Ericu950/classical-swedish-citations-v2

Classical-Swedish Citations (v2)

Cross-lingual citation alignments between classical Greek + Latin corpora and the Swedish literary corpus Litteraturbanken, found by multilingual neural retrieval and judged for genuine citation by a large language model.

This is version 2, superseding Ericu950/classical-swedish-citations. v1 is preserved unchanged for reproducibility of prior work; v2 is the recommended dataset for new analysis.

What changed in v2

• New bi-encoder models for retrieval: Ericu950/intertext-classical-swedish-sentence and Ericu950/intertext-classical-swedish-window, trained from scratch on classical-Swedish parallel data (no annotation contamination).
• New cross-encoder rerankers at sentence and window granularity, trained on the same parallel corpus.
• Multi-channel composite scoring: bi-encoder cosine + cross-encoder probability + corpus-relative rarity weighting, all z-fused.
• LLM-as-judge verification of every candidate alignment, with calibrated probability.
• 3,090 confirmed citations (llm_answer = TRUE) from ~10% of Litteraturbanken (open EPUBs only).
• Precision at 15.6% across the ranked output, up from 4.4% in v1.

LIMITATION

Note that the corpus was filtered by a LLM (Qwen3.6-27b and has false positives and negatives, use with care!)

Subsets

Two subsets, two retrieval granularities of the same underlying alignment problem.

sentences

Sentence-level alignments. Each record pairs one Greek/Latin sentence with one Swedish sentence.

windows

5-sentence window alignments. Each record pairs one source window (5 sentences of context) with one target window. Useful when the citation spans more than a single sentence on either side.

Record schema

Per row:

Field	Type	Description
rank	int	Position in the composite-score ranking
src_lang	str	greek or latin
score	float	Composite z-fused score across all channels
rerank	float	Cross-encoder probability (v4 reranker)
sent_sim	float	Bi-encoder cosine at sentence granularity
win_sim	float	Bi-encoder cosine at window granularity (windows subset only)
src_density	float	Mean cosine to random source-corpus sample
tgt_density	float	Mean cosine to random Swedish-corpus sample
src_rarity	float	1 - percentile-rank of src_density
tgt_rarity	float	1 - percentile-rank of tgt_density
src_idx	int	Source index in the v3 cache
tgt_idx	int	Target index in the v3 cache
src_bin	int	Source sentence index // 5 (bin used for dedup)
tgt_bin	int	Target sentence index // 5 (bin used for dedup)
src	dict	Source metadata: author, title, book, file, lang, lang_conf, sentence_index, sentence_text/window_text
tgt	dict	Target metadata, same structure
llm_answer	str	LLM judge verdict: TRUE or FALSE
llm_p_true	float	LLM-calibrated probability that this is a genuine citation

Filtering to confirmed citations

from datasets import load_dataset

# Sentence-level confirmed citations
ds = load_dataset("Ericu950/classical-swedish-citations-v2", "sentences", split="train")
ds_true = ds.filter(lambda r: r["llm_answer"] == "TRUE")

# Window-level confirmed citations
ds_w = load_dataset("Ericu950/classical-swedish-citations-v2", "windows", split="train")
ds_w_true = ds_w.filter(lambda r: r["llm_answer"] == "TRUE")

Splits by source language

greek = ds.filter(lambda r: r["src_lang"] == "greek")
latin = ds.filter(lambda r: r["src_lang"] == "latin")

Languages

• Source: Ancient Greek (grc), Latin (la)
• Target: Swedish (sv)

Source corpora

• Greek + Latin: Perseus Digital Library and Open Greek and Latin
• Swedish: Litteraturbanken, open EPUBs subset

Models used

• Bi-encoders: Ericu950/intertext-classical-swedish-sentence, Ericu950/intertext-classical-swedish-window
• Cross-encoder rerankers: v4 sentence and window rerankers (release separately)

License

CC BY 4.0.

Citation

If you use this dataset, please cite the accompanying paper (forthcoming).

Acknowledgments

Compute provided by NAISS (Sweden). Corpus access via Litteraturbanken and Perseus.