Thanks Will! It appeared in the ACL'24 proceedings. This post of mine was a repost from X/Twitter 6 months ago.

That's really interesting how you could normalise the disfluencies to get to something more readily translatable, and that this simulates what humans seem to be doing subconsciously.

For more of this story, please watch the 10 minute intro video, or dive into my ACL'24 paper "Must NLP be Extractive?" What do you think? Please let me know in the comments. 5/5
bit.ly/bird-must-nl... aclanthology.org/2024.acl-lon...

There are other ways of thinking about what language *is*. Our epistemology of language affects how we approach language technologies. The first wave and second waves of symbolic and subsymbolic NLP are familiar. I believe a third wave is emerging: Relational Language Processing. 4/5

When we do this, something interesting comes up: relationality. Language is intrinsically social, inseparable from culture, society, or nature. Language both reflects and creates context. This exceeds narrow epistemologies of "language as lexicogrammatical code", or "language as sequence data". 3/5

I find it instructive to look at Indigenous societies, who are guarding their agency in the face of relentless pressures from outside, enacting their sovereignty in shaping their lives, landscapes, and languages. We who increasingly cede our agency to AI may have something to learn. 2/5