SceneLex is an image-first English vocabulary trainer. It is based on a simple learning hypothesis: a word should connect directly to a scene or concept, without requiring an intermediate Japanese translation.
Live demo: https://yurikada.github.io/scenelex/
- Word → Scene: see an English word and choose the closest visual scene.
- Scene → Word: see a visual scene and choose the English word that belongs to it.
- Vector Rank (experimental): lazily loads a sentence-embedding model from a CDN and reranks Wikimedia candidates using their titles and descriptions.
After answering, the app gives a short English definition and example sentence. Session accuracy, streak, and median response time are tracked locally.
Images are searched live through the public Wikimedia Commons API. The creator, license, and source-page link are displayed for every image. Search results are cached in the browser. This repository does not redistribute a bundled image dataset.
The experimental vector feature ranks text metadata, not image pixels. A future CLIP-based pipeline could compare word/sense embeddings against visual embeddings directly, but would require a reviewed dataset and more client compute.
npm install
npm run devnpm run lint
npm test
npm run buildPush the repository to GitHub, then enable Settings → Pages → GitHub Actions. The included workflow builds and publishes the site.
One image cannot fully encode polysemy, function words, or abstract concepts. SceneLex therefore uses several images over time and reveals an English definition plus context sentence after every answer. The intended outcome is not image memorization; it is faster access to a concept from English.
MIT. Individual Wikimedia images retain their respective licenses.