julian-schelb/latin-classical-intertextuality-edit-scripts

Latin Intertextuality Edit Scripts

Word-level explanations of text reuse in Latin literature, predicted by retexo.

A later author who quotes or alludes to an earlier one has done something to the words: kept them, changed their inflection, replaced them. For each of the 1,490 references of the Loci Similes benchmark, this dataset holds the edit script that retexo predicts: every word of the reusing passage is aligned to the source word it takes up, or to none, and every link is labeled with the operation behind it (copy, inflection, substitution, split or merge).

• Documentation: julianschelb.github.io/retexo
• Review app: julianschelb.github.io/retexo/demo (browse the predicted script of every pair, with filters, arrows between the passages, and the Latin and English texts)
• Code: github.com/julianschelb/retexo

Related Datasets

This dataset belongs to the Latin intertextuality collection and refers to its records:

• Corpus: julian-schelb/latin-classical-intertextuality-corpus - Classical Latin texts
• Queries: julian-schelb/latin-classical-intertextuality-queries - Jerome texts used as queries
• Labels: julian-schelb/latin-classical-intertextuality-labels - the 1,490 references (benchmark_id below is the row's id there)

Dataset Structure

One record per reference, split into the benchmark's five folds, fold_0 to fold_4 (1,490 records in total: 279, 275, 336, 317 and 283). The script of a reference was predicted by a model that was not trained on that reference's fold, so every record is a held-out prediction.

Fields:

• id: identifier of the pair
• benchmark_id: id of the reference in the labels dataset
• fold: the benchmark fold of the reusing passage
• reference_type: cit. (citation) or cf. (allusion), copied from the labels dataset
• source, reuse: the earlier and the later passage, each with author, work, citation and tokens (the words)
• links: one entry per linked reuse word, with
  • reuse: index of the reuse word
  • source: index of the source word it takes up
  • label: COPY, INFLECT, SUBST, SPLIT or MERGE
  • relation: for a SUBST, the relation between the two words if the model names one (SYN, POS, NE-SUB, ...), else null
  • confidence: the model's probability for the link
• frame: citing formulas such as ut ait Maro, as half-open ranges {start, end} of reuse words
• insertions: reuse words that are neither linked nor part of a citing formula
• deletions: source words that no link claims

All indices are positions in the tokens of the passage they belong to, counted from 0.

Label	Meaning
COPY	the same form, spelling variants folded
INFLECT	the same lemma in another form
SUBST	another lemma in the same slot
SPLIT	two reuse words for one source word (the source index appears in two links), usually an enclitic
MERGE	one reuse word for two source words (the link points at the first, the second counts as a deletion)

Usage

from datasets import load_dataset

scripts = load_dataset("julian-schelb/latin-classical-intertextuality-edit-scripts", split="fold_4")
pair = scripts[0]

print(pair["source"]["citation"], "->", pair["reuse"]["citation"])
for link in pair["links"]:
    print(pair["reuse"]["tokens"][link["reuse"]], "<-", pair["source"]["tokens"][link["source"]], link["label"])

How the predictions were made

The model is retexo's stretch-and-link pointer: a pointer network over LatinBERT that chooses a source word, or none, for each word of the reusing passage and names the operation in the same step, with lexical evidence from Latin resources and a copy rule at decoding that counts spelling variants as copies. For test fold k it was trained on synthetic pairs and the annotated pairs of the three training folds, with fold (k + 1) mod 5 held out to set the link threshold.

Limitations

These are model predictions, not annotation. Most of the model's mistakes on inflections and substitutions are words it left without a link, so a missing link is the first thing to suspect. Read the scripts as a first reading to correct, not as ground truth.

Citation

The paper is in preparation, and its reference will be added here. Please also cite the Loci Similes benchmark, on which the references and their folds are based.