PoetryMTEB/ClassicalPoetryAllusionsRetrieval

Classical Poetry Allusions Retrieval

Retrieval benchmark for classical Chinese poetic allusions (用典) under PoetryMTEB / MTEB BEIR-style evaluation.

Given an allusion source query (典源词 + 典源内容), retrieve poem/verse examples that instantiate related allusion forms (典形词例句).

Upstream knowledge base: Ancient-Chinese-Allusion-Resource-Database (莫凯洁等, 中文信息学报, 2024).

Dataset Card

Item	Description
Dataset version (PoetryMTEB)	1.0.0
Task	Retrieval (document ranking)
Languages	Classical Chinese / Chinese (zh)
Query	Allusion source term + source passage (queries.text)
Corpus	Poem/verse example sentences using related allusion forms
Relevance	Same-allusion sampled examples (score=1)
Splits	test only for all configs (MTEB Retrieval)
Size	corpus=4838; queries=503; qrels=5013
Avg. relevant docs / query	9.97
Evaluation metrics	nDCG@10, MAP, Recall@k (MTEB Retrieval)

Sampling (prior packaging, preserved)

• Prefer diverse keyword examples per allusion.
• Drop allusions with fewer than 10 examples; keep up to 10 examples per allusion.

Configs

corpus (documents)

Field	Type	Description
id	string	Document id (e.g. P04047)
text	string	Retrieval document (poem/verse sentence; formerly sentence)
keyword	string	Trigger allusion keyword in the sentence
source	string	Bibliographic / poem attribution

queries

Field	Type	Description
id	string	Query id
text	string	Retrieval query: source_term + newline + source_content
source_term	string	典源词 (allusion name; formerly Source_Term)
source_content	string	典源内容 (source passage; formerly Source_Content)
derived_list	string	JSON string of 衍生列表 (formerly Derived_List)

qrels

Field	Type	Description
query-id	string	Matches queries.id
corpus-id	string	Matches corpus.id
score	float64	Relevance (1.0 for positive pairs)

How to load

from datasets import load_dataset

corpus = load_dataset("PoetryMTEB/ClassicalPoetryAllusionsRetrieval", "corpus")["test"]
queries = load_dataset("PoetryMTEB/ClassicalPoetryAllusionsRetrieval", "queries")["test"]
qrels = load_dataset("PoetryMTEB/ClassicalPoetryAllusionsRetrieval", "qrels")["test"]

print(queries[0]["source_term"])
print(queries[0]["text"][:200])
print(corpus[0]["text"], corpus[0]["keyword"])

Construction method

1. Load existing Hub corpus / queries / qrels test parquet.
2. Strip HTML <u> tags; add MTEB-standard text fields.
3. Snake-case query metadata (source_term, source_content, derived_list).
4. Write BEIR-style configs + Dataset Card + supporting materials.

Supporting materials

Path	Content
VERSION / LICENSE / CITATION.* / metadata.json	Packaging metadata
checksums.sha256	File checksums
scripts/evaluate_allusions_retrieval.py	Baseline BM25-style lexical probe (optional)

Citation

@article{Kaijie2024,
  title   = {古汉语典故资源库的构建及应用研究},
  author  = {莫凯洁 and 丘子靓 and 王予沛 and 胡韧奋},
  journal = {中文信息学报},
  year    = {2024},
  pages   = {27--34}
}

Upstream: https://github.com/KaijieMo-kj/Ancient-Chinese-Allusion-Resource-Database
This packaging: PoetryMTEB/ClassicalPoetryAllusionsRetrieval (version 1.0.0)

License

Follow the upstream GitHub project terms. Marked unknown on Hub pending explicit SPDX on the source repo.

---

古诗词典故检索（中文说明）

BEIR 式 Retrieval：用典源（典源词+典源内容）检索相关诗词例句。
corpus=4838 / queries=503 / qrels=5013（仅 test）。