Question re: `marginalized_responses`

[bkj]

Hello --

I'm trying to understand how the gradable adjectives get used to initialize the KDEs for the prior distribution.

I noticed that the responses get resampled to have length res:

• https://github.com/ml4ai/delphi/blob/master/lib/parameter_initialization.cpp#L256
• https://github.com/ml4ai/delphi/blob/master/lib/parameter_initialization.cpp#L282

Are you able to explain the reasoning for this resampling?

~ Ben

[adarshp]

@bkj - The gradable adjectives have different numbers of responses from the crowdsourcing experiment. We wanted to compensate for this and thus resample from a KDE that uses the existing responses as kernels, and use those samples as KDE kernels for future usage, so that all the adjectives have KDEs with the same number of kernels.

[bkj]

@adarshp -- thanks for the response!

It makes sense that you'd want to balance the weight of responses across the gradable adjectives.

Follow on question -- is there a reason you use the same parameter res for the number of KDE resamples and for the number of posterior samples here:

• https://github.com/ml4ai/delphi/blob/master/lib/train_model.cpp#L63

Eg, as opposed to using separate parameters for the two quantities?

Thanks again!

[adarshp]

@bkj - No particular reason - I think back in the day the number of pieces of evidence was small enough that we didn't care too much about specifying separate sampling resolutions for these two quantities, and using the same parameter simplified some function/method signatures.

The right thing to do, of course, is to use separate parameters. This is a piece of technical debt we should address eventually.