Core Concepts
DivLog proposes an effective log parsing framework based on in-context learning (ICL) of large language models (LLMs) to generate log templates in a training-free manner, achieving state-of-the-art performance.
Abstract
Log parsing is crucial for automated log analysis, converting semi-structured logs into structured logs.
DivLog utilizes large language models (LLMs) for in-context learning to generate log templates without training.
DivLog samples diverse logs for candidate examples and selects the most relevant ones for each target log during parsing.
The proposed prompt format restricts the output to enhance the quality of generated log templates.
DivLog outperforms existing log parsers in accuracy and stability across 16 datasets.
Stats
DivLog는 16개의 널리 사용되는 로그 데이터셋에서 Parsing Accuracy, Precision Template Accuracy, Recall Template Accuracy를 평균적으로 98.1%, 92.1%, 92.9% 달성
Quotes
"DivLog achieves state-of-the-art performance in log parsing accuracy across various datasets."
"The proposed prompt format restricts the output to enhance the quality of generated log templates."