PoetryTranslationEMNLP2021

Poetry-to-English parallel poetry bitext, reformatted for PoetryMTEB Bitext Mining / cross-lingual retrieval evaluation.

This Hub release is a structured redistribution of the parallel poetry corpus released with the EMNLP 2021 paper Don't Go Far Off: An Empirical Study on Neural Poetry Translation (Chakrabarty et al., 2021; source repository: tuhinjubcse/PoetryTranslationEMNLP2021).

---

Dataset Card

Item	Description
Source	PoetryTranslationEMNLP2021 (corpus_v2.1 + testdatawithtranslations); paper Chakrabarty et al., EMNLP 2021
Languages	Source → English: German (deu), Dutch (nld), Italian (ita), Portuguese (por), Russian (rus), Spanish (spa); test-only: Romanian (ron), Swedish (swe), Ukrainian (ukr)
Size	~192K aligned pairs across 9 configs (see per-config table below)
Label / field type	Parallel bitext (no classification labels). Fields: id (string), source_text (string), target_text (string, English)
Splits	train / validation (6 language pairs from corpus_v2.1); test (all 9 pairs from testdatawithtranslations)
Construction	See Construction method
License	CC BY-NC 4.0
Evaluation metrics	Bitext mining / retrieval: Recall@k, nDCG@k (PoetryMTEB). Original MT study also reports BLEU, BERTScore, COMET, and human faithfulness / poetic-style ratings

---

Features

Field	Type	Description
id	string	Unique pair identifier
source_text	string	Non-English source poetry text
target_text	string	English target / gold translation

Aligned with PoetryMTEB MultilingualPoetryBitextMining schema for drop-in bitext mining evaluation.

---

Configs, languages & size

Each Hugging Face config is one language pair {src}-eng (ISO 639-3).

Config	Source lang	Train	Validation	Test	Total
deu-eng	German	16,999	1,051	1,296	19,346
ita-eng	Italian	34,525	1,998	315	36,838
nld-eng	Dutch	23,376	1,000	159	24,535
por-eng	Portuguese	15,200	700	140	16,040
rus-eng	Russian	50,001	4,186	548	54,735
spa-eng	Spanish	37,744	2,059	536	40,339
ron-eng	Romanian	—	—	100	100
swe-eng	Swedish	—	—	96	96
ukr-eng	Ukrainian	—	—	99	99
All		177,845	10,994	3,289	192,128

---

Construction method

1. Upstream corpus
   Parallel poetry translations collected and released by Chakrabarty et al. (EMNLP 2021):

   • corpus_v2.1: sentence-level {lang}train.json / {lang}valid.json for 6 languages → English
   • testdatawithtranslations: poem-level source (test*) and gold English (gold*) for 6 + 3 additional languages

2. This release (PoetryMTEB packaging)

   • train / validation: read sentence pairs from corpus_v2.1, map fields to source_text / target_text, assign stable ids, write Parquet shards per config.
   • test: for each poem, split source and gold into non-empty lines; align lines within the poem (truncate to the shorter side if lengths differ); emit one bitext row per aligned line.
   • Schema unified with PoetryMTEB Bitext Mining (id, source_text, target_text).

We do not redistribute model-generated candidates from the original repo; only human gold references are used as target_text on the test split.

---

Evaluation metrics

Primary (PoetryMTEB Bitext Mining / retrieval):

Metric	Role
Recall@k	Fraction of queries whose gold translation is retrieved in top-k
nDCG@k	Ranking quality against gold bitext partners

Reported in the original EMNLP 2021 study (MT setting): BLEU, BERTScore, COMET, plus human evaluation of meaning faithfulness and poetic style (paper).

---

How to load

from datasets import load_dataset

ds = load_dataset("PoetryMTEB/PoetryTranslationEMNLP2021", "deu-eng")
print(ds)
print(ds["test"][0])
# {'id': '...', 'source_text': '...', 'target_text': '...'}

Available configs: deu-eng, ita-eng, nld-eng, por-eng, rus-eng, spa-eng, ron-eng, swe-eng, ukr-eng.

---

License

Distributed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).
Please also respect any terms associated with the original GitHub release.

---

Citation

If you use this dataset, please cite the original paper and, when relevant, this PoetryMTEB redistribution.

Paper (required):

@inproceedings{chakrabarty-etal-2021-dont,
    title = "Don{'}t Go Far Off: An Empirical Study on Neural Poetry Translation",
    author = "Chakrabarty, Tuhin  and
      Saakyan, Arkadiy  and
      Muresan, Smaranda",
    editor = "Moens, Marie-Francine  and
      Huang, Xuanjing  and
      Specia, Lucia  and
      Yih, Scott Wen-tau",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.577/",
    doi = "10.18653/v1/2021.emnlp-main.577",
    pages = "7253--7265"
}

Dataset source repository:

• GitHub: https://github.com/tuhinjubcse/PoetryTranslationEMNLP2021
• ACL Anthology: https://aclanthology.org/2021.emnlp-main.577/
• This Hub dataset: PoetryMTEB/PoetryTranslationEMNLP2021