Split-half reliability on SimLex999

[drussellmrichie]

Hi Felix,

Apologies for communicating with you through here. I tried writing to your Cambridge email address, but it bounced back, and I couldn't find your new email address at DeepMind. :-)

I have a (hopefully) quick question: do you know what the split-half reliability among your SimLex999 raters is, i.e., the correlation between the mean of half your raters with the mean of the other half of your raters (possibly over many random 50-50 splits)? I think for many purposes, this might be the right ceiling for model performance....

Thanks for reading, and sorry again for communicating like this. :-)

Best,

Russell