Temporal Fusion Strategies for NER in Historical Texts

This study investigates how temporal metadata can be structurally embedded into Named Entity Recognition (NER) models to address the challenge of entity drift in historical texts. The authors systematically evaluate lightweight fusion strategies, including cross-attention, adapters, and concatenation, within Transformer-based architectures.

• Experiments utilize both absolute and relative temporal representations injected via early or late fusion mechanisms.
• Evaluations are conducted on French and German historical datasets.
• Late fusion strategies demonstrate more robust and temporally generalizable performance, particularly in early and noisy periods.

The findings suggest that specific structural embeddings of temporal information significantly improve the ability of language models to reason about diachronic contexts.