CMER: A Context-Aware Approach for Mining Ethical Concern-related App Reviews

TL;DR

This paper introduces CMER, a novel, scalable method combining NLI and LLMs to automatically identify ethical concern-related app reviews, addressing the challenge of domain-specific language and overshadowing feedback.

Contribution

The study proposes a new context-aware approach that leverages NLI and LLMs to extract ethical concern reviews without labeled data, improving detection accuracy over previous methods.

Findings

Successfully extracted 2,178 additional privacy/security reviews

Demonstrated effectiveness of combining NLI and LLMs

Outperformed keyword-based approaches in identifying ethical concerns

Abstract

With the increasing proliferation of mobile applications in our daily lives, the concerns surrounding ethics have surged significantly. Users communicate their feedback in app reviews, frequently emphasizing ethical concerns, such as privacy and security. Incorporating these reviews has proved to be useful for many areas of software engineering (e.g., requirement engineering, testing, etc.). However, app reviews related to ethical concerns generally use domain-specific language and are typically overshadowed by more generic categories of user feedback, such as app reliability and usability. Thus, making automated extraction a challenging and time-consuming effort. This study proposes CMER (A \underline{C}ontext-Aware Approach for \underline{M}ining \underline{E}thical Concern-related App \underline{R}eviews), a novel approach that combines Natural Language Inference (NLI) and a decoder-only (LLaMA-like) Large Language Model (LLM) to extract ethical concern-related app reviews at scale. In CMER, NLI provides domain-specific context awareness by using domain-specific hypotheses, and the Llama-like LLM eliminates the need for labeled data in the classification task. We evaluated the validity of CMER by mining privacy and security-related reviews (PSRs) from the dataset of more than 382K app reviews of mobile investment apps. First, we evaluated four NLI models and compared the results of domain-specific hypotheses with generic hypotheses. Next, we evaluated three LLMs for the classification task. Finally, we combined the best NLI and LLM models (CMER) and extracted 2,178 additional PSRs overlooked by the previous study using a keyword-based approach, thus demonstrating the effectiveness of CMER. These reviews can be further refined into actionable requirement artifacts.