Hi guys, I am pretty new in NLP. I joined this challenge to put to test what I have learnt so far on nlp. I have tried my hands at every trick I learnt before joining this challenge and the ones I learnt during this challenge. I can't seem to get below a O. 46 score on LB. I have run out of ideas.
Can anyone recommend learning materials to improve my current knowledge state and any tips to improve my score on the LB.
I have used Roberta with simple transformers which has produced my best score so far, word embeddings with keras hasn't helped me very much.
I have really come to like nlp and what it is about. Please I need recommendations on materials that can help me improve my skills even beyond this competition. I would be very grateful for any suggestions.
Have you tried augmenting your test dataset?
I googled articles about augmentating text dataset, I couldn't find anything concrete, the augmention articles I saw were largely for computer vision.
But I will intensify looking into that.
I don't mind recommendations though.
In between, I appreciate you responding.
Hi Busayor,
Here is what I think could improve your score:
1- Spending a little time cleaning your data (indeed, this is the most important part in this competition) will massively improve your score. (Hope I'm going to get burned giving this advice 😂😅)
2- Setting up a strong CrossValidation scheme
3- Don't relying to much on the metric or the LB, it can sometimes be misleading. If you feels like your model is doing good and the LB isn't following, don't worry
4- Don't be afraid to be crazy. Try things out, you never know
5- I don't think any of the Top guys are using Data Augmentation (but like I said, go crazy, try it out)
6- As I like to say, don't start big. Start small and build up from there
Have fun
Wow. Thank you Muhamed_Tuo 😎. I really do appreciate. This is a big push you have given me. I have done little or nothing on data cleaning. I will use this opportunity to improve myself in that area. Thanks again for what you have said from 1 - 6. I really appreciate