LogGen: Integrating traditional model and LLM with code analysis for precise log generation

Abstract

Logging is a crucial method for the runtime monitoring and measurement of program states. However, in large-scale projects, variations in developer experience lead to significant differences in the key information within log statements, making it challenging to maintain uniform log formats. Therefore, we propose LogGen, a log generation framework that combines a traditional model with Large Language Models (LLMs). It learns from existing projects and automatically inserts log statements into new code, addressing the challenge of LLMs struggling to pinpoint accurate log positions. By leveraging LLM code summarization and integrating function-level semantic information, LogGen constructs more robust features, enabling the traditional model to achieve high accuracy in log positioning. Additionally, LogGen extracts log styles and function call chains from projects to create a knowledge database. LLMs can produce more precise log statements by combining our knowledge with the Retrieval-Augmented Generation (RAG) framework. We evaluate LogGen on long-term developed Java and C++ projects, showing that it surpasses the state-of-the-art log generation methods by 41.6% in F1 of log position and 36.2% in text BLEU-4 score, exhibiting stronger robustness against code and knowledge changes.