Tipoff
Five clues share a category and appear one at a time. Name the category as early as you can. The model is guessing too: call it before it does.
Calling it on partial evidence is what a churn or fraud early-warning model does: act while there is still time, and know when the evidence is enough. Wait for every clue and the customer is already gone. Here is how we build it for real.
Every word is a vector learned from how the language is written on the web. The model knows a few hundred category words. After each clue it averages the clues shown so far and compares that average with every category; a softmax over those similarities is its confidence.
It calls it once it is at least half sure of its top pick, the line on the bar. You don't see its pick until the end, only how sure it is getting. Type any form of the word; close synonyms count too. A wrong guess or a skip reveals the next clue.
Built with fastText Common Crawl vectors (CC BY-SA 3.0, Facebook AI Research) and Simplemma (MIT; word-form data from Michal Měchura's lemmatization lists, ODbL, and other open sources listed in its documentation).