Enhancing Log Template Generation with Part-of-Speech Awareness for Log Analysis

Log analysis is a fundamental component of system operations, enabling tasks such as health monitoring, anomaly detection, and root cause analysis. As systems grow in scale and complexity, the volume of generated log data has grown exponentially, making manual analysis increasingly infeasible. To address this challenge, automatic log analysis methods have been widely adopted. A common preprocessing step in these methods is to structure unstructured logs by applying log templates. While various methods have been proposed for automatic log template generation, most fail to incorporate information about system-state-indicative words, which can degrade accuracy. In this paper, we propose a novel method that leverages part-of-speech (POS) information to identify system-state-indicative words during template generation. Experiments conducted on 16 benchmark datasets demonstrate that our method improves the accuracy of log template generation compared to existing methods across 10 datasets, while achieving comparable accuracy on the remaining 6.