Chain of Association

Pre-AI word game

Type an English word and press Enter. ConceptNet answers with something it associates with your word; associate on that one, and the chain keeps growing.

Creative Coding Web Experiment ConceptNet API

Context

Where this comes from

This experiment is deliberately pre-LLM. ConceptNet grew out of Open Mind Common Sense, started at the MIT Media Lab in 1999, and the /related endpoint this page calls ranks words by embedding similarity: semantic relatedness as computational linguistics computed it through the 2010s, assembled from a knowledge graph and word vectors rather than asked of a language model. When ConceptNet's API drops out (it does, for days at a time), the page falls back on Datamuse, which draws on the same era's word2vec and WordNet, so the game keeps answering.

I worked on that question back then. My 2015 master's thesis, Konzeptionelle Dichte, examined how machines form human-like word associations, built around a prototype for automated image retrieval. Chain of Association replays the question as a game.

The hypernym chain above the word table, from entity down to furniture, with its senses beside it and its hyponyms below — Hypernyms above, senses beside, hyponyms below: how WordNet, one of the graphs folded into ConceptNet, structures what “table” means. Figure from Konzeptionelle Dichte (2015).

The question itself has not aged: what a machine associates with a word, and how far to trust it, is the judgment I now apply to agent output daily.