Dataset card and data licence: authors, versions, rules, baselines, citation

LICENSE-DATA.md

@@ -27,7 +27,7 @@ Please include the following when redistributing ParlIR data or derivatives:
 > Contains European Parliament written questions and official answers. Source: European Parliament Open Data
 > Portal (data.europarl.europa.eu), © European Union, 2014–2026, CC BY 4.0. Commission answers reused under
 > Decision 2011/833/EU (CC BY 4.0); Council answers reused under Decision (EU) 2017/1842. Texts were modified
-> (header removal, identifier masking). ParlIR annotations © TODO(authors), CC BY 4.0.
+> (header removal, identifier masking). ParlIR annotations © 2026 Marek Šuppa and Viktória Ondrejová, CC BY 4.0.
 
 ## Personal data
 
@@ -35,5 +35,6 @@ Author names are those of Members of the European Parliament acting in their off
 the Parliament. They are kept as metadata and removed from the canonical retrieval text where they appear in
 headers and co-signatory lines; residual mentions remain (e.g. 12 train queries with a "Question supported by …"
 list, and MEPs or other public figures named in question/answer content). `subject`, `header` and `raw_text` are
-unmasked. See docs/datasheet.md "Known issues in v1.0". A GDPR/legal-basis statement and a takedown contact are
-human-author decisions (docs/submission-checklist.md).
+unmasked. See docs/datasheet.md "Known issues in v1.0". Removal requests concerning personal data can be filed on
+the code repository's issue tracker (https://github.com/NaiveNeuron/ParlIR-ecir-2027/issues). They take effect in
+the next release.

README.md

@@ -63,126 +63,114 @@ configs:
 
 # ParlIR
 
-European Parliament written questions (terms 8–10, 2014–2026) paired with their official Commission/Council
-answers, released as an IR test collection. Code and full documentation:
-https://github.com/NaiveNeuron/ParlIR-ecir-2027 (private until the authors release it).
+ParlIR is a test collection built from European Parliament written questions (parliamentary terms 8–10,
+July 2014 – October 2026) and the official answers of the Commission, the High Representative and the Council.
+Relevance comes from the institutions' own links: each question is linked to its reply, and replies cite earlier
+answers. The qrels come from these records, and a sample of the referral citations was checked by a human.
+
+- Paper: *ParlIR: A Test Collection for Answer and Precedent Retrieval from European Parliament Written Questions*
+  (Marek Šuppa and Viktória Ondrejová, under review)
+- Code, pipeline, tutorial and datasheet: https://github.com/NaiveNeuron/ParlIR-ecir-2027
 
 ## Tasks
 
-All tasks have qrels files `qrels/<task>_<split>.tsv` (splits `train`, `dev`, `test`). PrecedentRet and AnswerRet
-are the co-headline tasks.
-
-- **AnswerRet** (`answer`, config `qrels-answer`): retrieve the official answer to a question from the full answer
-  corpus; binary known-item qrels; primary metric MRR@10.
-- **Lenient AnswerRet** (`answerlenient`): the official answer (rel 2) plus typed equivalent answers (rel 1):
-  `twin_question`, `duplicate_answer`, `near_duplicate_answer` (listed in `equivalents.jsonl`).
-- **PrecedentRet** (`precedent`, config `qrels-precedent`; *which earlier answers will its reply cite?*): retrieve earlier
-  answers that the official answer explicitly cross-references; eligible = answers published before the question
-  date (pre-submission, as-of ranking); metric: known-citation recall (R@10, R@100). `precedenthc` (config
-  `qrels-precedenthc`) is the referral-phrased subset.
-- **CiteRet** (`cite`, config `qrels-cite`; drafting-time variant): the same citation edges, eligible = answers
-  published before the official reply's date. `citehc` is the referral-phrased subset.
-
-**Example (test).** Question `P-9-2024-000684` (4 Mar 2024, asbestos screening proposal): its official reply
-`P-9-2024-000684-ASW` is a referral stub ("…would refer the Honourable Member to its answer to written question
-[REF]"), which is the AnswerRet target. The answer it cites, `E-9-2023-003699-ASW` (12 Mar 2024), is the CiteRet
-target; it is eligible at drafting time (before the reply) but not in PrecedentRet, the pre-submission task (after the question).
-
-## Files
-
-All files of a release live under `<version>/` (this card: `v1.0/`). Version tags are readable labels; pin downloads
-to the full commit hash, because tags can be deleted and recreated: `v1.0` (data) =
-`2176502d8ed76885c17fc85604be33d2cf429368`, `v1.0-textonly` = `925a8807a3b471beac0dc376826b2e0a086b0bc4`,
-`v1.0-evidence-r5` = `6b381010ef7b07927b30bd4c35bbb1d416f68baf`. The **evidence supplement** (every run with
-per-query metrics and meta, all LLM-audit records including the extended audit and pools, and the fine-tuned
-bi-encoder weights) is under `v1.0/experiments/` at `v1.0-evidence-r5` (earlier evidence tags lack later post-hoc runs and audit extensions: random-negative models, Qwen3, dense question→question, all-periods zero-shot runs). Embedding caches are not included; the pipeline rebuilds them.
-
-- `corpus.jsonl`: `_id` (`<QID>-ASW`), `text` (masked, canonical), `raw_text` (header-stripped, unmasked),
-  `date`, `institution`, `answers` (question ids), `header`, `source_url`.
-- `queries.jsonl`: `_id`, `subject`, `text` (masked subject + body), `raw_text`, `date`, `source_url`, `type`
-  (E/P), `term`, `addressee`, `author`, `orig_lang`, `orig_lang_src` (`csv`/`detail`), `answer`, `header`, `split`
-  (train/dev/test/null), `joint_size`.
-- `queries_xling.jsonl`: official non-English versions of questions: `_id`, `lang`, `text` (masked), `raw_text`.
-- `equivalents.jsonl`: `qid`, `equivalent_answers` (list of `{doc, type}`).
-- `precedent_meta.jsonl`: per query, citation edges (`cited`, `answer`, provenance `via` ⊆ {`hyperlink`, `url`,
-  `text`}, `referral` flag), with the eligible answers under the pre-submission and drafting cutoffs.
-- `qrels/<task>_<split>.tsv` for `answer`, `answerlenient`, `cite`, `citehc`, `precedent`, `precedenthc`: TREC
-  format `qid 0 docid rel` (whitespace-separated, no header), for `trec_eval`/`ir_measures`.
-- `qrels_jsonl/<task>_<split>.jsonl`: the same qrels as JSONL (`query_id`, `iteration`, `doc_id`, `relevance`);
-  these back the `qrels-*` dataset configs.
-- `build_manifest.json`: counts and sha256 checksums.
-- LLM pooled-audit pools (sensitivity analysis only; reproducible from the sample seed and the runs).
-
-**Masking.** Headers are removed structurally; every question/answer identifier in the masker's forms (including
-URL-embedded ids, hyphen variants, ids glued to a language code, and typo forms) is replaced by `[REF]`; term-8
-PE-number footers and co-signatory/supporter lines are removed. The validator checks for identifiers independently
-of the masker. Known v1.0 residue (to be fixed in the next release): rarer spellings survive in 255 queries (5 in test) and 48
-answers, none naming a dev/test citation target; `subject`, `header` and `raw_text` are unmasked (datasheet, "Known
-issues in v1.0").
-
-**Permitted inputs.** Systems may use only the `text` fields. Ids, dates, `header`, `subject`, `raw_text`,
-`source_url`, `answers` and the queries' `answer` field are forbidden: answer ids reuse the question id (that rule
-alone scores MRR@10 0.989) and a 0–180-day date window lifts BM25 from 0.547 to 0.760. At test time, dates may be used
-only for the as-of filters of the citation tasks; training may use training-split metadata (two post-hoc recipes
-select negatives by date).
-
-**Citation-task allowance.** For `precedent*` and `cite*`, a system may also use the archive's question→answer links
-of *eligible* answers (dated before the query's cutoff) and those earlier questions' `text`; the query's own `answer`
-link and anything after its cutoff stay forbidden (the as-of filter enforces the dates). Only the question→question
-baselines use this; all other baselines rank answer texts only. The text-only export does not contain these links,
-so question→question baselines cannot be rerun from it (use the v1.0 data, commit
-`2176502d8ed76885c17fc85604be33d2cf429368`).
-
-**Safe default for evaluation: the text-only export** (tag `v1.0-textonly`, `v1.0/textonly/`): corpus and queries with
-only a random opaque `_id` and the masked `text`; qrels on opaque ids; evaluator-only date files for the as-of rule in
-`eval/`; the opaque→source mapping as a separate download (`v1.0/textonly-provenance/`). Every main-experiment run scores
-identically on it.
-
-**Baseline protocol caveat.** Fine-tuning on the training split with BM25 hard negatives mined across the whole
-archive (the pre-registered protocol) hurts (MRR@10 0.424 ± 0.008 over 3 seeds vs 0.490 zero-shot). Post-hoc
-controls (3 seeds each): random negatives 0.669 ± 0.012, random negatives within ±365 days 0.712 ± 0.003, date-matched BM25
-negatives 0.637 (3 seeds). Why is not established; plausibly many mined negatives are relevant answers the qrels miss
-(LLM-judged test-query proxy: 18% "directly addresses", 51% "partially"), plus a date signal. Avoid naive
-hard-negative mining here; these results are post hoc, to be confirmed on the next test period. Best pre-registered baseline: RRF(BM25, BGE-M3), 0.588.
-Post-hoc zero-shot baseline: Qwen3-Embedding-8B (2025) reaches 0.714, with the same gain over BM25 (+0.17) on test
-questions answered after the model's release as before it. PrecedentRet test R@10: BM25 0.534, BGE-M3 0.585,
-Qwen3-Embedding-8B 0.632, BM25 question→question 0.712, BGE-M3 question→question 0.725. For CiteRet, dense
-question→question retrieval (BGE-M3) reaches R@10 0.744, on par with BM25 question→question (0.724).
-
-Splits are temporal (train 2014–2021, dev 2022–2023, test 2024–2026); 2015 questions are dated from their XML
-header.
-
-**Counts (v1.0).** Corpus 76,536 answers (Commission 70,397, VP/HR 4,181, Council 1,892, unknown 66); 78,314
-queries = 78,309 split-assigned (train 59,612, dev 7,735, test 10,962) plus 5 unsplit; `queries_xling.jsonl` 7,683 texts in 22 languages;
-equivalent answers for 1,396 queries (twin_question 419, duplicate_answer 852, near_duplicate_answer 296).
-Qrels (queries / pairs):
-
-| Task | train | dev | test |
+| Task | Query | Relevant documents | Metric |
+|---|---|---|---|
+| **AnswerRet** (`answer`) | a question | its official reply, among all 76,536 answers | MRR@10 |
+| **PrecedentRet** (`precedent`) | a new question | the earlier answers its reply cites, among answers published before the question | R@10, R@100 |
+| CiteRet (`cite`) | a question | the same citations, among answers published before the reply | R@10, R@100 |
+
+`precedenthc` and `citehc` keep only referral citations ("…refers the Honourable Member to its answer to…").
+`answerlenient` also accepts answers that are equivalent to the official one (`equivalents.jsonl`). Splits are
+by question date: train 2014–2021, dev 2022–2023, test 2024–2026.
+
+## Versions
+
+Pin downloads to the commit hash. Tags are readable labels, and the hash identifies the snapshot.
+
+| Tag | Contents | Commit |
+|---|---|---|
+| `v1.0` | data and qrels (`v1.0/`) | `2176502d8ed76885c17fc85604be33d2cf429368` |
+| `v1.0-textonly` | text-only evaluation export (`v1.0/textonly/`) | `925a8807a3b471beac0dc376826b2e0a086b0bc4` |
+| `v1.0-evidence-r5` | all runs, per-query metrics, audit records, fine-tuned weights (`v1.0/experiments/`, about 21 GB) | `6b381010ef7b07927b30bd4c35bbb1d416f68baf` |
+
+```python
+from datasets import load_dataset
+corpus = load_dataset("NaiveNeuron/parlir", "corpus", revision="2176502d8ed76885c17fc85604be33d2cf429368")
+```
+
+## Files (`v1.0/`)
+
+- `corpus.jsonl`: `_id` (`<QID>-ASW`), `text` (masked, use this), `raw_text` (unmasked), `date`, `institution`,
+  `answers`, `header`, `source_url`
+- `queries.jsonl`: `_id`, `subject`, `text` (masked, use this), `raw_text`, `date`, `source_url`, `type` (E/P),
+  `term`, `addressee`, `author`, `orig_lang`, `orig_lang_src`, `answer`, `header`, `split`, `joint_size`
+- `queries_xling.jsonl`: the official non-English versions of questions (7,683 texts, 22 languages)
+- `equivalents.jsonl`: equivalent answers per query (`twin_question`, `duplicate_answer`, `near_duplicate_answer`)
+- `precedent_meta.jsonl`: citation edges per query, with their source (`hyperlink`, `url`, `text`) and a
+  `referral` flag
+- `qrels/<task>_<split>.tsv`: TREC qrels (`qid 0 docid rel`), and `qrels_jsonl/` with the same qrels as JSONL
+- `build_manifest.json`: counts and SHA-256 checksums
+
+## Rules for systems
+
+Systems may use only the `text` fields. Do not use ids, dates, `header`, `subject`, `raw_text`, `source_url`,
+`answers` or the `answer` field as retrieval features, because some reveal the answer and others give shortcuts.
+An answer's id is its question's id plus `-ASW` (matching ids alone scores MRR@10 0.989), and a date window lifts
+BM25 from 0.547 to 0.760. Dates are used
+only by the evaluator's as-of filter for the citation tasks. For those tasks, systems may also use the links from
+earlier eligible answers to the questions they replied to, as the question-matching baselines do.
+
+The text-only export (`v1.0-textonly`) is the safest choice for evaluation. Corpus and queries carry only a random
+id and the masked text, the dates sit in evaluator-only files, and the id mapping is a separate download. All
+stored main-experiment runs score the same on it. It omits the links that question matching needs.
+
+## Baselines (test split)
+
+| System | AnswerRet MRR@10 | PrecedentRet R@10 |
+|---|---|---|
+| BM25 | 0.547 | 0.534 |
+| BGE-M3 | 0.525 | 0.585 |
+| Qwen3-Embedding-8B (zero-shot) | 0.714 | 0.632 |
+| BM25 question matching | n/a | 0.712 |
+| BGE-M3 question matching | n/a | 0.725 |
+
+PrecedentRet has 301 test queries, so please report confidence intervals. More results, including fusion,
+re-ranking, fine-tuning and a label audit with LLM judges, are in the paper and the evidence supplement.
+
+## Counts (v1.0)
+
+76,536 answers (Commission 70,397, High Representative 4,181, Council 1,892, unknown 66) and 78,309 questions with a
+split (train 59,612, dev 7,735, test 10,962).
+
+| Task | train (queries / pairs) | dev | test |
 |---|---|---|---|
 | `answer` | 59,612 / 59,612 | 7,735 / 7,735 | 10,962 / 10,962 |
 | `answerlenient` | 59,612 / 70,463 | 7,735 / 7,854 | 10,962 / 11,239 |
-| `cite` | 3,767 / 5,592 | 299 / 441 | 396 / 634 |
-| `citehc` | 2,991 / 4,199 | 186 / 245 | 256 / 352 |
 | `precedent` | 2,302 / 3,393 | 252 / 364 | 301 / 481 |
 | `precedenthc` | 1,788 / 2,462 | 159 / 208 | 192 / 269 |
+| `cite` | 3,767 / 5,592 | 299 / 441 | 396 / 634 |
+| `citehc` | 2,991 / 4,199 | 186 / 245 | 256 / 352 |
 
-PrecedentRet test has 301 queries (CiteRet 396), below the pre-registered 500 (H6 not met); report confidence
-intervals.
+Known issues of v1.0 (rare unmasked identifier spellings, a few odd dates) are listed in the datasheet. v1.0 stays
+frozen, and fixes go into the next release.
 
-## Licence and attribution
+## Licence
 
-Source data: European Parliament Open Data Portal, © European Union, CC BY 4.0; Commission answers under
-Decision 2011/833/EU (CC BY 4.0 via C(2019) 1655); Council answers under Decision (EU) 2017/1842. Texts were
-modified (header removal, identifier masking). ParlIR annotations: CC BY 4.0. Author names are MEPs acting in
-an official capacity; they are metadata only and not part of the retrieval text.
+Source texts: European Parliament Open Data Portal, © European Union, CC BY 4.0. Commission answers are reused
+under Decision 2011/833/EU (CC BY 4.0 via C(2019) 1655), Council answers under Decision (EU) 2017/1842. The texts
+were modified (headers removed, identifiers masked). ParlIR annotations: CC BY 4.0. See `LICENSE-DATA.md`.
+Questions are signed by MEPs acting as public office-holders. Removal requests can be filed on the code
+repository's issue tracker.
 
 ## Citation
 
 ```bibtex
-@inproceedings{parlir2027,
-  title     = {ParlIR: TODO(authors)},
-  author    = {TODO(authors)},
-  booktitle = {Advances in Information Retrieval (ECIR 2027)},
-  year      = {2027}
+@misc{suppa2026parlir,
+  title  = {{ParlIR}: A Test Collection for Answer and Precedent Retrieval from {European Parliament} Written Questions},
+  author = {{\v{S}}uppa, Marek and Ondrejov{\'a}, Vikt{\'o}ria},
+  year   = {2026},
+  note   = {Dataset: https://huggingface.co/datasets/NaiveNeuron/parlir},
+  url    = {https://github.com/NaiveNeuron/ParlIR-ecir-2027}
 }
 ```