[visarga]

They could have tried to have a dataset of bias triples (A in relation to C is like B in relation to C), and minimise the score on that by adding it to the loss function, so the model trains with minimal bias.