What Word Embeddings Show About Herodotus

I have a fastText model trained on Herodotus (using CADE, a diachronic embedding framework) and wanted to see whether it captures anything meaningful about how group difference works in the text — and whether it could be useful for a broader project on racializing discourse in ancient Greek sources.

The first thing the model shows is where βάρβαρος actually lives in Herodotean discourse. Its nearest neighbors are Σαλαμίς, στρατόπεδον, διώκω, ναῦς, ναυμαχία, φυγή, πολέμιος, Ἀρτεμίσιον. The word clusters almost entirely with military and naval engagement — pursuit, defeat, fleet, battle. This is not the βάρβαρος of philosophical anthropology (the being defined by nature or custom) but the βάρβαρος of the Persian Wars narrative: the adversary in combat. That is a real finding. Whatever conceptual elaboration of group difference happens in Herodotus happens elsewhere, not attached to the general label.

The more interesting signal comes from probing specific peoples against abstract concepts — slavery, freedom, tyranny, custom, nature, lineage. The results are unequal in ways that track the actual structure of the text:

Αἰγύπτιος scores 0.30 against νόμος — by far the highest association of any people with custom and law discourse. This reflects Book 2, where Egyptian customs are described at exhaustive length and consistently framed as the inversion of Greek practice. Egypt in Herodotus is characterized through cultural elaboration: what they eat, how they worship, how they bury their dead. The model picks this up as a corpus-level pattern.

Πέρσης scores highest against δοῦλος (0.24) and τύραννος (0.20). Persian identity in Herodotus is constructed through the vocabulary of slavery and despotic rule — both the freedom-versus-slavery antithesis that runs through the Persian War narrative and the internal structure of the Persian court, where subjects prostrate themselves and the king’s word is law. The model captures the political register of Persian characterization.

Σκύθης clusters with νομάς and γεωργός — nomad and farmer as structural opposites — and scores near zero against τύραννος. The Scythians in Herodotus Book 4 are defined by their resistance to conquest and their nomadic mobility; they cannot be enslaved precisely because they have no fixed land to lose. The negative τύραννος score fits.

Two other results are worth noting. γένος — lineage, descent — scores near zero for all peoples. Whatever Herodotus is doing when he characterizes non-Greek peoples, he is not primarily framing group membership as a matter of biological descent or blood. This matters for the question of whether ancient Greek ethnography constitutes a form of racialization: the answer the model suggests is that it works through practice and custom (νόμος) or through political structure (τύραννος, δοῦλος), not through naturalized hereditary essence. And αἷμα (blood) in the text clusters with ritual vessels, sacrificial animals, and weapons — it is a term of sacrifice and combat, not of descent.

The model has real limits. The vocabulary is 9,278 tokens, which is small; φύσις gives garbage neighbors, suggesting the word is too rare in Herodotus for reliable embedding. The text appears to be in surface form rather than lemmatized, so Πέρσης, Πέρσῃ, and Πέρσεων are three separate tokens splitting frequency. These are problems of corpus size and preprocessing that a single-author fastText model will always struggle with.

What the model cannot do is identify individual racializing passages — it captures distributional patterns across the whole corpus, not the structure of specific claims. But that macro-level view has its own value. It tells you which peoples are characterized through which conceptual domains, which is a different question from whether any given paragraph makes a claim about inherent group character. The domain profiles — Egypt through religion and custom, Persia through tyranny and slavery, Scythia through nomadic freedom — are not close-reading findings; they are aggregate discourse structures that tell you where to look and which annotation categories to expect in density.

The more promising use would be the CADE compass model also in the project, which is designed for diachronic comparison. Separate models trained on Herodotus, Euripides, Plato, and Xenophon would make it possible to ask whether βάρβαρος migrates from its military cluster in Herodotus toward a more normative or philosophical cluster in Plato — whether, in other words, the philosophical elaboration of the Greek/Barbarian distinction that Hall and Saïd both take for granted as a fifth-century development is actually visible in the distributional record. That question is worth pursuing.