search

vdobrovolskii / wl-coref

This repository contains the code for EMNLP-2021 paper "Word-Level Coreference Resolution"
MIT License
104 stars 37 forks source link

Are all models in Table 2 in their base version? #8

Closed yhcc closed 2 years ago

yhcc commented 2 years ago

Thank you for your time, but are all models in Table 2(as depicted in the following) in their base version (it is quite surprise that the RoBERTa-base got 80.72, while RoBERTa large got 81.0)? 8F3D3D34-5662-43B1-85FF-B75DBD81AA1F

vdobrovolskii commented 2 years ago

No, all the models are in their large versions. BCE stands for Binary Cross Entropy.

yhcc commented 2 years ago

Thanks for you timely reply. So, why this ``wl + RoBERTa'' is 80.72, while in Table 1, it is 81.0 ~

vdobrovolskii commented 2 years ago

Table 2 shows performance on the development set, Table 1 uses the test set.

yhcc commented 2 years ago

Sorry for this silly question. But for my reproduction, your model usually gets higher performance in the development set. The highest performance I got in the development set is about 81.6. Thanks again for your answering, I will close this issue.

vdobrovolskii commented 2 years ago

Well, it depends heavily on the initialization. I used the same random seed for all my experiments, probably there is a better one out there :)