Sentiment Analysis Gojek.
A BiLSTM-based natural language processing (NLP) pipeline classifying Gojek Google Play Store user reviews with informal Indonesian slang normalization and 88.99% test accuracy.
Informal Indonesian NLP Challenges
High-volume applications like Gojek receive thousands of daily user reviews on Google Play Store characterized by heavy Indonesian slang, informal abbreviations (e.g., bgt, bener2, ga bsa), and colloquial syntax. Manual sentiment tracking is infeasible, while simplistic keyword-based matching fails to understand contextual nuances and compound sentences.
This project implements a Recurrent Neural Network (BiLSTM) pipeline tailored specifically for colloquial Indonesian app reviews, achieving 88.99% test accuracy and a 0.89 Macro F1-Score.
Text Preprocessing Pipeline & Slang Normalization
Downstream sequence representation relies on structured text sanitization:
└── 1. Case Folding → Standardize all text characters to lowercase
└── 2. Text Cleaning → Strip mentions (@), URLs, numbers, and irregular emojis
└── 3. Slang Word Normalization → Map informal Indonesian slang to standardized root forms
└── 4. Tokenization & Padding → Word sequence indexing and sequence length standardization
LSTM Architecture & Convergence Evaluation
The model combines an Embedding layer, Bidirectional LSTM layer, and Dropout regularization to capture contextual bidirectional word dependencies:
| Sentiment Class | Precision | Recall | F1-Score |
|---|---|---|---|
| Negative | 0.89 | 0.89 | 0.89 |
| Positive | 0.89 | 0.89 | 0.89 |
| Overall Accuracy | 88.99% (Macro F1: 0.89) | ||
The confusion matrix confirms balanced precision and recall across both sentiment polarities, minimizing missed user bug reports and complaints.