Automating modern code review processes with code similarity measurement

Context:

Modern code review is a critical component in software development processes, as it ensures security, detects errors early and improves code quality. However, manual reviews can be time-consuming and unreliable. Automated code review can address these issues. Although deep-learning methods have been used to recommend code review comments, they are expensive to train and employ. Instead, information retrieval (IR)-based methods for automatic code review are showing promising results in efficiency, effectiveness, and flexibility.

Objective:

Our main objective is to determine the optimal combination of the vectorization method and similarity to measure what gives the best results in an automatic code review, thereby improving the performance of IR-based methods.

Method:

Specifically, we investigate different vectorization methods (Word2Vec, Doc2Vec, Code2Vec, and Transformer) that differ from previous research (TF-IDF and Bag-of-Words), and similarity measures (Cosine, Euclidean, and Manhattan) to capture the semantic similarities between code texts. We evaluate the performance of these methods using standard metrics, such as Blue, Meteor, and Rouge-L, and include the run-time of the models in our results.

Results:

Our results demonstrate that the Transformer model outperforms the state-of-the-art method in all standard metrics and similarity measurements, achieving a 19.1% improvement in providing exact matches and a 6.2% improvement in recommending reviews closer to human reviews.

Conclusion:

Our findings suggest that the Transformer model is a highly effective and efficient approach for recommending code review comments that closely resemble those written by humans, providing valuable insight for developing more efficient and effective automated code review systems.