RAG-ENHANCED ASPECT-BASED SENTIMENT ANALYSIS FOR MOBILE APPLICATION REVIEWS: A MULTIAGENT FRAMEWORK FOR DEVELOPER-ORIENTED INSIGHT GENERATION

ICLR 2026 Conference Submission19548 Authors

Keywords: Aspect-Based Sentiment Analysis, Retrieval-Augmented Generation, Multi-Agent Systems, LangGraph, LoRA Fine-Tuning, Mobile App Reviews, Developer Tooling, Code-Aware Solution Generation

TL;DR: RAG-enhanced ABSA system that filters, analyzes, and grounds mobile app reviews to generate developer-actionable bug insights, fixes, and test suggestions using LoRA-adapted LLMs and LangGraph multi-agent orchestration.

Abstract: Mobile app developers encounter a considerable challenge in understanding users' genuine perceptions of their programs. Manual analysis is not feasible, and coarse-grained sentiment labels are not effective, because the number of apps being analyzed and the number of app reviews are both in the millions, and the number of actionable engineering activities is expected to be small, thus a need to explore automated analysis. We introduce an Aspect-Based Sentiment Analysis system with RAG enhancements that directly correlate user complaints with developer fixes that may be made. Our end-to-end system consists of four contributions: (1) a contextual retrieval architecture that links complaints with a history of version and relevant documentation, with a dense retriever + RAG backbone; (2) resource-efficient adaptation with Low-Rank Adapters (LoRA) to LLaMA 3.1 8B, which dramatically reduces the footprint of deployable parameters, but does not affect predictive quality; (3) automated multi-agent orchestration (LangGraph) to refer developer queries to specialized agents helpful in relevant detection, ABSA inference, problem extraction and solution recommendation; and Our end-to-end system achieves good task-level performance (high sentiment accuracy and 82.3% aspect-extraction F1) a sampled set of 41,245 reviews of English apps and produces developer-actionable results, which could be checked through automated test-checks and by human developer study.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 19548