Deep Assessment of Code Review Generation Approaches: Beyond Lexical Similarity

Abstract

Code review is a standard practice for ensuring the quality of software projects, and recent research has focused extensively on automated code review. While significant advancements have been made in generating code reviews, the automated assessment of these reviews remains less explored, with existing approaches and metrics often proving inaccurate. Current metrics, such as BLEU, primarily rely on lexical similarity between generated and reference reviews. However, such metrics tend to underestimate reviews that articulate the expected issues in ways different from the references. In this paper, we explore how semantic similarity between generated and reference reviews can enhance the automated assessment of code reviews. We first present a benchmark called \textit{GradedReviews}, which is constructed by collecting real-world code reviews from open-source projects, generating reviews using state-of-the-art approaches, and manually assessing their quality. We then evaluate existing metrics for code review assessment using this benchmark, revealing their limitations. To address these limitations, we propose two novel semantic-based approaches for assessing code reviews. The first approach involves converting both the generated review and its reference into digital vectors using a deep learning model and then measuring their semantic similarity through Cosine similarity. The second approach generates a prompt based on the generated review and its reference, submits this prompt to ChatGPT, and requests ChatGPT to rate the generated review according to explicitly defined criteria. Our evaluation on the \textit{GradedReviews} benchmark indicates that the proposed semantic-based approaches significantly outperform existing state-of-the-art metrics in assessing generated code review, improving the correlation coefficient between the resulting scores and human scores from 0.22 to 0.47.