xl2248 / cpcc Goto Github PK

How to split dataset into En to Zh and Zh to En ?

If, as described in the paper "assume 50% speakers as Chinese speakers to keep data balance", then how to identify who is a native English speaker and who is a native Chinese speaker in the BMELD dataset?

Thanks.

Hello, thank you for your released code. This is great work.
There are some questions about the implementation of recognition and prior networks what was mentioned in your paper.

1. The two networks have the same structure in your paper, i.e , MLP and MLP + Softplus, right? But in your released code, It seems that the structure of recognition networks is Linear layer + Linear layer, and the one of prior network is an MLP.
2. It seems that the Sigma and Mu in Gussian distribution have been gotten from the same network in your released code. i.e, you passed the network once, then split the output of the network into two matrixes as Sigma and Mu, right? But in your paper, the Mu and Sigma have been gotten by MLP and Softplus + MLP, respectively.

So I wonder about which one is more effective and can reach the performance that has been reported in your paper.