[phowon]

With ELMo, the pretrained weights are frozen. Only the scalars for ELMo layers are tuned (as well as the additional top-level model, of course).