25 lines
2.1 KiB
Markdown
25 lines
2.1 KiB
Markdown
collapsed:: true
|
||
type:: [[REVIEWS]]
|
||
tags::
|
||
year:: 2025
|
||
venue:: [[ESEM]]
|
||
full-title:: When Retriever Meets Generator: A Joint Model for Code Comment Generation
|
||
date-start:: [[26-06-2025]] - 14:49
|
||
date-submitted::
|
||
external-links::
|
||
status:: [[DONE]]
|
||
deadline-submission::
|
||
file:: [[@ESEM25_paper_260]]
|
||
parent::
|
||
todoist:: https://app.todoist.com/app/task/260-when-retriever-meets-generator-a-joint-model-for-code-comment-generation-6c6VQJ4qGr8rx7mc
|
||
|
||
- ### [[Highlights]]
|
||
- ### [[Comments]]
|
||
- Summary: Ther paper presents RAGSum, an approach based on retrieval-augmented generation to automatically generate code comments. The approach is compared with different baselines with respect to standard metrics and different benchmark datasets. The performed experiments shows improvements and shows that different components of the approach contributing to acheve the measured performance.
|
||
- COMMENTS: The paper tackles an important problem in software engineering, namely the generation of meaningful and accurate comments for source code snippets. While the motivation and the general idea of integrating retrieval and generation using a unified CodeT5 backbone is promising, some sections of the paper would benefit from improved clarity:
|
||
- In the methodological discussion, there is a reference to retrieving comments from similar source code snippets. The paper should better support the hypotesis whether code similarity reliably implies comment similarity, as this assumption underpins the retrieval component’s effectiveness.
|
||
- When reporting results (e.g., RAGSum's improvements of 4.1%, 5.31%, etc. across different metrics), the paper lacks a qualitative discussion that contextualizes these gains. Are such improvements statistically significant? Do they translate into meaningfully better developer experience or understanding?
|
||
- An important point is raised regarding the potential mismatch between metrics like ROUGE-L and true semantic equivalence. This issue deserves further investigation (not in this paper indeed), especially since some generated comments may be semantically adequate yet penalized due to surface-level differences with the reference.
|
||
-
|
||
- ---
|
||
- |