taoshi1998 / sslcl Goto Github PK

Hello @TaoShi1998,

Thanks for confirmation that we could use either pre-trained or non pre-trained network for 'ERC model'.

Additionally, requesting your to please confirm on below queries:

1. We are training both 'ERC model' and 'Label learning' model using back propagation. The SSLCL + cross entropy loss is calculated for ERC model and same is back propagated through both ERC network and label learning network using two different optimizers in PyTorch (code screenshot attached). Please could you provide feedback if this training approach matches your paper's recommendations.

2. If a non pre-trained ERC model is used then the parameters of the ERC model are to be optimized based on SSLCL + cross entropy loss. The SSLCL loss is dependent on the label embeddings. The label embeddings inturn need to be optimized based on the SSLCL + cross entropy loss from ERC model(which is also not yet optimized). With such interdependency, how is it guaranteed that the optimal label embeddings are learned ?

Thanks and regards,
Supriya

Very nice work! I was wondering if you have any plans to open the entire code? Thanks

Hello @TaoShi1998,

We are trying to implement the proposed framework using a transformer based model as 'An Arbitrary ERC Model'(figure 2). We are jointly training 'ERC Model' and the 'Label Learning' model. 'ERC Model' is trained from scratch. We are not using a pre-trained model.

Please could you confirm if above implementation is in line with the proposed framework. Or do we have to use a pre-trained ERC model as 'An Arbitrary ERC Model' in the framework?

Thanks and regards,
Supriya