elmiram/anime-professions

Which real-world jobs does anime actually show? ~22k titles classified to ISCO-08. Live explorer + dataset. - elmiram/anime-professions

0 stars · 0 forks · 0 watchers

README

Anime Professions Which real-world jobs does anime actually show? This project scans 21,988 anime and counts how often each occupation appears — a complete lookup table keyed to ISCO-08, the UN/ILO's standard occupation classification. The headline finding is the absence list: the jobs anime never names at all.

🔎 Live explorer: https://elmiram.github.io/anime-professions/  ·  search any profession, see which anime feature it, and how it trends by decade. 📊 The deck: a 12-slide fun tour of the findings (PDF).

A few things it found:

Teachers are the most-named profession (in ~1,370 anime, 7.6%) — anime lives at school. In the 2020s, shopkeepers overtook teachers for #1, on the back of the isekai/slice-of-life boom. If you count fantasy roles, being royalty (1,792 titles) out-appears every real job. Of the ~150 jobs a character could plausibly be named as, anime never names 17 — even across millions of lines of dialogue: financial analysts, systems analysts, driving instructors, printers, fast-food workers, ships' engineers… the back-office and service tail. (See The occupation list below.) Your niche jobs, quantified: software developers 41, insurance reps 18 (Trigun's leads are literally insurance agents), pharmacists 15, tram/bus drivers 5.

What this is A hobby data project. It builds a taxonomy of occupations, detects mentions of them in anime, classifies each mention with an LLM, and produces per-occupation counts, a decade-by-decade breakdown, precision/recall estimates, and three CSVs. A small local/static web app explores the results. Real occupations are kept strictly separate from a fantasy stratum (ninja, mage, adventurer, royalty, samurai…) — the fictional roles are counted but never folded into the real-job totals. How it works The occupation list — 436 trimmed to ~150. Occupations are classified against ISCO-08, the UN/ILO's list of 436 official job codes. Most of those can't be named in a story — nobody announces "I'm a mineral-processing-plant operator" — so a zero for them would say nothing about anime. So the list is trimmed to the ~150 jobs a character could plausibly be identified as (the "renderable" set); the other ~286 are reported as untestable, not absent. Anime names 133 of the 150 — the 17 it never names are the headline finding, because they could plausibly appear yet don't. (Every anime is still scanned for all ~150; the trimming is of job categories, not titles.) Detection itself is text-first, in three layers:

Corpus + protagonists — every anime's plot summary (via AniList). Supporting characters — AniList character bios, with a role field (main / supporting / background) so a doctor who appears in six episodes still counts as representation. Dialogue — ~128,000 Japanese subtitle files (from a community mirror), so a job mentioned only in passing gets caught too.

Detection is a two-stage pipeline:

Stage 1 — high recall. A keyword lexicon (English + Japanese surface forms mapped to ISCO codes) casts a wide net over the text, accepting false positives. Stage 2 — precision. Every ambiguous candidate sentence is judged in context by an LLM (Claude Haiku): does a character here really hold this job? This kills the classic traps — negation ("I'm not a doctor"), hypotheticals ("if I were a pilot"), polysemy ("officer", "band"), and fantasy homonyms (a mecha pilot is not an airline pilot).

A stronger model (Claude Opus) then grades a stratified sample of the confirmed hits to estimate precision per occupation, and AniList tags provide a recall benchmark. Occupations whose precision falls below a floor are flagged unreliable in the output rather than reported as fact. Reading the numbers (important caveats)

Named, not shown. We detect jobs that are named in text. A profession shown on screen but never labeled — a festival food stall, a hotel receptionist standing in frame — is invisible to the method. The "never-named" list is exactly that: never named. Unreliable flags are real. ~a quarter of measured occupations are flagged unreliable (e.g. "aircraft pilot" is mostly mecha pilots) — treat those counts as upper bounds. Franchise vs entry. Counts deduplicate to franchise by default (Detective Conan's ~28 films = 1), but every entry is kept as its own row so you can expand or collapse. Denominator is printed next to every percentage (of ~18,127 titles with scannable text).

Data sources & ethics

Anime metadata, titles, character bios, tags: AniList (GraphQL API). Occupation taxonomy: ISCO-08 (International Labour Organization). Dialogue: a community archive of Japanese anime subtitles.

Subtitles are copyrighted works. This project derives counts and locators only — it never stores or redistributes subtitle text. The exported dataset and the web app contain derived statistics, titles, and AniList ids; no subtitle text, and not even the raw AniList descriptions. Non-commercial fan project. The explorer Two ways to run the same UI: python ap.py web # local server at http://127.0.0.1:8000 python ap.py export-static # precompute a static ./site for GitHub Pages (no backend) The static build turns every API response into JSON and answers the same calls client-side, so it hosts anywhere as plain files. Run the pipeline yourself Requires Python 3.11+ and an Anthropic API key (for the LLM stages). python3 -m venv .venv && ./.venv/bin/pip install -r requirements.txt echo "sk-ant-..." > api_key.txt # or export ANTHROPIC_API_KEY

python ap.py collect # AniList corpus + characters + tags (free, ~15 min) python ap.py franchise # cluster franchises python ap.py extract # Stage-1 lexicon candidates (free) python ap.py estimate # exact candidate count + projected batched cost (free) python ap.py adjudicate # Stage-2 Claude Haiku via Batch API (spends) python ap.py measure # precision/recall (Claude Opus gold + tag recall) python ap.py analyse # coverage report + zero-mention list python ap.py export # occupation_counts.csv, title_occupation.csv, evidence.csv

Optional Layer C — subtitle dialogue:

Credits & license Built with the Claude API. Data © their respective providers (AniList; ILO for ISCO-08). Code released for reference; the data shown is derived and non-commercial. Not affiliated with AniList, MyAnimeList, or any studio.

source