Discovering types for entity disambiguation

blog.openai.com

85 pointsgdb6 comments

> Pick a list of ~100 categories to be your “type” system, and optimize over this choice of categories so that they compactly express any entity

So it's not generic then.

Heh. 100 categories allow you to express 2^100 different things. That's ~10^30, BTW.

Even up to 4 categories means you can describe 100 million different expressions.

I think that means this number of different categories make description generic enough.

That only works if categories are mostly independent of each other. If you have 100 mutually exclusive categories (as in this paper), you only have those 100 different options.

However, for the task of disambiguating ambiguous words, you don't need that many categories. You only need to make sure that most different meanings of each word are in different categories. Given that most words have a very low number of alternate meanings, finding a good categorization to distinguish them isn't too hard.

This is really cool, though the interactive elements seem to be broken in Firefox.

Works for me in Firefox 58 in Ubuntu. Don't know what more to say to you. Yes, very cool research.

Works fine for me too, Firefox Nightly (60.0a1) on Linux.

Reply on news.ycombinator.com