Why Log Semantics Matter More Than Sequence Data in Detecting Anomalies

This study explores how semantic information within log messages enhances anomaly detection, often outperforming models that rely solely on sequential or temporal data. Through Transformer-based experiments on public datasets, the authors find that event occurrence and semantic cues are more predictive of anomalies than sequence order. The research underscores the limits of current datasets and calls for new, well-annotated benchmarks to evaluate log-based anomaly detection more effectively, enabling models that fully leverage log semantics and event context.