What Word Embeddings Show About Euripides

A previous note looked at what a fastText model trained on Herodotus reveals about the discourse of group difference in that corpus. The same framework applied to Euripides produces a useful comparison — not because the models are large enough to settle interpretive questions, but because systematic differences between two authors at the corpus level can generate hypotheses that close reading can then test.

The Euripides model has 13,189 tokens against Herodotus’s 9,278, which reflects the larger size of the extant tragic corpus. The same caveats about surface forms and corpus size apply: rare terms give noisy results, and many inflected forms are split across entries. The high-frequency terms — βάρβαρος, Ἕλλην, δοῦλος, νόμος, φύσις, γένος — are reliable; the rest should be treated with caution.

The most important finding is what happens to βάρβαρος. In Herodotus, its nearest neighbors are almost entirely military: Salamis, fleet, naval battle, pursuit, defeat. The word is doing adversarial-military work in that corpus. In Euripides, the neighborhood is different: Φρυξί, Φοῖνιξ, Ἑλληνικός, Θηβαῖος, μιγάς. The word has become primarily a categorical ethnic term, placed in the company of other identity markers rather than in the company of battle verbs. And the term μιγάς — mixed, hybrid, of dual origin — appears at 0.414 proximity to βάρβαρος in Euripides and has no equivalent presence in the Herodotean model. The dissolving boundary between Greek and Barbarian that Suzanne Saïd identifies as a defining feature of Euripidean dramaturgy appears here as a distributional fact: there is a word for the mixed category, and it lives near the term it complicates.

The cross-author comparison of βάρβαρος against a set of conceptual terms makes the shift concrete:

Every concept rises — except φύσις, which barely moves. That exception is the analytically important result. If the Greek/Barbarian distinction in Euripides were operating as a naturalized racial category — groups defined by their inherent constitution, by what they are by nature — you would expect φύσις to rise alongside the political and legal concepts. It doesn’t. What rises instead is the political register: tyranny (+0.184), slavery (+0.134), custom and law (+0.243), blood (+0.213), freedom (+0.101). Euripides is loading βάρβαρος with political-social meaning, not naturalizing it. This fits Saïd’s argument that what Euripides does is foreground and expose the political rhetoric of the distinction — the claim that Greeks are natural commanders and Barbarians natural slaves — rather than grounding it in nature. The model can’t see the dramatic irony, but it can see the discourse field the category is embedded in.

The φύσις embedding itself is worth noting. In Euripides, φύσις clusters with γνώμη (judgment), σοφός (wisdom), εὐγένεια (noble birth), τρόπος (character and manner), σωφρονέω (to be prudent) — a constellation of terms linking nature to intelligence, character, and inherited social standing. This is the sophistic discourse of nature and nurture that Saïd identifies as Euripides' intellectual context. The concept is active in the text; it just isn’t strongly attached to the Greek/Barbarian axis.

The picture for γένος is more complicated. In Herodotus, γένος clusters with ἀπόγονος (descendant, 0.542), τύραννος (0.419), γαμβρός (son-in-law), ἄποικος (colonist) — a dynastic and political vocabulary of lineage and succession. In Euripides, γένος clusters with φύω (to grow, to beget, 0.533), σπέρμα (seed, 0.381), σπορά (sowing), φυτεύω (to plant), and φύσις (0.243). The shift is from political inheritance to biological reproduction. Whatever naturalization is present in Euripides' treatment of group difference, it comes through γένος — through the language of seed, begetting, and natural growth — rather than through βάρβαρος or φύσις directly. The annotation scheme’s distinction between genealogical and naturalistic modes of racialization maps onto this difference: Herodotus handles lineage politically, Euripides handles it biologically.

Blood also shifts register between the two authors. In Herodotus, αἷμα clusters with ritual vessels, sacrificial animals, libation implements. In Euripides it clusters with φόνος (murder), φόνιος (murderous), σφαγή (slaughter), and μυσαρός (polluted, abominable). Tragic blood is not sacrificial blood; it is the blood of violent death and the ritual pollution that follows from it. This difference matters for how bodily harm is coded in each genre: in Herodotus the body appears in ethnographic description (diet, burial, circumcision); in Euripides it appears in the aftermath of killing, carrying moral weight.

The limits of the model remain those of the Herodotus analysis. Several terms of obvious relevance — ἔθνος, ἄνανδρος, τρυφή, ὕβρις, προσκυνέω — give high-cosine-similarity neighbors that are clearly noise rather than semantic signal, meaning the model is interpolating from character n-grams for words it has not seen in sufficient context. The result is consistent with a small corpus trained on surface forms: reliable for frequent, morphologically stable terms; unreliable for anything at the margins of the vocabulary.

What the two models together provide is the beginning of a diachronic discourse map: βάρβαρος shifts from military category (Herodotus) to political-social category (Euripides) without gaining naturalistic grounding. The next step — which the CADE compass framework was designed for — is to track whether the concept continues to shift in Plato and Xenophon, where the philosophical elaboration of the Greek/Barbarian distinction becomes explicit. Whether that elaboration shows up as a φύσις signal, and whether the political terms hold or recede, is a question the per-author models are positioned to answer.