Notes : the narrations is in Indonesia https://www.kaggle.com/code/nadyadtm/distilbert-with-class-weight/notebook
| Date | Progress |
|---|---|
| 16/05/2023 | I found that doing downsampling isn't a good idea. The accuracy is very low (25%) and the loss is more than one (>1) in every epoch. Maybe there are some loss information. So I'll try to use class weight in the loss function. The file will be push soon. |
| 17/05/2023 | Have using class weight in loss function, but the accuracy is in 50%. The file has been pushed (distilbert-with-class-weight.ipynb and distilbert-with-downsampling.ipynb). Next will try to labeling the sentiment based on rating threshold and need to do more preprocessing. Also I'll try to train with more epoch. UPDATED for downsampling, the result is around 40% |
| 21/05/2023 | Try to labeling the sentiment based on rating threshold (4-5 positive, 3 neutral, 1-2 negative). After I labeled the sentence, the accuracy changed to around 80% |
| 22/05/2023 | Add more text analysis (wordcloud) and some preprocessing and simple narration, try to increase epoch, using parameter based on paper (optimizer Adam and learning rate 5-e5), but the precision and recall in negative and neutral class is very low. I have been update it to distilbert-with-class-weight.ipynb |
| 24/05/2023 | Add comparison to BERT based |
| 25/05/2023 - 13/07/2023 | I lost my progress, but I tried to improve the narrations, techniques, and analysis |
| 14/07/2023 | Improve the narrations |