LogVm: Variable Semantics Miner for Log Messages

Abstract

Modern automated log analytics rely on log events without paying attention to variables. However, variables, such as the return code (e.g., “404”) in logs, are noteworthy for their specific semantics of system running status. To unlock the critical bottleneck of mining such semantics from log messages, this study proposes LogVM with three components: (1) an encoder to capture the context information; (2) a pair matcher to resolve variable semantics; and (3) a word scorer to disambiguate different semantic roles. The experiments over seven widely-used software systems demonstrate that Log Vm can derive rich semantics from log messages. We believe such uncovered variable semantics can facilitate downstream applications for system maintainers.