ModernBERT bests 6 models in AAM sentiment analysis of 306k texts

A team of researchers from the University of New Mexico (Esrat Farhana Dulia, Amina Dhaher, Raiful Hasan, Syed Arbab Mohd Shihab) published a paper on arXiv (cs.CL/2606.20751) analyzing public sentiment toward Advanced Air Mobility (AAM). They scraped 306,009 human-generated texts from Reddit and Quora spanning 2008 to 2025. To pick the best classifier, they benchmarked seven models: VADER, TextBlob (lexicon-based); Logistic Regression, SVM (ML); LSTM (deep learning); BERT, ModernBERT (transformer). ModernBERT won on F1-score and was used to label the entire dataset.

Using Latent Dirichlet Allocation (LDA) per sentiment class, the team identified 20 latent topics and six major clusters of public concern. Workforce and skill development accounted for 25.29% of discussion, followed by regulation and compliance (24.64%), technical performance of drones (20.99%), military/geopolitics/defense (14.58%), safety and operational risks (8.51%), and noise and disturbance (5.98%). The paper proposes actionable strategies—like transparent regulation, noise mitigation standards, and public awareness campaigns—to improve public acceptance and ultimately the commercial viability of AAM deployments.