---
pretty_name: Document-Level Word Alignment
license: other
license_name: mixed-per-subset
license_link: LICENSE.md
language:
  - en
  - fr
  - ro
  - ja
  - zh
  - gr
  - la
  - 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–Ancient 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)).
This dataset accompanies the paper [*"Scaling Unsupervised Word Alignment to Documents with Structural Constraints"*](https://arxiv.org/abs/2608.21023).


## Dataset Structure

Each configuration corresponds to a language pair, e.g. `en-ja`. Each row is one document pair. Splits are `dev` and `test`, except `en-cz`, which is test-only.

### 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) | Mean doc length (chars) |
|---|---:|---:|
| en-fr | 1 / 1 | 22,408 |
| en-ro | 1 / 10 | 2,901 |
| en-ja | 8 / 7 | 5,516 |
| en-zh | 47 / 52 | 646 |
| la-gr | 1 / 1 | 5,837 |
| en-cz | — / 33 | 6,318 |

Lengths are target-side (`text_b`) character counts. 

## Sources & Licensing

This is a derived/reconstructed dataset. Each language pair originates from a separate, independently licensed source.

| Lang pair | Source | License |
|---|---|---|
| en-ja | [Neubig (2011)](https://www.phontron.com/kftt) | CC BY-SA 3.0 |
| la-gr | [Yousef et al. (2022)](https://aclanthology.org/2022.lrec-1.634/) | CC BY-SA 4.0 |
| en-fr, en-ro | [Mihalcea & Pedersen (2003)](https://aclanthology.org/W03-0301/) | Not specified by the source |
| en-zh | [Liu and Sun (2015)](https://nlp.csai.tsinghua.edu.cn/~ly/systems/TsinghuaAligner/TsinghuaAligner.html) | Not specified by the source |
| en-cz | [Mareček (2011)](https://ufal.mff.cuni.cz/czech-english-manual-word-alignment) | CC BY-SA 4.0 |

## 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. 

## 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
@misc{wastl2026scalingunsupervisedwordalignment,
      title={Scaling Unsupervised Word Alignment to Documents via Structural Constraints}, 
      author={Michelle Wastl and Jannis Vamvas and Rico Sennrich},
      year={2026},
      eprint={2608.21023},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.21023}, 
}
```