François CholletVerified account

All NLP is about mapping the internal statistical dependencies of language, missing the point that language is a *communication protocol*.

You cannot study language without considering *agents* communicating *about something*.

The only reason language even has any statistical dependencies to study is because it's imperfect.

A maximally efficient communication protocol would look like random noise, out of context (besides error correction mechanisms).

why should human language be maximally efficient?

it shouldn't. The point is that it could, and it would still serve the same purpose.

... maybe I'm being slow, but how should acknowledgement of this fact about efficient communication protocols affect NLP?

you can't understand language without considering it in context: agents communicating about something.