Understanding Evolving Sentiment Dynamics in Indonesian Disaster-Related Tweets

Social media platforms generate massive amounts of unstructured data during crises. Most prior work treats sentiment analysis and topic detection as separate problems. This paper combines them — using aspect-sentiment analysis to capture more detailed, nuanced views that aggregate scoring tends to miss.

Background

Indonesia is one of the most disaster-prone countries in the world. During and after crisis events, Twitter (now X) becomes an informal real-time feed of public reaction — frustration with aid logistics, reports of unmet needs, gratitude for responders, and everything in between. Aggregating this signal can give authorities a faster, finer-grained picture of public sentiment than surveys or news coverage.

Previous work tended to tackle either what people are talking about (topic modeling) or how they feel (sentiment classification) — not both simultaneously. Aspect-based sentiment analysis (ABSA) addresses this by linking a sentiment to a specific target: not just "negative," but "negative toward government response efforts."

Dataset

We collected 4,506 tweets spanning July 2023 to March 2025, filtered to Indonesian-language posts mentioning major disaster events. Tweets were labeled across multiple aspect categories (infrastructure, response efforts, aid distribution, etc.) and three sentiment classes (positive, negative, neutral).

Methodology

We fine-tuned indobenchmark/indobert-base-p1 — a BERT-based model pre-trained on a large Indonesian corpus — separately for aspect classification and sentiment classification. Temporal analysis was layered on top to track how sentiment distributions shifted across the observation window.

The pipeline: tweet → aspect classifier (what topic?) → sentiment classifier (how do people feel about it?) → temporal aggregation (how does this change week over week?).

Results

• Aspect classifier: weighted F1-score of 0.73
• Sentiment model: F1-score of 0.81

One of the clearest findings: a sustained cluster of negative sentiment toward response efforts, particularly in the weeks immediately following disaster onset. This pattern signals public dissatisfaction — either with the speed of aid delivery or perceived gaps in coordination — that authorities could act on in near real-time.

Temporal analysis also revealed that sentiment tends to shift from predominantly negative (acute crisis phase) toward more neutral and mixed (recovery phase) as weeks pass — a pattern consistent with general disaster response literature but now quantifiable at scale.

Why It Matters

The combined aspect-sentiment approach gives emergency managers a richer signal than headline sentiment scores. Instead of knowing that "people are unhappy," they can see that "people are unhappy specifically about aid distribution in region X." That specificity is actionable.