3 ms·
> the issue is that, across a wide variety of contexts, the model prefers the heteronormative output. ...of course it does? If you're a queer data scientist (I
by smoldesu 4y ago
> the issue is that, across a wide variety of contexts, the model prefers the heteronormative output.
...of course it does? If you're a queer data scientist (I'm only queer, unfortunately), you should be able to piece together the fact that any model trained on heteronormative data will give you heteronormative results. Queer individuals are still a minority, in real life and the internet, so it makes complete sense that the models responses would default to heterosexual, perhaps even homophobic behavior.
So, therein lies the problem space. Their 'solution', which amounts to feeding it more queer-specific content, strikes me as a particularly poor response. You're now diluting a dataset that was generalized, biasing it to queer people on Twitter. Not only does that not reflect all queer people (I don't have a Twitter account, for example), but it also soils your model in a completely different direction!
If your model is saying truly despicable things, then censor it's training data. That's at least better than dumping in completely unrelated data to 'fix' a comparatively small problem.