Skip to main content
Back to timeline
arXivSource publication:

Latin text reuse gets a word-by-word account: copy, inflection or substitution, matched with 200 corrected pairs

Synopsis

Computational work on Latin intertextuality could previously only say whether two passages are related, or roughly split references into verbatim citation and allusion, without explaining what the later author changed word by word. A pointer network links each word of the reusing passage to a source word or to none and names the operation, such as copy, inflection or substitution, in the same step; synthetic pairs built by applying known operations to Latin text train a first model, and active learning selects the pairs an annotator corrects. On the 1,490 references of the Loci Similes benchmark, 200 pairs chosen by uncertainty bring the model level with one trained on every annotated pair (macro F1 84.58 against 84.44).

Source-provided article image: Learning Word-Level Explanations of Text Reuse in Latin Literature with Little Annotation
arXiv

Interpretation

Text-reuse analysis moves from deciding whether two passages are related to explaining word by word how a later author reworked the source: each word of the reusing passage is linked to its source word or marked as having none, and carries an operation label such as copy, inflection or substitution, together forming an edit script. Earlier computational work on Latin intertextuality sorted a passage pair into no reference, verbatim citation or allusion, a label that says whether the passages are related but not how they were reworked; word-alignment methods come mostly from translation and give links without labels, and word-level annotation for Latin existed only for 100 Bible reuses. On the 1,490 expert-verified references of the Loci Similes benchmark the model reaches the best macro F1 at 85.37 and leads on both hard labels, inflection at 78.36 and substitution at 53.60, while a copy floor that links only identical forms reaches macro F1 53.47.

The model decides in one step whether a word has a source at all and which source word it is and what operation was applied, rather than aligning first and naming links with a fixed rule. Aligning first and naming links by a fixed rule loses most on substitutions: the rule treats a shared lemma as an inflection and everything else as a substitution, and cannot use meaning relations. On the same links, a classifier trained afterwards reaches macro F1 72.00 and the fixed rule 81.04, while the model that labels while it links reaches 84.44; averaging its link scores with the span aligner raises link F1 to 95.47 but, named by the rule, lowers macro F1 to 83.53.

Annotation cost falls to a few hundred pairs: synthetic pairs first teach the shape of the task, and active learning then has annotators correct only the pairs the model is least sure of. Uncertainty sampling usually fails from a cold start, when the model has not yet learned the task; the synthetic stage exists to remove that cold start. Starting from the model trained on synthetic pairs alone, 200 pairs chosen by uncertainty reach macro F1 84.58, against 84.44 for training on all annotated pairs and 83.54 for random selection; the two other strategies do not exceed random selection.

Perspective

The demonstrated scope is the 1,490 Latin literary references of the Loci Similes benchmark, 858 verbatim citations and 632 allusions, evaluated on the benchmark's five folds; the predicted scripts and the review app cover all 1,490 references, and a reader can inspect each script word by word there.

The corrections of the annotation loop were made by a single annotator, so the annotation reflects one reading of the guideline and no inter-annotator agreement could be measured; where the guideline leaves room, as in naming the relation of a substitution or marking a link as possible, another annotator may decide differently. A relation the Latin lexical resources do not record cannot be named by the rule and leaves the lexical evidence silent, and most of the substitutions the model misses are of this kind.