IFEST 2021 Sentiment Analysis: E-Wallets and PPKM Policy

A team's winning entry in a national data analysis competition: CNN-LSTM for e-wallet reviews in the selection phase, then IndoBERTweet for PPKM policy tweets in the final.

1st place
IFEST 2021 Data Analysis Competition, final round
84.8%
IndoBERTweet validation accuracy in the final
9,810
Labelled PPKM tweets in three classes

Role: Team member in a team of four
Period: Aug 2021 to Oct 2021
Project type: Featured
Room in the 3D HQ: L3:ifest-2021-sentiment-analysis

Results

1st place
IFEST 2021 Data Analysis Competition, final round
Organized by Universitas Padjadjaran; the judges reported the team's 0.79 test accuracy as the highest of all teams (repository README).
84.8%
IndoBERTweet validation accuracy in the final
On 1,962 validation tweets after three epochs (fine-tuning notebook output).
9,810
Labelled PPKM tweets in three classes
The competition's training data after cleaning (the unbalanced training file in the repository).
82.8%
CNN-LSTM validation accuracy in the selection phase
On 1,108 validation texts, F1 0.83 (training log in the CNN-LSTM notebook).

Problem

The IFEST 2021 Data Analysis Competition asked teams to turn Indonesian public opinion into insight: first app-store reviews of four e-wallets, then, in a five-hour final, a sentiment model for tweets about the government's PPKM COVID-19 policy, scored on hidden test data.

Contribution

In the selection phase the team crawled Google Play reviews of four e-wallets, grouped them into topics with keyword rules and trained a CNN-LSTM sentiment model on public labelled data plus 201 reviews labelled by hand.
The team presented the results in a Streamlit app with Plotly charts and later wrote them up as an IEEE conference paper.
In the final the team compared seven cleaning variants on balanced and unbalanced data across IndoBERTweet, IndoBERT, LSTM, BiLSTM, SVM and Naive Bayes, then submitted IndoBERTweet predictions.