
- the proposal", "General context").
105
- - **JA** (e-Gov): Japanese laws and their amending acts from 2019-2025.
106
- Rationales are extracted from the "Reasons" section of the corresponding
107
- bill text posted on the House of Representatives' page.
108
-
109
- **Article alignment**: pre-/post-revision versions of an amended act are
110
- split into articles (number, caption, body) and aligned via Dice
111
- similarity (word-level for EU, character-level for JA, threshold ≥ 0.7),
112
- falling back to Simpson similarity ≥ 0.95 for articles longer than 40
113
- characters. Article-number matches are used only when no article-number
114
- movement is detected within the same act. Typographical corrections,
115
- cross-reference renumbering, and pairs with Dice ≥ 0.99 are discarded.
116
- Newly added articles (no pre-revision version) are excluded.
117
-
118
- **Article-level corpus**: for each amended act, we use the version
119
- immediately before its earliest amendment included in our collection
120
- window; for acts not amended during the window we use the version in force
121
- at the start of the window (2010 for EU, 2019 for JA). This gives each
122
- article a single version.
123
-
124
- ## Validation
125
-
126
- Quality of relevant articles was verified on a stratified sample of **200
127
- query-article pairs** (50 per task-language combination: 25 relevant + 12
128
- hard negatives sampled from BM25 top-100 excluding target articles + 13
129
- random negatives sampled from the remaining corpus). Two authors judged
130
- every pair under full label blinding (positive/negative status, hard /
131
- random-negative origin, and amendment provenance were hidden).
132
-
133
- - **Overall agreement**: 72.5%
134
- - **Cohen's κ (overall)**: 0.368
135
- - **Cohen's κ (Rat2Rev)**: 0.461
136
- - **Cohen's κ (Rev2Rev)**: 0.270
137
 
138
- Agreement is higher for Rat2Rev than for Rev2Rev, consistent with the task
139
- definitions: Rev2Rev often requires judging whether two articles are
140
- related through an implicit revision relationship rather than through
141
- direct lexical evidence.
142
-
143
- ## License
144
-
145
- ReCaRe is built from publicly available legal documents and reused under
146
- the licenses of the original sources:
147
 
148
- - **EU content** (corpus-en, metadata-en, Rat2Rev rationales, Rev2Rev
149
- articles): reused under the [EUR-Lex legal notice](https://eur-lex.europa.eu/content/legal-notice/legal-notice.html),
150
- i.e. **CC BY 4.0** for EU-owned consolidated texts and **CC0 1.0** for
151
- metadata.
152
- - **Japanese statutory articles** (corpus-ja, metadata-ja, Rev2Rev
153
- articles): reused from [e-Gov](https://laws.e-gov.go.jp/) under its
154
- **CC BY 4.0** terms.
155
- - **Japanese bill texts** (used to extract Rat2Rev rationales): legislative
156
- bill texts placed in the legislative process; reuse relies, where
157
- applicable, on **Article 13 of the Japanese Copyright Act** (texts of
158
- laws, ordinances, notifications and the like enacted by national or
159
- local government bodies are not the subject of copyright).
160
 
161
- The dataset may contain names of public officials and administrative
162
- bodies as part of the legislative record, but does not include personal
163
- data newly collected from research participants.
 
164
 
165
- ## Intended use and limitations
166
 
167
- ReCaRe is a benchmark for research on retrieval, document update, and
168
- legal information access. Retrieved candidates represent **potential
169
- revision targets for human expert review**, not authoritative
170
- determinations. The dataset is not suitable for automated legal drafting,
171
- legal advice, or production deployment without expert oversight. Findings
172
- should not be extrapolated beyond the EU and Japanese legal jurisdictions
173
- represented.
174
 
175
- ## Loading
 
 
 
176
 
177
- Using the Hugging Face Hub directly:
178
 
179
  ```python
180
  from huggingface_hub import hf_hub_download
181
  import json
182
 
183
  def load_jsonl(filename):
184
- path = hf_hub_download(
185
- repo_id="kasys/ReCaRe",
186
- filename=filename,
187
- repo_type="dataset",
188
- )
189
  with open(path, "r", encoding="utf-8") as f:
190
  return [json.loads(line) for line in f]
191
 
@@ -194,20 +280,204 @@ queries = load_jsonl("queries-rat2rev-en/queries.jsonl")
194
  qrels = load_jsonl("qrels-rat2rev-en/test.jsonl")
195
  ```
196
 
197
- Or using the `datasets` library:
198
-
199
- ```python
200
- from datasets import load_dataset
201
-
202
- corpus = load_dataset("kasys/ReCaRe", data_files="corpus-en/corpus.jsonl", split="train")
203
- queries = load_dataset("kasys/ReCaRe", data_files="queries-rat2rev-en/queries.jsonl", split="train")
204
- qrels = load_dataset("kasys/ReCaRe", data_files="qrels-rat2rev-en/test.jsonl", split="train")
205
- ```
206
-
207
  Reference baselines (BM25, dense retrievers, cross-encoder rerankers,
208
  RankGPT, query/document expansion, and dense-retriever domain adaptation)
209
  are available in [github.com/kasys-lab/ReCaRe](https://github.com/kasys-lab/ReCaRe).
210
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
211
  ## Citation
212
 
213
  ```bibtex
 
1
  ---
2
+ # For reference on dataset card metadata, see the spec: https://github.com/huggingface/hub-docs/blob/main/datasetcard.md?plain=1
3
+ # Doc / guide: https://huggingface.co/docs/hub/datasets-cards
 
4
  language:
5
+ - en
6
+ - ja
7
+ license:
8
+ - cc-by-4.0
9
+ - cc0-1.0
10
+ tags:
11
+ - legal
12
+ - law
13
+ - multilingual
14
+ - bilingual
15
+ - retrieval
16
+ - benchmark
17
+ - eu-law
18
+ - japanese-law
19
+ language_creators:
20
+ - expert-generated
21
+ pretty_name: "ReCaRe: A Bilingual Legal Benchmark for Revision Candidate Retrieval"
22
  size_categories:
23
+ - 100K<n<1M
24
+ source_datasets:
25
+ - original
26
  task_categories:
27
+ - text-retrieval
28
+ configs:
29
+ - config_name: corpus-en
30
+ data_files:
31
+ - split: corpus
32
+ path: corpus-en/corpus.jsonl
33
+ - config_name: corpus-ja
34
+ data_files:
35
+ - split: corpus
36
+ path: corpus-ja/corpus.jsonl
37
+ - config_name: queries-rat2rev-en
38
+ data_files:
39
+ - split: queries
40
+ path: queries-rat2rev-en/queries.jsonl
41
+ - config_name: queries-rat2rev-ja
42
+ data_files:
43
+ - split: queries
44
+ path: queries-rat2rev-ja/queries.jsonl
45
+ - config_name: queries-rev2rev-en
46
+ data_files:
47
+ - split: queries
48
+ path: queries-rev2rev-en/queries.jsonl
49
+ - config_name: queries-rev2rev-ja
50
+ data_files:
51
+ - split: queries
52
+ path: queries-rev2rev-ja/queries.jsonl
53
+ - config_name: qrels-rat2rev-en
54
+ data_files:
55
+ - split: train
56
+ path: qrels-rat2rev-en/train.jsonl
57
+ - split: validation
58
+ path: qrels-rat2rev-en/validation.jsonl
59
+ - split: test
60
+ path: qrels-rat2rev-en/test.jsonl
61
+ - config_name: qrels-rat2rev-ja
62
+ data_files:
63
+ - split: train
64
+ path: qrels-rat2rev-ja/train.jsonl
65
+ - split: validation
66
+ path: qrels-rat2rev-ja/validation.jsonl
67
+ - split: test
68
+ path: qrels-rat2rev-ja/test.jsonl
69
+ - config_name: qrels-rev2rev-en
70
+ data_files:
71
+ - split: train
72
+ path: qrels-rev2rev-en/train.jsonl
73
+ - split: validation
74
+ path: qrels-rev2rev-en/validation.jsonl
75
+ - split: test
76
+ path: qrels-rev2rev-en/test.jsonl
77
+ - config_name: qrels-rev2rev-ja
78
+ data_files:
79
+ - split: train
80
+ path: qrels-rev2rev-ja/train.jsonl
81
+ - split: validation
82
+ path: qrels-rev2rev-ja/validation.jsonl
83
+ - split: test
84
+ path: qrels-rev2rev-ja/test.jsonl
85
+ - config_name: metadata-en
86
+ data_files:
87
+ - split: metadata
88
+ path: metadata-en/dataset.jsonl
89
+ - config_name: metadata-ja
90
+ data_files:
91
+ - split: metadata
92
+ path: metadata-ja/dataset.jsonl
93
  ---
94
 
95
  # ReCaRe: A Bilingual Legal Benchmark for Revision Candidate Retrieval
96
 
97
+ <!-- Provide a quick summary of the dataset. -->
98
+
99
+ ReCaRe (pronounced "re-care") is a bilingual legal benchmark for **Revision
100
  Candidate Retrieval (RCR)**: given information about a change, retrieve and
101
  rank documents that are plausible candidates for expert inspection and
102
+ possible revision. It spans European Union law (EUR-Lex, English) and
103
+ Japanese law (e-Gov, Japanese), with **703 amendment events (340 EU + 363
104
+ JA)** and **181,531 articles (91,361 EU + 90,170 JA)**, supporting two
105
+ retrieval tasks (Rat2Rev and Rev2Rev) over bilingual corpora.
 
 
106
 
107
  - **DOI**: [10.57967/hf/8642](https://doi.org/10.57967/hf/8642)
108
  - **Code**: [github.com/kasys-lab/ReCaRe](https://github.com/kasys-lab/ReCaRe)
109
 
110
+ ## Dataset Details
111
+
112
+ ### Dataset Description
113
+
114
+ <!-- Provide a longer summary of what this dataset is. -->
115
+
116
+ Document corpora in regulated domains evolve: statutes are amended,
117
+ internal policies revised, software specifications updated. Yet most
118
+ information retrieval research has framed retrieval as a one-shot
119
+ question-answering problem over a frozen corpus, leaving the IR aspects
120
+ of *document maintenance* — finding which documents need to change, and
121
+ which other documents must change with them — comparatively underexplored.
122
+
123
+ ReCaRe formalizes two complementary RCR tasks over bilingual corpora of EU
124
+ and Japanese law. Let `D` denote the article-level corpus, `q` a query, and
125
+ `D⁺_q ⊂ D` the set of relevant articles for `q`. For each amendment event
126
+ `e`, let `r_e` denote its amendment rationale and let `A_e ⊂ D` denote the
127
+ set of target articles (the pre-revision versions of articles modified by
128
+ the amending act of `e`).
129
+
130
+ - **Rat2Rev (Rationale-to-Revision).** Given the amendment rationale `r_e`
131
+ as the query `q`, retrieve the target articles from `D`, where
132
+ `D⁺_q = A_e`.
133
+ - **Rev2Rev (Revision-to-Revision).** Given an observed revision of a
134
+ target article `a ∈ A_e` as the query `q` — represented by its pre- and
135
+ post-revision versions — retrieve the other target articles in the same
136
+ amendment event from `D \ {a}`, where `D⁺_q = A_e \ {a}`.
137
+
138
+ Properties:
139
+
140
+ - **Curated by:** Takumi Ito, Yuma Kurokawa (University of Tsukuba),
141
+ Makoto P. Kato (University of Tsukuba / National Institute of
142
+ Informatics), Sumio Fujita (LY Corporation).
143
+ - **Language(s):** English (`en`, EU subset) and Japanese (`ja`, Japanese
144
+ subset).
145
+ - **License:** CC BY 4.0 (EU-owned consolidated texts, Japanese e-Gov
146
+ consolidated statutes) / CC0 1.0 (metadata from EUR-Lex). See
147
+ [Licensing Information](#licensing-information) for per-source detail.
148
+
149
+ ## Uses
150
+
151
+ <!-- Address questions around how the dataset is intended to be used. -->
152
+
153
+ ### Direct Use
154
+
155
+ <!-- This section describes suitable use cases for the dataset. -->
156
+
157
+ ReCaRe is intended as a research benchmark for:
158
+
159
+ - Training and evaluating retrieval models on legal text where queries
160
+ and target documents differ markedly in length and register.
161
+ - Studying *document maintenance* retrieval: surfacing revision candidates
162
+ for expert review, distinct from question-answering or paragraph-level
163
+ retrieval.
164
+ - Comparative benchmarking against general-domain IR resources (e.g. BEIR)
165
+ to characterise the difficulty of low-overlap, multi-target retrieval
166
+ with implicit dependency.
167
+ - Ablation of long-context vs. short-context retrieval models, given that
168
+ Rat2Rev queries are substantially longer than typical IR queries.
169
+
170
+ ### Out-of-Scope Use
171
+
172
+ <!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->
173
+
174
+ ReCaRe is **not** designed for, and should not be used as, a basis for:
175
+
176
+ - Automated legal drafting, automated revision recommendation, or any other
177
+ production legal workflow without expert review. Retrieved articles are
178
+ *candidates* for expert review, not authoritative outputs.
179
+ - Provision of legal advice to end-users.
180
+ - Inference of personal or demographic information about individuals — the
181
+ data contains only public legal text, but personal names appearing in
182
+ legal records (e.g. drafters, ministers) should not be used to build
183
+ profiles of those individuals.
184
+ - Generalisation claims about legal systems other than the EU and Japan,
185
+ or about time periods outside the included windows (EU 2010-2025; JP
186
+ 2019-2025).
187
+
188
+ ## Dataset Structure
189
+
190
+ <!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->
191
+
192
+ The dataset is organised as **12 configs** sharing two language-aligned
193
+ corpora:
194
+
195
+ | Config | Splits | Schema | Records |
196
+ | --- | --- | --- | --- |
197
+ | `corpus-en` | `corpus` | `{_id, text}` | 91,361 |
198
+ | `corpus-ja` | `corpus` | `{_id, text}` | 90,170 |
199
+ | `queries-rat2rev-en` | `queries` | `{_id, text}` | 340 |
200
+ | `queries-rat2rev-ja` | `queries` | `{_id, text}` | 363 |
201
+ | `queries-rev2rev-en` | `queries` | `{_id, text}` | 1,509 |
202
+ | `queries-rev2rev-ja` | `queries` | `{_id, text}` | 1,653 |
203
+ | `qrels-rat2rev-en` | `train` / `validation` / `test` | `{query-id, corpus-id, score}` | 2,063 / 1,948 / 2,080 |
204
+ | `qrels-rat2rev-ja` | same | same | 3,228 / 2,501 / 3,395 |
205
+ | `qrels-rev2rev-en` | same | same | 12,088 / 8,189 / 8,156 |
206
+ | `qrels-rev2rev-ja` | same | same | 15,054 / 13,591 / 14,853 |
207
+ | `metadata-en` | `metadata` | 16-field amendment metadata (see below) | 91,361 |
208
+ | `metadata-ja` | `metadata` | 17-field amendment metadata (see below) | 90,170 |
209
+
210
+ All files are JSONL with one JSON object per line. The schemas follow BEIR
211
+ conventions (`_id`, `text`, `query-id`, `corpus-id`, `score`) so that
212
+ existing tooling such as `ir_measures`, `pytrec_eval`, and Pyserini works
213
+ without adapter code.
214
+
215
+ ### Query and qrel statistics
216
+
217
+ | Per (task, lang) | Rat2Rev EU | Rat2Rev JA | Rev2Rev EU | Rev2Rev JA |
218
+ |-----------------------------|-----------:|-----------:|-----------:|-----------:|
219
+ | # queries (train) | 113 | 121 | 506 | 563 |
220
+ | # queries (validation) | 114 | 121 | 503 | 555 |
221
+ | # queries (test) | 113 | 121 | 503 | 535 |
222
+ | # amendment events (total) | 340 | 363 | 340 | 363 |
223
+ | queries per event (mean) | 1.00 | 1.00 | 4.45 | 4.55 |
224
+ | \|D⁺_q\| mean (test) | 18.41 | 28.06 | 16.21 | 27.76 |
225
+ | \|D⁺_q\| median (test) | 8 | 14 | 8 | 14 |
226
+ | \|D⁺_q\| min / max (test) | 2 / 177| 2 / 219| 1 / 176| 1 / 222|
227
 
228
  Each amendment event with `|A_e| ≥ 2` provides up to 5 Rev2Rev queries by
229
  random target-article selection. Splits are constructed at amendment-event
230
  granularity into approximately equal thirds; all queries from the same
231
  event are assigned to the same split to prevent event-level leakage. EU
232
+ and JA amendment events are disjoint by construction (independently
233
+ sourced corpora) and cross-split overlap is zero.
 
 
 
 
 
 
 
 
 
 
234
 
235
+ ### Metadata schema
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
236
 
237
+ **`metadata-en`** (16 fields) and **`metadata-ja`** (17 fields, additionally
238
+ including `amendment_reason_url`) carry the full revision metadata, one
239
+ row per article version: `amendment_law_id`, `amendment_reason`,
240
+ `amendment_reason_url` (JA only), `law_id`, `type_of_change`,
241
+ `law_title_before`, `revision_id_before`, `article_id_before`,
242
+ `article_number_before`, `caption_before`, `text_before`,
243
+ `law_title_after`, `revision_id_after`, `article_id_after`,
244
+ `article_number_after`, `caption_after`, `text_after`.
 
245
 
246
+ Records with `amendment_law_id == null` are corpus-only articles (≈ 93-94 %
247
+ of metadata rows); records with non-null `amendment_law_id` are revisions
248
+ traceable to a specific amending act. The metadata configs are provided
249
+ for downstream provenance analyses and are not required to run the
250
+ retrieval tasks themselves.
 
 
 
 
 
 
 
251
 
252
+ **Splits.** The `qrels-*` configs use train / validation / test splits.
253
+ The `corpus-*`, `queries-*`, and `metadata-*` configs are not split — the
254
+ full corpus is used at retrieval time, and queries split membership is
255
+ encoded through the qrels.
256
 
257
+ ### Quick start
258
 
259
+ ```python
260
+ from datasets import load_dataset
 
 
 
 
 
261
 
262
+ corpus = load_dataset("kasys/ReCaRe", "corpus-en", split="corpus")
263
+ queries = load_dataset("kasys/ReCaRe", "queries-rat2rev-en", split="queries")
264
+ qrels = load_dataset("kasys/ReCaRe", "qrels-rat2rev-en") # train/validation/test
265
+ ```
266
 
267
+ Or using the Hugging Face Hub directly:
268
 
269
  ```python
270
  from huggingface_hub import hf_hub_download
271
  import json
272
 
273
  def load_jsonl(filename):
274
+ path = hf_hub_download(repo_id="kasys/ReCaRe", filename=filename, repo_type="dataset")
 
 
 
 
275
  with open(path, "r", encoding="utf-8") as f:
276
  return [json.loads(line) for line in f]
277
 
 
280
  qrels = load_jsonl("qrels-rat2rev-en/test.jsonl")
281
  ```
282
 
 
 
 
 
 
 
 
 
 
 
283
  Reference baselines (BM25, dense retrievers, cross-encoder rerankers,
284
  RankGPT, query/document expansion, and dense-retriever domain adaptation)
285
  are available in [github.com/kasys-lab/ReCaRe](https://github.com/kasys-lab/ReCaRe).
286
 
287
+ ## Dataset Creation
288
+
289
+ ### Curation Rationale
290
+
291
+ <!-- Motivation for the creation of this dataset. -->
292
+
293
+ Existing legal-IR benchmarks (e.g. COLIEE, BSARD, LeCaRD, STARD,
294
+ LegalBench-RAG) cover statutory or case-law question answering but not
295
+ *document maintenance* — finding which provisions must change when a
296
+ corpus evolves. ReCaRe was constructed to give the IR community a shared,
297
+ bilingual benchmark for two practically motivated retrieval tasks
298
+ (Rat2Rev, Rev2Rev) over legal corpora, in two jurisdictions whose
299
+ legal-revision practices differ (EU consolidation tradition vs. Japanese
300
+ amending-act tradition).
301
+
302
+ ### Source Data
303
+
304
+ <!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->
305
+
306
+ The dataset contains **only public legal text** drawn from official portals:
307
+
308
+ - **EU subset:** [EUR-Lex](https://eur-lex.europa.eu/) — English Legislative
309
+ Acts (Regulations, Directives, Decisions) published between 2010 and
310
+ 2025. The amendment rationale is taken from sections of the
311
+ corresponding COM document (the originating European Commission
312
+ proposal) that describe the proposal motivation, such as
313
+ "Reasons for and objectives of the proposal" and "General context".
314
+ - **Japanese subset:** [e-Gov 法令検索](https://laws.e-gov.go.jp/) —
315
+ amending acts and amended acts from 2019-2025. For each amending act,
316
+ the corresponding bill is identified via the [Japanese Laws and
317
+ Regulations Index](https://hourei.ndl.go.jp/) and the bill text is
318
+ collected from the
319
+ [House of Representatives' bill page](https://www.shugiin.go.jp/); its
320
+ "Reasons" section is used as the amendment rationale.
321
+
322
+ #### Data Collection and Processing
323
+
324
+ <!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->
325
+
326
+ Acquisition spanned 2025-Q3 to 2026-Q1.
327
+
328
+ **Amendment events.** Each amendment event comprises (1) an amending act,
329
+ (2) an amendment rationale, (3) the amended acts modified by the amending
330
+ act, and (4) the target articles (pre-revision versions of articles
331
+ modified by the amending act).
332
+
333
+ **Article alignment.** Each version of an amended act is split into
334
+ articles represented by article number, caption, and body. Unchanged
335
+ articles and articles with unchanged body text are removed. Pre- and
336
+ post-revision articles are then aligned using:
337
+
338
+ - **Body similarity by Dice coefficient** (word-level tokens for EU,
339
+ character-level tokens for JA) with threshold ≥ 0.7, greedily accepted
340
+ as one-to-one alignments.
341
+ - For lower-Dice pairs, the **Simpson coefficient** is used to retain
342
+ cases where one version preserves a substantial part of the other,
343
+ accepting pairs with Simpson ≥ 0.95 and article length above 40
344
+ characters.
345
+ - **Article-number matches** are used only when no article-number movement
346
+ is detected within the same act.
347
+
348
+ Thresholds and filtering rules were tuned through manual inspection,
349
+ prioritising the avoidance of incorrect alignments. Minor edits (typo
350
+ corrections, cross-reference renumbering) and near-identical pairs
351
+ (Dice ≥ 0.99) are discarded. Newly added articles are excluded because
352
+ they have no pre-revision version. The pre-revision articles in the
353
+ remaining aligned pairs constitute the **target articles**.
354
+
355
+ **Article-level corpus.** Each article in the corpus has a single fixed
356
+ version. For each amended act, we use the version immediately before its
357
+ earliest amendment included in the collection window; for acts not amended
358
+ during the window we use the version in force at the start of the window
359
+ (2010 for EU, 2019 for JA). If an article is amended multiple times,
360
+ only the earliest amendment is retained, because later amendments would
361
+ refer to a pre-revision version that differs from the corpus version.
362
+
363
+ **Queries and qrels** are then derived deterministically from this
364
+ alignment:
365
+
366
+ - **Rat2Rev queries** are the amendment rationale of each amending act
367
+ (one query per amendment).
368
+ - **Rev2Rev queries** are individual revised articles (up to five queries
369
+ per amendment, randomly sampled from `A_e` when `|A_e| > 5`).
370
+ - A `(rationale, article)` pair is positive iff the article is a target
371
+ of that amendment.
372
+ - A `(revised-article, article)` pair is positive iff both are targets of
373
+ the same amendment event.
374
+
375
+ #### Who are the source data producers?
376
+
377
+ <!-- This section describes the people or systems who originally created the data. -->
378
+
379
+ The original legal text was authored by the **European Union legislative
380
+ bodies** (Council, Parliament, Commission, etc.) and by **the Government
381
+ of Japan** (Diet, ministries, etc.), and is published in their respective
382
+ official portals as public law. Producers are state-level institutions
383
+ acting in a public legislative or regulatory capacity; no individual
384
+ authorship metadata is included in the dataset beyond what appears in the
385
+ text of the law itself.
386
+
387
+ ### Annotations
388
+
389
+ <!-- If the dataset contains annotations which are not part of the initial data collection, use this section to describe them. -->
390
+
391
+ #### Annotation process
392
+
393
+ <!-- This section describes the annotation process such as annotation tools used in the process, the amount of data annotated, annotation guidelines provided to the annotators, interannotator statistics, annotation validation, etc. -->
394
+
395
+ Relevance labels (qrels) are produced **automatically** from the official
396
+ amendment alignment described above; there is no per-pair human labelling
397
+ in the released qrels.
398
+
399
+ To verify the quality of relevant articles, we conducted a manual
400
+ **validation study** on a stratified sample of **200 query-article pairs**
401
+ (50 per task-language combination). Each combination contained 25
402
+ relevant articles, 12 hard negatives sampled from BM25 top-100 results
403
+ excluding target articles, and 13 random negatives sampled from the
404
+ remaining corpus. Two authors judged every pair under full label blinding
405
+ (positive/negative status, hard/random-negative origin, and amendment
406
+ provenance were hidden).
407
+
408
+ | | Overall | Rat2Rev | Rev2Rev |
409
+ |---|---:|---:|---:|
410
+ | Raw agreement | 72.5 % | — | — |
411
+ | Cohen's κ | 0.368 | 0.461 | 0.270 |
412
+
413
+ Agreement is higher for Rat2Rev than for Rev2Rev, consistent with the task
414
+ definitions: Rev2Rev often requires judging whether two articles are
415
+ related through an implicit revision relationship rather than through
416
+ direct lexical evidence.
417
+
418
+ #### Personal and Sensitive Information
419
+
420
+ <!-- State whether the dataset contains data that might be considered personal, sensitive, or private (e.g., data that reveals addresses, uniquely identifiable names or aliases, racial or ethnic origins, sexual orientations, religious beliefs, political opinions, financial or health data, etc.). If efforts were made to anonymize the data, describe the anonymization process. -->
421
+
422
+ The dataset contains **only public legal text** released by the European
423
+ Union and the Japanese government. The dataset may contain names of
424
+ public officials and administrative bodies as part of the legislative
425
+ record, but does not include personal data newly collected from research
426
+ participants. References to officials, ministers, or other named parties
427
+ appear because they are part of the public legislative record; they are
428
+ not subject to research-purpose privacy obligations and have not been
429
+ anonymised.
430
+
431
+ ## Bias, Risks, and Limitations
432
+
433
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
434
+
435
+ - **Legal scope.** The dataset is grounded in EU and Japanese law and
436
+ inherits the conventions, drafting traditions, and any latent biases of
437
+ those legal systems. Conclusions should not be transferred to other
438
+ jurisdictions or to non-statutory legal text without further validation.
439
+ - **Not a substitute for legal advice.** The dataset is a research
440
+ resource. Retrieved articles are revision *candidates* for expert review.
441
+ - **Automatic qrels.** Relevance labels are derived from official
442
+ amendment alignment, not from per-pair human judgment. The validation
443
+ study above quantifies inter-annotator agreement on a stratified sample
444
+ but does not exhaustively re-label the qrels.
445
+
446
+ ## Licensing Information
447
+
448
+ ReCaRe is built from publicly available legal documents and reused under
449
+ the licenses of the original sources:
450
+
451
+ - **EU content** (corpus-en, metadata-en, Rat2Rev rationales, Rev2Rev
452
+ articles): reused under the [EUR-Lex legal notice](https://eur-lex.europa.eu/content/legal-notice/legal-notice.html),
453
+ i.e. **CC BY 4.0** for EU-owned consolidated texts and **CC0 1.0** for
454
+ metadata.
455
+ - **Japanese statutory articles** (corpus-ja, metadata-ja, Rev2Rev
456
+ articles): reused from e-Gov under its **CC BY 4.0** terms.
457
+ - **Japanese bill texts** (used to extract Rat2Rev rationales):
458
+ legislative bill texts placed in the legislative process; reuse relies,
459
+ where applicable, on **Article 13 of the Japanese Copyright Act** (texts
460
+ of laws, ordinances, notifications and the like enacted by national or
461
+ local government bodies are not the subject of copyright).
462
+
463
+ ## Glossary
464
+
465
+ <!-- If relevant, include terms and calculations in this section that can help readers understand the dataset or dataset card. -->
466
+
467
+ - **RCR (Revision Candidate Retrieval).** The retrieval task of locating
468
+ documents that constitute plausible candidates for an authoritative
469
+ revision.
470
+ - **Rat2Rev (Rationale-to-Revision Retrieval).** Given an amendment's
471
+ textual rationale, retrieve the concrete articles to be modified.
472
+ - **Rev2Rev (Revision-to-Revision Retrieval).** Given an article to be
473
+ revised, retrieve the other articles revised in the same amendment
474
+ event.
475
+ - **Amendment event.** One amending act (a single Regulation, Directive,
476
+ or domestic amending law) that modifies one or more prior provisions.
477
+ - **`A_e`.** Set of target articles of amendment event `e` (pre-revision
478
+ versions of articles modified by the amending act).
479
+ - **`D⁺_q`.** Set of relevant articles for query `q`.
480
+
481
  ## Citation
482
 
483
  ```bibtex