You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I try to reproduce the cnn/pcnn with tensorflow. In test phase, I treat a sentence as a special bag, and draw Precision-Recall curve. Surprisingly, single CNN/PCNN get better performance than CNN/PCNN+ONE/ATT。
Another question:the performance of various models look very similar,while in papers, the performance of various models look very different?
The text was updated successfully, but these errors were encountered:
I try to reproduce the cnn/pcnn with tensorflow. In test phase, I treat a sentence as a special bag, and draw Precision-Recall curve. Surprisingly, single CNN/PCNN get better performance than CNN/PCNN+ONE/ATT。

Another question:the performance of various models look very similar,while in papers, the performance of various models look very different?
The text was updated successfully, but these errors were encountered: