At this final competition, we are given time for 5 hours with a task to make the best model for sentiment analysis from the given train and test data (non-label) about the Indonesia government rule, PPKM.

Authors (Yaudalah Team) :

1. Ananda Affan Fattahila
2. Fendi Irfan Amorokhman
3. Kaenova Mahendra Auditama
4. Kurniadi Ahmad Wijaya

Actually, At this phase, we tried several data cleansing techniques to find the best model accuracy. The list of it is :

• Stemming + Delete PPKM Words + Stopwords
• Stemming + Minimum 2 Words + Stopwords
• Stemming + Stopwords
• Non-Stopwords + Stemming + Minimum 2 Words
• Non-Stopwords + Minimum 2 Words
• Delete PPKM Words + Minimum 2 Words + Stopwords
• Non Stemming + Stopwords

We also do benchmark the model into the balance and unbalance data. So, all of it was consist 16 Combination For each Models 😱

We actually tuning the models for about 4 hours lol.

After benchmarking, we find out that Unbalance, Non-Stopwords, Stemming, and Minimum 2 Words cleansing give us the best accuracy. Below is the list of the best accuracy we got in some models.

As seen above the IndoBERTweet models give the highest accuracy. Then, because no time left for another benchmark 😋 So, it's time for us to predict the data test.

In the end, the judge tells that we got the highest accuracy from other teams in the data test which has the accuracy of 0.79 and got the 1st Place, Horaay!!!