HenryPengZou / JointMatch

[EMNLP 2023] Official Code of "JointMatch: A Unified Approach for Diverse and Collaborative Pseudo-Labeling to Semi-Supervised Text Classification"
https://aclanthology.org/2023.emnlp-main.451
17 stars 0 forks source link

About ema_mode #2

Closed hsannn closed 8 months ago

hsannn commented 9 months ago

Thank you for your thoughtful response to my previous question!

I have another question: Can you explain why the IMDB dataset is the only one with ema_mode set to True?

HenryPengZou commented 8 months ago

Hi @hsannn, sorry for the late reply. We used EMA only on the IMBD dataset primarily because: We noticed that performance on the IMDB dataset varied significantly with different runs. EMA often results in more stable results. Therefore, we explored the use of EMA on the IMDB dataset. However, even after employing EMA, we noticed the performance variance across different runs is still quite high compared to performance variance on other datasets even without EMA. We didn't use EMA on other datasets because the performance variance on them is small without EMA. But you can also explore EMA on other datasets, which might help you further boost the performance. Hope this information can help~