Log anomaly detection in large-scale digital service systems is critical for reliable operations and effective maintenance, yet remains challenging due to the heterogeneous formats and semantic diversity of logs. Traditional methods often fail to correlate diverse log types or capture their contextual variability. To address this, we propose HEDGE, a unified framework that constructs a dynamic heterogeneous log graph to model fine-grained semantic and spatio-temporal relationships. HEDGE integrates a dual-tower SemanticFormer for semantic alignment with a graph-based model for temporal dependency learning, enabling more accurate anomaly detection. To further support operators in real-world maintenance scenarios, we extend our framework with HEDGE-QA, a log-based question answering system that retrieves relevant log traces and generates natural language responses. This enables intuitive interaction with system logs and provides interpretable explanations grounded in detection outcomes. Extensive experiments on public benchmark datasets demonstrate that our framework significantly outperforms state-of-the-art baselines in both detection accuracy and operational interpretability.
