Closed kunalpagarey closed 3 years ago
Hi! You might want to use the Speaker utterance as a label if you're training a model on both Speaker and Listener utterances, as we do in the paper. (Empirically, we found that results were slightly better at least for the retrieval model when we did so.)
Hi @EricMichaelSmith, I am assuming that the model BERT, fine-tuned provided here https://dl.fbaipublicfiles.com/parlai/empatheticdialogues/models/bert_finetuned.mdl is trained in both the roles. Please let me know if that is correct.
Thank you.
Yes @kunalpagarey that's correct.
Hi @EricMichaelSmith, I was going through the data preparation for ED dataset in empchat.py file and found that the speaker utterance is also taken as label with prev conversations utterances as context. I am a little confused as to why would you take speaker utterance as a label when you want a response in listener role only?
Is data preparation different for generation and retrieval tasks?
Please clarify this. Thank you.