search

gkoumasd / MSAF

Fusion Modality Approaches for sentiment analysis and emotion recognition task.
12 stars 2 forks source link

confusion about the performance on IEMOCAP #4

Open 201528014227051 opened 3 years ago

201528014227051 commented 3 years ago

Thank u for the great work. When I run the MARN model on the IEMOCAP data set, I find that the performance of in the output .csv is higher than the results reported in the paper. the accuracy is 0.824627 F1 score for every class is 0.7105674147704142, 0.8270833752719677, 0.8452668586306232, 0.8754207538192843 am I missing something?

gkoumasd commented 3 years ago

Hi,

thank you for the feedback. Indeed, the performance is too high for a multi-class classification task. According to the literature, MULT is the sota approach. But, still, your results are pretty higher from Mult's performance. Have you changed the configuration settings (i.e., label) from sentiment to emotion?

201528014227051 commented 3 years ago

I have changed the label. After comparing the hyperparameter settings, I changed the attn_num to 1, It seems overfitting if atten_num is larger. and another difference is the epoch number in run.ini file, I changed it from 1 to 5