PhiloBERTA

The formal object is a similarity score between a Greek term g and a Latin term l. Rather than compare a single embedding for each, the model averages over many contexts in which each term actually occurs, computing the mean similarity across every pairing of Greek and Latin context vectors. Averaging over contexts is what lets the score absorb polysemy instead of being defeated by it: a word that means several things contributes all of them, weighted by how often each sense appears.

The case for contextual embeddings over older static ones is not merely stylistic. On the model's own evaluation, moving from static FastText vectors to contextual BERT representations cuts the out-of-vocabulary rate from 18.7% to 4.2%, lifts part-of-speech accuracy from 76.3 to 85.1, and more than doubles the cross-lingual correlation (0.31 → 0.68). Static spaces, trained separately for each language, simply never align; the contextual model shares parameters and so shares a geometry.

The deeper trick addresses the scarcity of parallel Greek–Latin text. The model is trained by multilingual knowledge distillation (Reimers and Gurevych, 2020; Krahn et al., 2023a), using English as a pivot: an English teacher model's representations guide the student so that Greek and Latin are aligned through English rather than directly. The authors report that this reduces the parallel data required by 83% against direct alignment. Genre is handled as a source of variance in its own right — an analysis of variance attributes roughly 63% of the spread in similarity to genre rather than time (F = 8.92, p = 0.002) — and missing genre metadata is modelled with adversarial dropout during training rather than simply ignored.

PhiloBERTA is a compact pipeline built on a familiar base. A multilingual tokenizer segments both Greek and Latin into subwords; a transformer core — bert-base-multilingual-cased, 768-dimensional, twelve attention heads, about 110 million parameters — produces contextual representations; a small temporal projection layer (a 768 → 768 linear map) absorbs diachronic variation; and the result is a cross-lingual embedding in which Greek and Latin share one semantic space. Figure 1 sets out the four stages.

Against a plain multilingual BERT baseline the gains are consistent rather than marginal: cross-lingual accuracy rises from 0.78 to 0.92 (+18%), intra-language cohesion from 0.71 to 0.85 (+20%), and genre robustness from 0.63 to 0.89 (+41%). The temporal objective also guards against a characteristic failure of modern-trained models — semantic anachronism. A contemporary multilingual sentence model maps ἄτομον to the modern atom rather than to the Latin individualis that the ancient sense demands; PhiloBERTA's temporal masking reduces such errors by 37%. The lesson, which recurs in classical NLP, is that training data drawn from the wrong epoch quietly imports the wrong meanings (Riemenschneider and Frank, 2023).

Two embeddings can be compared in several ways; the choice matters more than it looks. PhiloBERTA uses angular similarity — one minus the arccosine of the cosine, rescaled to the unit interval — rather than raw cosine similarity. The reason is empirical: embedding magnitudes vary far more in ancient texts than in modern ones (a magnitude variance of 0.18 against 0.05), and the angular form is insensitive to magnitude, reading only the direction in which a meaning points. It is, in effect, a measure built for a noisy corpus.

The data set is small by design and large in care. Ten term pairs are studied — five etymologically related, five controls. For each term the model draws fifty contextual windows from the Perseus corpus, segmented with CLTK's sentence tokenizer for Greek and Latin, balanced so that no term is starved of context. The texts are sampled from the canonical authors of each tradition: Plato, Aristotle, Homer, Thucydides, and Herodotus on the Greek side; Seneca, Cicero, Virgil, Lucretius, and Tacitus on the Latin. Each term's embedding is the average of its fifty context vectors, taken at the model's [CLS] position; the cross-lingual score is the angular similarity between the two averages.

The control pairs are the quiet hero of the design. Without them, a high similarity score proves nothing — any two reasonable embeddings are somewhat alike. It is the gap between etymological and control pairs, and the difference in how tightly each clusters, that turns a pleasant number into evidence.

The central finding is a clean separation between the two pair types. Etymologically related pairs reach a mean similarity of 0.814 with a standard deviation of 0.003; control pairs sit lower and far looser, at 0.780 ± 0.023. A two-sample t-test gives t = 3.219, p = 0.012 — the difference is unlikely to be an accident of sampling. Figure 2 shows the two distributions; the story is in their shape as much as their position.

Table 1 makes the comparison numeric. What is most telling is not the 34-thousandth gain in the mean but the collapse in spread: the etymological pairs are nearly an order of magnitude more consistent than the controls. A tight cluster at a high value is the signature of a process, not a coincidence.

The preservation is also uneven, and the unevenness is interpretable. The strongest alignments fall on the most abstract philosophical concepts: ἐπιστήμη / scientia at 0.820, δικαιοσύνη / iustitia and ἀλήθεια / veritas at 0.814, ψυχή / anima at 0.812. Figure 3 lays the five pairs out as a radar, where the etymological set traces a near-regular pentagon and the control set buckles inward — most sharply at justice.

Resolved term by term in Figure 4, the same pattern recurs: the abstract concepts on the left hold their cross-lingual alignment while the etymological advantage over the control widens for terms whose mappings between the languages were less formalized. The error bars are 95% confidence intervals from ten thousand bootstrap resamples; the gaps that matter clear them.

Three conclusions follow. First, systematic preservation: the extremely low variance of the etymological pairs (σ = 0.003) points to a structured transfer of meaning rather than random convergence. Second, an abstract-concept advantage: philosophical vocabulary, lacking everyday equivalents, was preserved more faithfully than concrete terms. Third, statistical robustness: the significance (p = 0.012) together with the tightness of the etymological cluster gives the pattern evidential weight beyond any single pair.

The most significant number in the study is the smallest: σ = 0.003. A spread that tight, around a high mean and significantly apart from the controls, is hard to explain as chance. It suggests that the passage of these concepts from Greek into Latin was a structured process — that translators were not improvising independently but converging on shared solutions for the core vocabulary of philosophy.

Yet the preservation is concept-dependent, and that nuance is where the interpretation lives. Abstract terms — central to philosophical discourse, and often without ready equivalents in ordinary speech — appear to have been translated with more deliberation and less drift than concrete ones. This fits the conceptual-metaphor tradition (Lakoff and Johnson, 1980; Gaber, 2024): if abstract concepts are understood through stable underlying metaphors, the stability of the metaphor may be exactly what survives the crossing into another language.

For intellectual history, the implication is a measurable kind of continuity. A high, regular alignment in the key philosophical concepts supports the long-held picture of a sustained transmission of ideas from Greek thought into the Roman world (Slisli, 2024; Delignon, 2025; Arzhanov, 2024) — Horace's Graecia capta rendered as a number. The model measures similarity, not identity: ancient translators were creative adapters, not word-for-word copyists. What the results suggest is that beneath that creative adaptation a core semantic kernel was, for the specialized vocabulary of philosophy, reliably preserved. The natural next step is to ask whether different schools — Stoic against Epicurean — preserved their terms differently, a comparison this framework is built to support.

The honesty of the result depends on naming what it does not show. Four limits deserve to be stated plainly.

The term set is small and curated.

Ten pairs, five of them controls, were chosen to represent key philosophical concepts; they are not a sample of the lexicon. The pattern is sharp, but its generality is a hypothesis until the analysis is scaled to a larger and automatically selected set of terms (Copeland, 2024).

The corpus carries its own biases.

Everything rests on the digitized Perseus texts, which are extensive but incomplete. Whatever is over- or under-represented there is over- or under-represented in the embeddings, and the result inherits those tilts.

English is a pivot, and pivots leak.

Aligning Greek and Latin through an English teacher is what makes training feasible (Reimers and Gurevych, 2020), but a modern pivot can introduce subtle modern bias. Multiple pivots, or a method for direct Greek–Latin alignment, would test how much the choice of bridge shapes the view from it.

Time is handled only obliquely.

Diachronic variation is absorbed through genre-conditioning rather than an explicit temporal model — a pragmatic choice, since precise dating is unavailable for most classical works. An explicit temporal embedding would let the evolution of these relationships be tracked rather than averaged over (Perrone et al., 2021; McGillivray et al., 2019). Beyond it lie multimodal evidence from manuscripts and the school-by-school comparisons the discussion gestures toward.

PhiloBERTA shows that a transformer model, properly adapted for the ancient languages, can distinguish etymologically related Greek–Latin term pairs from controls with statistical significance (p = 0.012), and that the etymological pairs preserve their meaning with remarkable consistency (σ = 0.003), most of all for abstract philosophical concepts like ἐπιστήμη / scientia and δικαιοσύνη / iustitia.

The contribution is less a single number than a method. By joining cross-lingual alignment to a tolerance for diachronic and genre variation, the model turns a qualitative claim about the transmission of philosophy — that Greek concepts passed into Latin with their cores intact — into something that can be measured, contested, and extended at scale. The arts that captive Greece brought into rustic Latium left a trace precise enough, it turns out, to read off the geometry of an embedding space.