X-ray for sentences
Stanza, Stanford's multi-language NLP library, gets auditioned as Severo's grammar x-ray: dependency arrows showing that a whole sentence hangs off its verb (Obama pays attention to born), German words carrying their nominative or accusative case and gender in color. Three visualization drafts die on camera, and the surviving problem is not the linguistics but the labels: S, VP, NP and 'nsubj' read like another foreign language stacked on the one you're learning, so the design work becomes translation, friendly names, arrows that appear one at a time on tap, and a for-nerds toggle with a legend for whoever wants the raw taxonomy.
// trace: where this idea came from
- ↳ video diary @ 33:43 (la versión dos, las flechitas)
- ↳ video diary @ 34:17 (¿qué es S? ¿qué es VP? el susto de las siglas)
- ↳ Entry 226-1: A translator, not a dictionary (el clic que traduce, ahora con anatomía)
The tap that translates a word gets an ambitious sibling: showing how the sentence hangs together. The instrument is Stanza, the natural-language-processing library he’d been probing, and the audition is a chain of generated visualizations graded live. Version one, a table, boring, discarded ▸ 33:24. Version two draws the dependency arrows and earns interest, tap a word and its relations unfold below ▸ 33:43. Version three, nested constituency boxes, is rated by its own author as something you look at and get very scared ▸ 34:05.
What the arrows actually teach survives the fright: a sentence is centered on its verb, everything in Obama was born in Hawaii points at born, who, Obama, where, Hawaii ▸ 35:32, and in the German demo each word wears its case and gender, Hauptstadt nominative feminine, hence die, accusative elsewhere, hence die again ▸ 37:45, exactly the invisible machinery a learner never gets shown.
la gramática con rayos X →
The real defect is the labels. Julia’s reaction is the spec: “¿Qué es S? ¿Qué es VP? ¿Qué es NP?” ▸ 34:17, and he concedes that nsubj and its cousins read like a second foreign language stacked on the first, one even he can’t parse ▸ 36:44. So the roadmap becomes translation, not linguistics: friendly names for the tags, arrows that appear one at a time when a word is pressed instead of all at once ▸ 36:51, and the full taxonomy demoted to a for-nerds toggle with a legend explaining itself ▸ 37:25. The x-ray works; the radiologist’s report still needs subtitles…