A Hybrid LLM and NLP Framework for Aspect-Based Sentiment Analysis of Mobile Application Reviews

Abstract

Feedback from mobile app reviews helps uncover how applications perform, feel, and function in real-world usage. However, interpreting large volumes of unstructured textual data remains a significant challenge. A hybrid approach combining traditional Natural Language Processing techniques with Large Language Models forms the foundation of this methodology. The process begins with automated data collection followed by text preprocessing to improve data quality. Aspect extraction is primarily performed using LLM-based semantic understanding, with fallback statistical methods ensuring robustness and completeness. Sentiment classification relies on lexicon-based approaches to determine emotional polarity efficiently. Summarization is achieved using transformer-based models capable of generating concise and meaningful representations of user feedback. Visualization techniques, such as word clouds and sentiment distribution charts, highlight frequently discussed topics and overall sentiment trends. By integrating statistical and semantic techniques, the system improves resilience and contextual accuracy compared to standalone approaches. Experimental results demonstrate effective identification of key aspects such as performance issues and functional defects, along with their associated sentiment polarity. These insights support developers in aligning application improvements with actual user feedback.