Deep Learning · Sentiment Analysis · NLP

Reading the
Crisis in Real Time

How do Indonesian netizens feel about the economic crisis — and which deep learning model reads them best? A head-to-head study of Bi-LSTM versus IndoBERT on public comments from X (Twitter).

“Can a machine learn to feel the mood of a nation in crisis — from nothing but its tweets?”
01

The Dataset

1,539
Comments collected from X (Twitter)
2025-2026
Collection period
3 Classes
Sentiment categories
363 Positive 493 Neutral 683 Negative
Manual + AI
The data was labelled with the help of ChatGPT and manually verified to ensure correctness
Curated & Clean
No slang
No code-switching
No typos
Standard Indonesian only
02

The Contenders

CHALLENGER

Bi-LSTM

Bidirectional Long Short-Term Memory

A recurrent neural network that reads each comment word by word — forwards and backwards — building meaning through sequential memory. It learns from word embeddings trained on the dataset itself.

ArchitectureRecurrent (RNN)
ContextSequential, both directions
Pre-trainingNone — learns from scratch
CONTENDER

IndoBERT

Transformer for Indonesian Language

A transformer pre-trained on a massive Indonesian text corpus, then fine-tuned on the labelled comments. It reads the whole comment at once, weighing every word against every other word.

ArchitectureTransformer
ContextFull-sentence attention
Pre-trainingLarge Indonesian corpus
03

The Method

1
Collect
Scrape crisis-related comments from X
2
Clean
Remove noise, normalise text
3
Label
Hand-tag 3 sentiment classes
4
Tokenize
Embeddings vs WordPiece
5
Train
Fit both models, same split
6
Evaluate
Compare on the test set
04

Head-to-Head Results

IndoBERT Bi-LSTM
Accuracy
IndoBERT
83.0%
Bi-LSTM
73.0%
Precision (macro)
IndoBERT
83.0%
Bi-LSTM
74.7%
Recall (macro)
IndoBERT
83.0%
Bi-LSTM
73.0%
F1-Score (macro)
IndoBERT
83.0%
Bi-LSTM
73.7%

F1 is the metric to highlight if your classes are imbalanced — crisis comments usually skew negative, so accuracy alone can be misleading.

05

What Indonesia Felt

SENTIMENT
across all
comments
Based on 1,539 labelled comments
42.6%
Negative · 656 comments

Anxiety over prices, jobs and the rupiah dominated the conversation.

26%
Neutral · 400 comments

Factual reporting, news links and questions without clear emotion.

31.4%
Positive · 483 comments

Hope, support for policy responses, or optimism about recovery.

06

The Verdict

IndoBERT led on accuracy and F1 — but the better choice depends on what the project values most.

Pre-training on Indonesian gave the transformer an edge in understanding context and subtle wording. Yet the lighter Bi-LSTM trained faster and on far less compute, making it a practical option when resources are tight.

IndoBERT — Strengths

  • Higher accuracy & F1
  • Understands context & nuance
  • Benefits from pre-training

Bi-LSTM — Strengths

  • Faster, cheaper to train
  • Lighter to deploy
  • Simpler to interpret
Indonesia Economic Crisis · Sentiment Analysis
Clarence, Daniel, Joel, Kenzie · Deep Learning · 2026
Tools: Python · TensorFlow / PyTorch · HuggingFace Transformers · scikit-learn