Rethinking App Review Mining with LLMs: A Human-Centric Perspective

Bachelor & Master Thesis

App review mining has long been used to support software maintenance and evolution by extracting insights from user feedback. Most existing approaches focus on the text of the review itself—classifying sentiment, clustering topics, or identifying feature requests—while largely ignoring the reviewer who authored it. This review-centric mindset limits our ability to fully interpret the nuances, intent, and context behind user feedback.

This project proposes a paradigm shift: using Large Language Models (LLMs) to mine app reviews from a human-centric perspective. The goal is to model not just what is written, but why it is written, by uncovering implicit reviewer traits such as expectations, domain knowledge, frustration tolerance, and usage context. LLMs are particularly well-suited for this task due to their ability to infer latent variables, perform reasoning over unstructured text, and adapt to instruction-style prompts. For example, prompting an LLM to "summarize this review from the perspective of a frustrated power user" can yield far richer insights than traditional sentiment classification.

In this thesis, you will explore how to use LLMs to simulate, infer, and transfer reviewer perspectives in app review datasets. By developing models that embed human-centric signals, your work can enable more empathetic and actionable understanding of user feedback—supporting not only technical issue resolution but also user experience design and community management.
 

Required knowledge:

Python programming and data processing, experience with NLP and LLMs, web crawling techniques, experience with mining software repositories

What you will do will include:

• Crawl and preprocess app review datasets (e.g., from Google Play or Apple App Store), and optionally collect related app metadata or update logs.
• Design prompt-based or fine-tuned LLM workflows to infer reviewer attributes and contextual signals.
• Build reviewer personas or intent models based on LLM-inferred insights.
• Compare the effectiveness of human-centric review mining with traditional sentiment/topic-based methods.
• Evaluate your approach via case studies or simulated developer usage scenarios.

For more details about this topic, please contact Dr. Shengcheng Yu (shengcheng.yu[at]tum.de).