Add dataset card

README.md

@@ -0,0 +1,146 @@
+---
+pretty_name: Document-Level Word Alignment
+license: other
+license_name: mixed-per-subset
+license_link: LICENSE.md
+language:
+  - en
+  - fr
+  - ro
+  - ja
+  - zh
+  - grc
+  - la
+  - pt
+  - cs
+multilinguality: multilingual
+task_categories:
+  - token-classification
+task_ids: []
+size_categories:
+  - 1K<n<10K
+configs:
+  - config_name: en-fr
+    data_files:
+      - split: dev
+        path: documents/en-fr/dev.json
+      - split: test
+        path: documents/en-fr/test.json
+  - config_name: en-ro
+    data_files:
+      - split: dev
+        path: documents/en-ro/dev.json
+      - split: test
+        path: documents/en-ro/test.json
+  - config_name: en-ja
+    data_files:
+      - split: dev
+        path: documents/en-ja/dev.json
+      - split: test
+        path: documents/en-ja/test.json
+  - config_name: en-zh
+    data_files:
+      - split: dev
+        path: documents/en-zh/dev.json
+      - split: test
+        path: documents/en-zh/test.json
+  - config_name: la-gr
+    data_files:
+      - split: dev
+        path: documents/la-gr/dev.json
+      - split: test
+        path: documents/la-gr/test.json
+  - config_name: en-cz
+    data_files:
+      - split: test
+        path: documents/en-cz/test.json
+---
+
+# Document-Level Word Alignment
+
+Document-level word alignment data for six language pairs — English–French (en-fr), English–Romanian (en-ro), English–Japanese (en-ja), English–Chinese (en-zh), Latin–Greek (la-gr), and English–Czech (en-cz) — reconstructed from existing sentence-level, human-annotated word alignment gold standards.
+
+Document-level examples are built by grouping sentence-level annotations by document membership and sentence order; the dataset does not contain new human annotation beyond what already exists in the source corpora (see [Sources & Licensing](#sources--licensing)). A matching sentence-level version of this data is used in the accompanying paper but is not included in this release.
+
+This dataset accompanies the paper *"Scaling Unsupervised Word Alignment to Documents with Structural Constraints"* (\[citation TODO\]).
+
+> **Draft / unpublished:** this repository is currently private and not part of any public release. Remove this notice once finalized for public/camera-ready release.
+
+## Dataset Structure
+
+Each configuration corresponds to a language pair, e.g. `en-ja`. Each row is one document pair (sentences concatenated in original order). Splits are `dev` and `test`, except `en-cz`, which is **test-only** (the source corpus has no dev partition).
+
+### Data Fields
+
+| Field | Type | Description |
+|---|---|---|
+| `sent_id` | string | Unique identifier for the document |
+| `doc_id` | string | Identifier of the parent document |
+| `lang` | string | Language pair code, e.g. `"en-ja"` |
+| `text_a` | string | Source-language text (whitespace-tokenized) |
+| `text_b` | string | Target-language text (whitespace-tokenized) |
+| `sure_align` | string | Space-separated `i-j` sure alignment links (0-indexed word positions, `i` in `text_a`, `j` in `text_b`) |
+| `possible_align` | string | Space-separated `i-j` possible alignment links |
+
+### Example
+
+```json
+{"sent_id": "trial", "doc_id": "trial", "lang": "en-fr",
+ "text_a": "source document text ...", "text_b": "target document text ...",
+ "sure_align": "0-0 3-5", "possible_align": "1-1"}
+```
+
+### Dataset Statistics (dev / test)
+
+| Lang pair | Docs (dev/test) | Sentences (dev/test) | Mean doc length (chars) | Mean sent. length (chars) |
+|---|---|---|---|---|
+| en-fr | 1 / 1 | 37 / 447 | 22,408 | 92 |
+| en-ro | 1 / 10 | 17 / 248 | 2,901 | 119 |
+| en-ja | 8 / 7 | 653 / 582 | 5,516 | 66 |
+| en-zh | 47 / 52 | 450 / 450 | 646 | 70 |
+| la-gr | 1 / 1 | 25 / 75 | 5,837 | 116 |
+| en-cz | — / 33 | — / 1,636 | 6,318 | 126 |
+
+Lengths are target-side (`text_b`) character counts. `la-gr`'s dev/test split is a **custom 25%/75% split** created for this release (the original Ugarit gold standard is not pre-split). `en-fr` and `la-gr` test sets each consist of a **single document pair** — treat document-level, per-language-pair results for these two pairs as illustrative rather than statistically robust (see Limitations).
+
+## Sources & Licensing
+
+This is a derived/reconstructed dataset. Each language pair originates from a separate, independently licensed source. **License terms are per-subset, not uniform across the dataset** — check the table below before reusing a given configuration.
+
+| Lang pair | Original source | License | Required attribution |
+|---|---|---|---|
+| en-ja | KFTT alignment data (Neubig, Kyoto Free Translation Task) | CC BY-SA 3.0 | Graham Neubig, [phontron.com/kftt](http://www.phontron.com/kftt) |
+| la-gr, en-gr, pt-gr\* | Ugarit Alignment Gold Standards (Palladino, Shamsian, Yousef et al.) | CC BY-SA 4.0 | Palladino, C., Shamsian, F., Yousef, T., Wright, D.J., d'Orange Ferreira, A. and dos Reis, M.F., 2023. *Translation Alignment for Ancient Greek: Annotation Guidelines and Gold Standards.* Journal of Open Humanities Data, 9(1), p.22. |
+| en-fr, en-ro | Mihalcea & Pedersen, HLT-NAACL 2003 Workshop *"Building and Using Parallel Texts"* shared task | Not formally specified by the source; \[maintainer note: confirm redistribution terms with original organizers before release\] | Mihalcea, R. and Pedersen, T., 2003. *An evaluation exercise for word alignment.* Proceedings of the HLT-NAACL 2003 Workshop on Building and Using Parallel Texts. |
+| en-zh | Tsinghua Aligner evaluation set (Liu, Xue et al.) | Not formally specified by the source; \[maintainer note: confirm redistribution terms with original authors before release\] | Liu, Y. and Xue, X. |
+| en-cz | Word alignment gold data based on Mareček's diploma thesis, distributed via LINDAT | \[maintainer note: verify exact LINDAT license before release\] | Mareček, D., 2011. *Automatic Alignment of Tectogrammatical Trees from Czech-English Parallel Corpus.* Diploma thesis. |
+
+\* `en-gr` and `pt-gr` are included in the raw data but excluded from the paper's evaluation and from `en-gr`/`pt-gr`-specific processing in the accompanying pipeline; included here for completeness under the same CC BY-SA 4.0 terms as `la-gr`.
+
+By using any configuration of this dataset, you agree to comply with the license and attribution terms of that configuration's original source, in addition to any terms stated here.
+
+**If you are the maintainer preparing this for release:** the three subsets marked "not formally specified" above should not be republished as raw text without either (a) written permission from the original authors, or (b) reducing that configuration to alignment indices only, with `text_a`/`text_b` reconstructed locally by users via the existing `scripts/preprocessing.py` download-and-build pipeline. Resolve this before making the repository public.
+
+## Dataset Creation
+
+Document-level examples were constructed by concatenating sentence-level source/target texts within a document in their original order and re-indexing `sure_align`/`possible_align` accordingly. No new annotation was performed; annotation quality, guidelines, and inter-annotator agreement are inherited from each original source (see citations above for details on each). Full preprocessing code is available at \[code repository TODO\].
+
+## Limitations
+
+- Document structures in this dataset are reconstructed from sentence-level annotations and are consequently near-monotonic; they do not reflect reordering, insertion, or deletion phenomena that occur in naturally-authored document pairs.
+- Sentence-level annotation may not capture document-level phenomena such as long-range reordering, cross-sentence dependencies, or omissions/additions spanning multiple sentences.
+- Some splits contain very few document pairs (`en-fr` test: 1 document; `la-gr` test: 1 document), limiting the statistical reliability of document-level, per-language-pair evaluation on these pairs.
+- `en-cz` has no dev split; hyperparameter selection for this pair in the accompanying paper was transferred from other language pairs rather than tuned directly.
+
+## Citation
+
+If you use this dataset, please cite both this work and the original source(s) listed in [Sources & Licensing](#sources--licensing) for the language pair(s) you use.
+
+```bibtex
+@inproceedings{TODO,
+  title     = {Scaling Unsupervised Word Alignment to Documents with Structural Constraints},
+  author    = {TODO},
+  booktitle = {TODO},
+  year      = {2026}
+}
+```