Measuring how five English translations diverge from Nietzsche's German using sentence embeddings.
Compare how different translators render the same aphorism. Search for keywords, filter by divergence, or hit "Surprise me" for a random high-divergence passage.
UMAP projection of 1,356 embeddings reveals distinct clusters. Each translator leaves a stylistic signature the model can detect.
Hover to isolate clusters. German sits at center. Kaufmann and Hollingdale cluster nearby.
Pairwise cosine similarity reveals Hollingdale as the gravitational center—closest to the German (0.806) and closest to every other translator.
Hover for exact scores. Kaufmann-Hollingdale (0.887) cluster tightest.
I kept switching translations—Kaufmann felt like a professor footnoting away the danger, Hollingdale felt rawer, Zimmern softened everything. Could NLP quantify what I felt?
"Was sich am schlechtesten aus einer Sprache in die andere übersetzen lässt, ist das tempo ihres Stils..."— BGE §28
"That which translates worst from one language into another is the tempo of its style..."
Nietzsche argues that tempo is rooted in "the average tempo of its metabolism"— a language carries the physiological signature of its speakers. The translator who captures words but loses tempo has captured nothing.
Some passages translate consistently. Others scatter wildly—where the German underdetermines the English and translators must make choices no dictionary dictates.
French phrases, cultural critique
Talent reveals character
Genius requires gratitude
Short aphorisms diverge more (less context = more ambiguity). Passages with embedded French scatter wildly. Self-referential passages about language itself prove hardest to render consistently.
Self-taught translator, literary over academic. Sits at the semantic center— closest to German and everyone else. The faithful middle.
Academic standard. Careful, scholarly, heavily footnoted. Clusters tightly with Hollingdale (0.887)—same era, same context.
Oxford World's Classics. Accuracy over style. Reliable but rarely surprising.
Cambridge edition. Takes interpretive risks, modernizes idioms. Drifts further from German but reads more easily.
First major translation. Victorian sensibilities filter the viciousness. Most distant from all others—the machine sees every year of that century.
Sentence embeddings capture semantic similarity in web-text space, not philosophical fidelity. The model learned "meaning" from Wikipedia, not Zarathustra. A philosophy-tuned model was better at Nietzsche's concepts but worse at cross-lingual alignment. You cannot optimize for both.
What I measure: relative divergence patterns. Where translators cluster and scatter. Not which translation is "best"—but that interpretive schools exist and fingerprints are real.
Nietzsche wrote before the 1901 spelling reform. The model sees "Werth" and "Wert" as different words. I built a 95-rule normalizer that improved alignment by 0.002-0.003. Small, but rankings stayed stable.
Translators have fingerprints. Hollingdale sits at the semantic center. And §28, where Nietzsche writes about tempo and the untranslatable? One of the least divergent passages (σ = 0.024). Poetically fitting.