Hybrid Deep Learning for Sentiment Analysis: A Bibliometric Perspective on Research trends And Gaps

Abstract

This bibliometric study examines hybrid deep learning architectures for sentiment analysis by analysing 18,373 publications (2010-2025) from Scopus through descriptive analysis, citation metrics, and keyword co-occurrence methods. Hybrid architectures mix different deep learning components, with ensemble methods and transformer-based models emerging as the two primary approaches researchers use. Ensemble methods showed up most frequently at 35%, but transformers grew fastest, increasing from 2.4% to 20.7% during the study period. China published the most with 5,909 papers, India came in second with 3,949, and the United States had 1,523. Across all the research, 12 main themes stood out. Ensemble methods appeared 3,675 times and transformers appeared 3,200 times across the papers reviewed. Among specific techniques, BERT variants grew from nearly invisible (0.1%) to 10.4%, while attention mechanisms climbed from 1.9% to 14.0%. Some areas got much less attention, though. Low-resource languages grew only 2.1%, domain adaptation just 0.3%, and real-time processing 2.4%, all falling far behind mainstream methods. These findings document how hybrid deep learning in sentiment analysis has evolved while identifying underexplored areas that present opportunities for methodological innovation and broader practical application.