search

chujiezheng / CoMAE

Codes and data for the ACL 2021-Findings paper: CoMAE: A Multi-factor Hierarchical Framework for Empathetic Response Generation
39 stars 6 forks source link

Can't find speaker embeddings. #1

Closed thomas-happify closed 3 years ago

thomas-happify commented 3 years ago

Hi there,

Amazing work where you include cognitive science into the model!

I have 2 questions:

  1. I didn't find any speaker embedding in the code where src_additive_embeds only includes word, positional, DA, and EM embeddings.

  2. What's the intuition of doing DA/EM/CM classifications first? It looks very similar to an autoencoder since inputs == outputs. Or did I misunderstood?

Thanks!

chujiezheng commented 3 years ago

Thanks for your interest!

  1. src_additive_embeds only contains DA and EM embeddings. The speaker embeddings are added through token_type_ids. You can see that I add the token_type_embeddings here
  2. Sorry, I do not understand what you mean. I first use the last hidden state of the context to predict the DA/EM/CM adopted in the response, and then use the predicted factors to control the response generation.
thomas-happify commented 3 years ago

@chujiezheng Thanks for the reply! it's really helpful.

I misunderstood the 2nd question yesterday so we are good.