Curated Word Lists

[koaning]

Place to discuss a linter to check for biases in word embeddings.

[koaning]

The idea is to pass it a language backend and then just run a lot of tests to indicate what types of bias may exist in the word-embeddings that you pass it. We can have a linters for different languages and they can be used to generate a report of sorts to demonstrate some of the potential downsides in the dataset.

[koaning]

Here's some tests that come to mind.

We could project some professions to the man-woman axis in the language embedding. This axis allows us to do some hypothesis tests.

• There's a set of professions like "nurse" that technically should be gender-neutral. If this is not the case -> flag it.
• There's a set of descriptions like "beautiful", "handsom" that might also suggest gender imbalance.

[koaning]

The more that I think about this the more that I wonder if it is better to just add word-lists so people can more easily make comparisons. There's not really a consensus to measuring bias.

[koaning]

The more and more that I think about it ... a linter is risky. Even a linter will have blind spots in it and we do not want to give a false suggestion here. Instead it may be more appropriate to supply the user with curated word lists.