search

PaperDiscovery / CRECL

3 stars 0 forks source link

Question about classification and relation prediction. #2

Open xiujiesong opened 1 year ago

xiujiesong commented 1 year ago

Hi,

thanks for ur excellent work. I have a question about the structure of CRECL. According to the paper, in inference stage u use cosine similarity to get the relation of a instance, right? My question is that why do u still have a classification part? Is classification network necessary or can we just remove it?

PaperDiscovery commented 1 year ago

In fact classification network is to simulate a real-word scenario which uses classification network, so according to your network architecture, it may not be necessary, I recommend another paper: Consistent Representation Learning for Continual Relation Extraction that only uses contrastive network with supervised contrastive learning loss. On the other hand, we do experments in ablation study as CRECL(C), classification network is not able to prevent catastrophic forgetting.