Empowering code review automation : a data-centric, multi-task-driven, and human-aware approach
Year: 2025
Type: dissertation
Abstract: The integration of artificial intelligence (AI), and more recently large language models (LLMs), into software engineering has enabled new levels of automation in traditionally human-intensive tasks. Among these, code review stands out as a crucial yet laborious activity, essential for ensuring code quality, maintainability, and team collaboration. While early tools focused on rule-based detection of superficial issues, the emergence of LLMs has paved the way for generative approaches capable of producing human-like review feedback. However, the transition toward trustworthy, human-aligned, and end-to-end automated code review remains hindered by four critical challenges: the rigidity of rule-based analyzers, the isolated modeling of interdependent review tasks, the noise and inconsistency of available training data, and the limited scope of evaluation metrics that overlook psychological and human-centered dimensions. This thesis addresses these limitations through four key contributions. First, it introduces a multi-step learning framework that uses pretrained language models to identify the type and location of review issues, serving as a flexible and interpretable alternative to static analysis. Second, it proposes a unified framework that jointly models code review subtasks, namely comment generation, code refinement, and quality estimation, via cross-task knowledge distillation and reinforcement learning, aligning generation behavior with downstream task utility. Third, it presents a data curation strategy that leverages LLMs to assess, filter, and rephrase noisy review comments, resulting in higher-quality training data that improves model performance. Finally, it develops SEWELL-CARE, a novel evaluation framework that captures not only technical effectiveness but also human aspects (e.g., developer well-being, autonomy, and cognitive load) in AI-assisted settings. Together, these contributions offer a holistic approach to code review automation that balances general-purpose language understanding with domain-specific adaptability, addresses both technical and human concerns, and provides actionable pathways toward the design of intelligent, efficient, and developer-friendly code review assistants.
DOI: 10.71781/11039
Author: Ben Sghaier, Oussama
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7104490331