DLCoG: A Novel Framework for Dual-Level Code Comment Generation based on Semantic Segmentation and In-Context Learning
In large software projects with collaborative development, comprehensive code comments are crucial for code readability and maintainability. Code comments mainly include method comments and inline comments, where the former describes the functionality globally, and the latter describes the implementation details locally. Existing methods typically generate these two kinds of comments independently, which results in weak correlations between comments and code context, as well as high model inference costs due to long token inputs. To address these issues, we define the combination of inline comments and method comments as Dual-Level Code Comments. We formulate the novel task of automatically generate dual-level code comments based on given code and propose an approach named DLCoG (Dual-Level Code Comment Generation) to automate this task. First, a semantic segmentation model is proposed to identify code segments requiring inline comments. Next, we retrieve similar samples to adopting the in-context learning paradigm, which can enhance the generation quality of large language models (LLMs) in specific domains. Finally, the LLM is guided to generate dual-level code comments using a chain-of-thought prompts that first produce inline comments, followed by method comments. We manually constructed a high-quality clean Java dataset consisting of <Method, Method Comment, < Snippet, Inline Comments>*> based on open-source Java projects by (i) determining comments type and (ii) manually associating inline comments with their corresponding code. Here, * indicates that a single method comment may correspond to multiple inline comments. Then, we trained a multi-task learning model based on CodeBERT to automatically take the two steps needed, termed ICSA (Inline Comment Classification and Scope Association), thus to expand to a dataset containing 80k dual-level code comments. Experimental results on clean and extended datasets show that DLCoG outperforms all baselines by substantial margins. The contextual information provided by DLCoG can effectively improve the inline comments generated by LLM. Coordinated generation of dual-level comment also brings effective improvements to method comments, which is particularly significant when there are few contextual examples. Our work fills the long-standing gap in the dual-level code comment generation field, and can provide insights for future research in this direction. We provide open-source datasets and source code for future research.
Sun 27 AprDisplayed time zone: Eastern Time (US & Canada) change
16:00 - 17:30 | Summarisation, Natural Language GenerationResearch Track / Early Research Achievements (ERA) / Replications and Negative Results (RENE) at 205 Chair(s): Oscar Chaparro William & Mary, Coen De Roover Vrije Universiteit Brussel, Gema Rodríguez-Pérez Department of Computer Science, Mathematics, Physics and Statistics, University of British Columbia, Okanagan Campus | ||
16:00 10mTalk | Optimizing Datasets for Code Summarization: Is Code-Comment Coherence Enough? Research Track Antonio Vitale Politecnico di Torino, University of Molise, Antonio Mastropaolo William and Mary, USA, Rocco Oliveto University of Molise, Massimiliano Di Penta University of Sannio, Italy, Simone Scalabrino University of Molise | ||
16:10 10mTalk | CMDeSum: A Cross-Modal Deliberation Network for Code Summarization Research Track Zhifang Liao Central South University, Xiaoyu Liu Central South University, Peng Lan School of Computer Science and Engineering, Central South University, Changsha, China, Song Yu Central South University, Pei Liu Monash University | ||
16:20 10mTalk | CLCoSum: Curriculum Learning-based Code Summarization for Code Language Models Research Track Hongkui He South China University of Technology, Jiexin Wang South China University of Technology, Liuwen Cao South China University of Technology, Yi Cai School of Software Engineering, South China University of Technology, Guangzhou, China | ||
16:30 10mTalk | DLCoG: A Novel Framework for Dual-Level Code Comment Generation based on Semantic Segmentation and In-Context Learning Research Track Zhang Zhiyang , Haiyang Yang School of Computer Science and Engineering, Central South University, Qingyang Yan Central South University, Hao Yan Central South University, Wei-Huan Min Central South University, Zhao Wei Tencent, Li Kuang Central South University, Yingjie Xia Hangzhou Dianzi University | ||
16:40 10mTalk | Explaining GitHub Actions Failures with Large Language Models: Challenges, Insights, and Limitations Research Track Pablo Valenzuela-Toledo University of Bern, Universidad de La Frontera, Chuyue Wu University of Bern, Sandro Hernández University of Bern, Alexander Boll University of Bern, Roman Machacek University of Bern, Sebastiano Panichella University of Bern, Timo Kehrer University of Bern | ||
16:50 10mTalk | Large Language Models are Qualified Benchmark Builders: Rebuilding Pre-Training Datasets for Advancing Code Intelligence Tasks Research Track Kang Yang National University of Defense Technology, Xinjun Mao National University of Defense Technology, Shangwen Wang National University of Defense Technology, Yanlin Wang Sun Yat-sen University, Tanghaoran Zhang National University of Defense Technology, Yihao Qin National University of Defense Technology, Bo Lin National University of Defense Technology, Zhang Zhang Key Laboratory of Software Engineering for Complex Systems, National University of Defense Technology, Yao Lu National University of Defense Technology, Kamal Al-Sabahi College of Banking and Financial Studies Pre-print | ||
17:00 10mTalk | Extracting Formal Specifications from Documents Using LLMs for Test Automation Research Track Hui Li Xiamen University, Zhen Dong Fudan University, Siao Wang Fudan University, Hui Zhang Fudan University, Liwei Shen Fudan University, Xin Peng Fudan University, Dongdong She HKUST (The Hong Kong University of Science and Technology) | ||
17:10 6mTalk | Using Large Language Models to Generate Concise and Understandable Test Case Summaries Early Research Achievements (ERA) Natanael Djajadi Delft University of Technology, Amirhossein Deljouyi Delft University of Technology, Andy Zaidman TU Delft Pre-print | ||
17:16 6mTalk | Towards Generating the Rationale for Code Changes Replications and Negative Results (RENE) Francesco Casillo Università di Salerno, Antonio Mastropaolo William and Mary, USA, Gabriele Bavota Software Institute @ Università della Svizzera Italiana, Vincenzo Deufemia University of Salerno, Carmine Gravino University of Salerno | ||
17:22 8mTalk | Session's Discussion: "Summarisation, Natural Language Generation" Research Track | ||