PoetryMTEB/MultilingualPoetryRetrieval

Multilingual Poetry Retrieval

BEIR-style multi-aspect poetry retrieval benchmark for PoetryMTEB / MTEB.

Chinese natural-language queries retrieve poems from a shared 24-language corpus along four semantic aspects (emotion / intent / theme / thought), under MONO (single target document language) or CROSS (full multilingual library) settings.

Item	Description
Dataset version	1.2.0
Hub repo	PoetryMTEB/MultilingualPoetryRetrieval
Task	Retrieval (graded relevance, score ∈ {0,1,2,3})
Corpus	138,604 poems (poems_all.jsonl)
Query language	Chinese
Aspects × modes	4 × 2 = 8 evaluation settings
Splits	train (LLM), silver (LLM test), gold (human RAA)
Metrics	nDCG@10, MAP, Recall@k (MTEB Retrieval)

Counts (queries / qrels)

qrels rows = graded positives (score 1–3) plus hard negatives (score 0).

Setting	train q	silver q	gold q	train qrels	silver qrels	gold qrels
emotion_mono	1652	220	48	34692	4620	1008
emotion_cross	5665	1570	48	118965	32970	1008
intent_mono	4928	674	48	103488	14154	1008
intent_cross	5441	1115	48	114261	23415	1008
theme_mono	1721	223	48	36141	4683	1008
theme_cross	5196	1567	48	109116	32907	1008
thought_mono	2840	359	48	59640	7539	1008
thought_cross	5317	1428	48	111657	29988	1008

Totals: queries ≈ 40,300 · qrels ≈ 846,300 · corpus = 138,604

Configs

Config	Splits	Description
corpus	test	Shared multilingual document library
queries_<dim>_<mode>	train / silver / gold	Chinese queries for one aspect × mode
qrels_<dim>_<mode>	train / silver / gold	Labels with score ∈ {0,1,2,3} (0=hard negative; 1–3=graded positive; source poem usually 3)

Placeholders:

• <dim> ∈ {emotion, intent, theme, thought}
• <mode> ∈ {mono, cross}

All query / qrels configs

Aspect	MONO	CROSS
emotion	queries_emotion_mono, qrels_emotion_mono	queries_emotion_cross, qrels_emotion_cross
intent	queries_intent_mono, qrels_intent_mono	queries_intent_cross, qrels_intent_cross
theme	queries_theme_mono, qrels_theme_mono	queries_theme_cross, qrels_theme_cross
thought	queries_thought_mono, qrels_thought_mono	queries_thought_cross, qrels_thought_cross

Relevance scale

Score	Meaning
3	Strong match (incl. source-poem self pair)
2	Moderate match
1	Weak match
0	Hard negative (pooled distractor)

MONO vs CROSS

• MONO: candidates constrained to target_document_language.
• CROSS: candidates from the full multilingual corpus.

How to load

from datasets import load_dataset

corpus = load_dataset("PoetryMTEB/MultilingualPoetryRetrieval", "corpus", split="test")

queries = load_dataset("PoetryMTEB/MultilingualPoetryRetrieval", "queries_emotion_cross", split="gold")
qrels = load_dataset("PoetryMTEB/MultilingualPoetryRetrieval", "qrels_emotion_cross", split="gold")

print(len(corpus), corpus[0]["language"], corpus[0]["content"][:80], qrels[0])

Silver / train:

q_silver = load_dataset("PoetryMTEB/MultilingualPoetryRetrieval", "queries_theme_mono", split="silver")
q_train = load_dataset("PoetryMTEB/MultilingualPoetryRetrieval", "queries_intent_cross", split="train")

Corpus fields

Unified schema for all languages (no per-language extras):

Field	Type	Description
id	string	Poem id
title	string	Title
author	string	Author
content	string	Poem body
language	string	3-letter language code (from poem_id, e.g. ARA / ENG)
token_count	int64	Token count
link	string	Source URL

Corpus languages

code	English	中文	n
ARA	Arabic	阿拉伯语	10000
CES	Czech	捷克语	10000
DEU	German	德语	10000
ENG	English	英语	10000
FAS	Persian	波斯语	10000
FRA	French	法语	10000
ITA	Italian	意大利语	10000
JPN	Japanese	日语	10000
RUS	Russian	俄语	10000
SPA	Spanish	西班牙语	6676
BUL	Bulgarian	保加利亚语	5641
TUR	Turkish	土耳其语	5324
TEL	Telugu	泰卢固语	5227
POR	Portuguese	葡萄牙语	4986
VIE	Vietnamese	越南语	3483
HIN	Hindi	印地语	3227
SLV	Slovenian	斯洛文尼亚语	2804
NOR	Norwegian	挪威语	2353
HUN	Hungarian	匈牙利语	2259
URD	Urdu	乌尔都语	2198
BEN	Bengali	孟加拉语	1789
PAN	Punjabi	旁遮普语	1538
ZHO	Chinese	中文	631
KOR	Korean	韩语	468

Annotation prompts

Per-dimension 4-scale (0–3) labeling prompts are shipped under prompts/:

Path	Content
prompts/prompt_specs.json	Structured ZH/EN rubrics for emotion / intent / theme / thought
prompts/<dim>/system_zh.txt / system_en.txt / system_bilingual.txt	System prompts
prompts/<dim>/user_template_zh.md / user_template_en.md / user_template_bilingual.md	User prompt templates
prompts/prompts.py	Source used by the labeling pipeline

<dim> ∈ {emotion, intent, theme, thought}.
Default LLM labeling language for train/silver: English (prompt_lang=en).

Construction (summary)

1. Sample ~10k poems/language (cap) → shared embedding corpus (sampled_embedding_N10000).
2. Cluster-stratified 8:2 query split → train / test; gold subset from test.
3. Ten-model consensus–difference pooling → ≤20 candidates/query (vote_top20_cap80).
4. 4-scale (0–3) labeling: LLM (gpt-5.6-luna) for train/silver; human RAA for gold.
5. Package BEIR configs for PoetryMTEB.

Note: Train currently includes completed LLM labels only; re-run packaging after more labels finish.

Citation

@misc{multilingual_poetry_retrieval_poetrymteb,
  title     = {{Multilingual Poetry Retrieval: Multi-Aspect Graded Retrieval for Multilingual Poetry}},
  author    = {{PoetryMTEB Contributors}},
  year      = {2026},
  version   = {1.2.0},
  publisher = {{Hugging Face}},
  url       = {https://huggingface.co/datasets/PoetryMTEB/MultilingualPoetryRetrieval},
  license   = {CC-BY-NC-SA-4.0},
  note      = {PoetryMTEB BEIR-style retrieval; emotion / intent / theme / thought; MONO and CROSS; graded relevance 0--3; splits train / silver / gold}
}

License

CC BY-NC-SA 4.0. Poem texts additionally inherit constraints of the upstream compliance-filtered corpus.