ConceptNet: common sense assembled from visitors' sentences
In October 2004 Hugo Liu and Push Singh of the MIT Media Lab described ConceptNet 2.0, a base of everyday knowledge holding 1.6 million assertions over 300,000 nodes. It was not written by hand: it was extracted automatically from the 700,000 English sentences that more than 14,000 people had typed into the Open Mind Common Sense website.
Why it matters
Cyc had spent twenty years arguing that common sense must be written down as axioms, and in twenty years reached a figure of the same order. ConceptNet showed another road to it: ask passers-by in ordinary sentences and parse them with a program. It is the same bet on a crowd and an automatic parser that was placed on Wikipedia three years later.
The paper appeared in the BT Technology Journal, volume 22, number 4, from page 211. Its figures: 1.6 million assertions, of which about 1.25 million are k-lines, that is generic connections between concepts rather than facts about them; over 300,000 nodes; twenty relation types - "The present relational ontology consists of twenty relation-types". The Open Mind Common Sense site had been running since 2000, and its corpus at this point was 700,000 sentences from more than 14,000 contributors. The paper places itself beside its two neighbours in its own words: "While WordNet is optimised for lexical categorisation and word-similarity determination, and Cyc is optimised for formalised logical reasoning, ConceptNet is optimised for making practical context-based inferences over real-world texts." What this record does not claim. The node count is not 280,000. That figure is in the paper, but it is the edge count of the earlier OMCSNet - "a semantic network with 280 000 edges and 80 000 nodes" - not the nodes of ConceptNet 2.0. Nor does the record claim the base is accurate: the authors report that human judges rated a sample of the underlying OMCS corpus and found 75 per cent of items largely true.