LSTM-Based Customer Feedback Classification in Indonesian
A study of Indonesian customer-feedback classification using FastText embeddings, a bidirectional LSTM, and cross-validation.
Problem
Can a supervised sequence model organize Indonesian customer feedback into four operational categories?
The project implements a deep learning pipeline for:
- Akun (Account) - Issues related to user accounts
- Aplikasi (Application) - Problems with the app itself
- Layanan Pelanggan (Customer Service) - Customer support related
- Transaksi (Transaction) - Issues with payments and transactions
Key Features
- 🧹 Text Preprocessing - Cleaning, normalization, and tokenization for Indonesian text
- 📊 FastText Embeddings - Pre-trained Indonesian word vectors (300 dimensions)
- 🧠 LSTM Architecture - Bidirectional processing with avg/max pooling
- 📈 5-Fold Cross Validation - Robust model evaluation
- 🎯 Reported result - The notebook reports approximately 85% accuracy; the evaluation protocol and class-level results should be read from the notebook.
Product value and limits
The pipeline shows how Indonesian support text can be mapped to categories that a service team can act on. The result is specific to the dataset, preprocessing, split, and labels used in the notebook. It does not establish performance on new support channels or justify automatic routing without monitoring and human review.
Interactive Notebook
The complete experiment notebook with code, outputs, and visualizations:
Loading notebook...