Sentiment Analysis for Public Transport using DeBERTa

The exponential growth of metro systems as a backbone for urban mobility has generated a large volume of unstructured passenger feedback online. Traditional sentiment analysis methods, which assign a single polarity to an entire review, are insufficient for this domain, as passengers often express conflicting sentiments about different aspects of the service within a single comment. This paper proposes a comprehensive framework for Aspect-Based Sentiment Analysis (ABSA) tailored for metro service reviews. A novel domain-specific dataset of 5,000+ annotated reviews is curated by collecting data from public forums across multiple Indian cities. After rigorous comparative evaluation of transformer architectures including BERT, RoBERTa, and ALBERT, the fine-tuned DeBERTav3-base-absa-v1.1 achieved strong performance with a weighted F1-score of 0.91 on the curated and labelled dataset. The disentangled attention mechanism of the model proved to be highly effective in mapping nuanced sentiments to specific aspects such as cleanliness, crowd management, and safety. The research is materialized through a full-stack web application that provides metro authorities with granular, actionable insights through sentiment analysis and enables anonymous feedback contributions from passengers. By applying ABSA, we can convert raw passenger opinions into a structured knowledge base, providing a powerful tool for data-driven public transport management.