The Representation Problem

A computer cannot hold a word. It holds numbers. So before a network can read anything, every word has to be swapped for a number, and somebody has to decide which one.

That sounds like a clerical job. Hand out the numbers in any order, keep a list, done. It is not. The numbers do more than name things. Once inputs are numbers, some of them sit near each other and some sit far apart, and a learning system treats that nearness as information. It assumes near things behave alike, and it spreads what it learns about one to the other.

So an arbitrary numbering is not neutral. It is a set of claims about which words resemble which, made by whoever handed out the numbers, and almost all of those claims are false. The system then reasons from them faithfully.

This is the representation problem. Find numbers that tell each word apart and whose arrangement also reflects how the words are actually used.

The obvious repair is to arrange them by hand, putting related words near each other on purpose. That fails, and it fails because relatedness pulls in several directions at once, and there is not enough room in a simple arrangement to satisfy all of them. Giving words more room to spread out solves that, but leaves a harder question. Nobody can place a hundred thousand words by hand, and no dictionary lists how near two words ought to be.

The evidence has to come from somewhere else. It comes from usage. Words that keep the same company tend to be used in the same way, and that is a pattern a machine can count.