Daniel Jeffries (Unknown)

You've done loads and loads of work in the data collection part of that predictive pipeline that we were talking about.

Can you tell us about that?

Where did those labels come from?

If I understand correctly, there was something called WordNet, which was like a lexical database.

Didn't they just like steal some categories from that?

Yeah, because that must be a constant problem.

Yeah.

And what are the adverse effects of that?

Like, you know, sort of de-biasing the data implicitly.

And in this case as well.

So it's built into the loss function.

And I guess it gets harder and harder to balance the weighting of terms when you have more of them.

So for example, you might do, you know, like self-selection bias and you might do another bias.

And then you've got this kind of problem that one might dominate the other.

Amazing.

Andrew, this has been so, so brilliant.

How can people find out more about yourself and your lab?

That's amazing.

So what's going to be your research plan?

Are you interested in reaching out to industry?