Understanding Public Perception of Crime in Bangladesh: A Transformer-Based Approach with Explainability

Social media is a valuable data source for understanding public safety concerns: The study demonstrates that analyzing comments on social media platforms can reveal how public perception of crime evolves over time, providing policymakers with real-time insights into community concerns that might inform crime prevention strategies.

Advanced AI models can work effectively with non-English languages: The transformer-based model achieved 97% accuracy in classifying sentiment in Bangla-language comments, showing that sophisticated AI tools are not limited to English-speaking contexts and can be deployed in low-resource language communities to understand local public opinion.

Explainability is essential for trustworthy AI applications: The researchers specifically employed explainable AI techniques to identify which features most influenced the model's sentiment classifications, addressing a key responsible AI principle—ensuring that automated decision-making systems can be understood and audited rather than operating as "black boxes."

High accuracy alone doesn't guarantee responsible deployment: While the model achieved 97% accuracy, the emphasis on explainability suggests the researchers recognized that accuracy metrics are insufficient; understanding why the model makes decisions is crucial before using such tools to inform real-world policy decisions affecting public safety.

Responsible AI requires careful dataset curation: The study involved creating a newly curated dataset of 28,528 comments, highlighting that responsible AI applications depend on thoughtful data collection and preparation, particularly when working with sensitive topics like crime where biased or unrepresentative data could lead to flawed policy recommendations.