A temporal fusion model and automated annotation framework for computing Li across time

Abstract Ancient Chinese word meanings undergo diachronic change that conventional models pretrained on static modern corpora fail to capture. We propose the temporal fusion pretrained language model (TFPLM), which progressively integrates continuous temporal embeddings and temporal self-attention through token- and layer-wise gating while preserving contextual semantics. To support diachronic sense research, we propose a human-in-the-loop annotation framework that combines prototype matching, discriminative classification, and LLM-based arbitration with multi-stage confidence calibration. Using this framework, we construct the first large-scale diachronic corpus for the core character “Li”, containing more than 100,000 high-quality sense instances across six historical periods. Experiments show that TFPLM outperforms baseline models on Ancient Chinese word sense disambiguation. The automatically annotated “Li” corpus reveals diachronic shifts in sense distribution, providing interpretable evidence for the semantic evolution of ritual culture and a reusable framework for computational diachronic semantics in digital humanities.