Enhancing Log Abstraction with Semantic Variable Naming via Large Language Models

초록

Log abstraction is a crucial process in log analysis, aiming to extract dynamic variables and create log templates composed of static words. However, recent studies underline the significance of understanding the semantics of dynamic variables. A major limitation in contemporary semantic log abstraction research is the manual labeling of dynamic variables' semantics. To overcome this challenge, we introduce a novel framework that capitalizes on developers' inherent habit of assigning meanings to variables during code writing. In our framework, variable names within logging statements serve as semantic labels for dynamic variables. Given the unavailability of the source code for a system under analysis, our framework transforms log templates into Python logging statements and uses them as prompts for Large Language Models (LLM) to generate semantically meaningful variable names. This approach greatly improves the effectiveness of semantic log analysis and contributes significantly to the progress of automated log analysis.