The Marvel map

The course page ends week 6 with a map: characters as dots, axes made of word lists — the average embedding of science, technology minus the average of magic, spell, sorcerer — and one honest warning: none of these positions is a fact about the characters. So we shipped the vectors into the browser and let you draw your own axes. Plus the second half of the week's toolkit: a Scattertext-style term-rate contrast between communities.

The recipe is the course page's, run on our own Marvel-trained skip-gram vectors (100 dims, window 5, seed 0 — no GloVe download needed):

1. Every character keeps a fingerprint — the words with the highest TF-IDF weight once proper names are removed, shipped with their weights (top 12 per character).
2. An axis is a direction in embedding space: the average vector of the positive pole's word list minus the average of the negative pole's, normalized.
3. A character's projection is the TF-IDF-weighted average of their fingerprint vectors, dotted with the axis direction. The map plots two such axes, X and Y — presets or any two words you type.

The vectors themselves ship to your browser quantized to int8 — that's what makes the type your own axis box possible without pre-baking anything.

The Scattertext panel: pick any two of the eight largest Louvain communities and look at their word rates side by side. Each dot is a term; the axes are uses per 10,000 tokens on a log-log scale; the dashed diagonal is equal rates. Red dots are relatively distinctive of the X-community, blue of the Y-community.

science ↔ magic works. Doctor Strange lands deep in the magic end (his magic, mystical and ancient do the pulling) while Machine Man, Supreme Intelligence-adjacent tech characters, and gadget heroes hold the science end; Spider-Man drifts somewhere in the middle — a street kid with a lab and a magic-adjacent rogues' gallery.

war ↔ home is the bias axis, exactly as the course page warned. The home end collects female-gendered pulls — female-titled pages, the women of Wakanda, the Shuris and Maris of the list — while the war end is almost all soldiers. It's GloVe's Wikipedia fingerprint re-negotiated on a much smaller corpus: the bias is in how the pages are written, not in the embedding tool.

Term contrasts answer a different question. Between the Wolverine-led and Spider-Man-led communities, the diagonal catches the shared Wikipedia dialect (series, character, team), Wolverine's corner floats x-men, logan, adamantium-family words toward its end, and the Spider-corner pulls spider, peter, venom. The same words the embeddings turn into geometry show up here as raw rates — two lenses, one landscape.

One fitted Word2Vec run (seed frozen); change the seed and dot positions drift. Axes made of single words are noisier than the 5-word presets; a pole word that is also a character name behaves differently than an abstract pole. And 727k tokens is small: every position carries corpus noise the course page's GloVe map hid behind a 128 MB training set. The point of the page is exactly that difference — check both, believe neither.