Context Matters — Dataset

Latin Intertextuality Test Cases (with controlled modifications) This is a small qualitative dataset of Latin sentence pairs used in the paper Context Matters: Probing the Robustness of Sentence Embeddings for Intertextuality Detection in Latin Texts to study how sentence-embedding similarity reacts to controlled modifications of a query sentence in intertextual links (literal quotation, paraphrase, allusion). The dataset contains 11 base cases, each paired with 12 modified variants of the query sentence (plus one random-control pairing), for a total of 143 rows. Contents context_matters_dataset.csv — the test cases and all controlled query modifications (no cosine similarities included) Intended use The dataset is designed for qualitative / diagnostic evaluation of embedding-based similarity in Latin intertextuality detection, including: reductions to specific word material (content words / function words / citation-only), controlled morphological variation (case, tense, mood, number), semantic(-syntactic) variation (paraphrase, limited reordering), comparison against a random corpus sentence (control). Data format The CSV has 6 columns: case_idIdentifier for the row. The numeric part denotes the base case (1–11). Some rows may append a suffix (e.g., 1a, 1b) but the definitive variant type is given by operation_description. query_idCitation identifier for the later text (the reusing passage), e.g., hier. epist. 84.3.5.3. source_idCitation identifier for the earlier text (the reused source), e.g., verg. aen. 11.282.1. originalThe source sentence (classical author).Note: in the random_corpus_sample condition, this field contains a random sentence sampled from the corpus. paraphrasedThe query sentence (late antique author), in its original or modified form. operation_descriptionThe modification applied to the query (or the control condition). Each base case appears once for each of the following operations: original — unmodified query and source reduction_to_content_words — query reduced to content words only reduction_to_function_words — query reduced to function words only reduction_to_citation — query reduced to citation-relevant material only morph_number — controlled number changes morph_tenses — controlled tense changes morph_moods — controlled mood changes morph_cases — controlled case changes morph_all — combination of multiple morphological changes paraphrase_1 — mild paraphrase (near-synonym substitutions) paraphrase_2 — stronger paraphrase (additional substitutions) paraphrase_2_reorder — paraphrase plus local reordering random_corpus_sample — query paired with a random corpus sentence from the same author (control) Notes on the text Text is provided as Latin strings and may preserve orthography and punctuation from the working corpus. Some modified variants are intentionally not fully grammatical; they are constructed to isolate the effect of minimal controlled changes. This dataset contains Latin excerpts only and does not include modern translations. The Latin text were drawn from freely accessible editions, downloaded from Corpus Corporum, the Tesserae Project and the OpenGreekandLatin Project: Virgil: Greenough 1900, Lucretius: Martin 1934, Jerome letters: Hilberg 1910, other works: Patrologia Latina 1945.