EV Sentiment Analysis (Undergraduate Thesis)

My undergraduate thesis: comparing CNN, LSTM, BiLSTM, and a hybrid CNN+BiLSTM-Attention model to classify Indonesian sentiment toward electric vehicles from Platform X, deployed as a live Streamlit demo.

Preprocessing: cleaning (URLs, mentions, hashtags, emoji, non-alphanumeric characters), slang normalization against a custom slang dictionary, case folding, and stemming via Sastrawi. Text is converted to fixed-length sequences (128 tokens, padded) through a trainable Keras embedding layer, with no pretrained word vectors.

Training

Each of the four architectures was trained at three epoch budgets (50, 100, 150) with the Adam optimizer, categorical cross-entropy loss, class weights to counter label imbalance, and EarlyStopping + model checkpointing to avoid overfitting and keep the best-performing weights.

Results

The hybrid model won on every metric. Single-direction LSTM effectively failed: it collapsed into predicting the majority class and never recovered, performing no better than a coin flip (AUC 0.50), regardless of how long it trained.

What the epoch sweep revealed (Table 4.6) was more interesting than the headline numbers. CNN's F1 score actually fell as training went from 50 → 150 epochs (0.744 → 0.721), a textbook overfitting curve. BiLSTM stayed flat across all three (a stable plateau). The hybrid model was the only one that kept improving with more training (0.751 → 0.760), meaning it had learning capacity the other architectures didn't, and it was chosen at 150 epochs for deployment.

Where it still struggles: even the winning hybrid model's biggest source of error is confusing Positive with Neutral. Its recall on the Positive class alone is just 0.63. Distinguishing mild positivity from neutral commentary turned out to be the hard part of this task, not separating clearly positive from clearly negative sentiment.

Live Demo

The best-performing model, tokenizer, and label encoder were exported and wired into a Streamlit web app. Type in a review, hit analyze, and the app runs the same preprocessing pipeline used in training before returning a prediction with a confidence score.

Worth being precise about: the research and evaluation, meaning every number above, comes from the custom-trained TensorFlow/Keras CNN+BiLSTM-Attention model described in the thesis. The public demo app, for practical deployment reasons, calls a pretrained HuggingFace sentiment model (PyTorch) instead of loading the custom Keras model directly. The UI and pipeline shown are the same ones built for the thesis; the inference backend behind the public link is a lighter swap to keep a free-tier deployment fast and reliable.

Tech Stack

Links

• Source: github.com/fatihaljabar/project-TA
• Live demo: project-ta-fatih.streamlit.app
• Published journal article: doi.org/10.36040/jati.v10i1.17094