LI-HDSRM: Linguistically-Informed Hierarchical Dynamic Semantic Representation Model — Source Code and Supplementary Materials

Source code and supplementary materials for the paper: "LI-HDSRM: A Linguistically-Informed Hierarchical Dynamic Semantic Representation Model Integrating Cognitive Linguistic Theory with Neural Computation" (Language Resources and Evaluation, under review). LI-HDSRM is a hybrid framework integrating linguistic theory with neural computation for semantic representation. The model comprises three core modules: (1) Sememe Decomposition Module with Prototype-Contrastive Learning, (2) Syntactic Encoding Module with Relation-Typed GAT and JK-Net aggregation, (3) Context Modeling Module with Relevance-Gated Sparse Attention, and (4) Multi-Task Learning Framework with Uncertainty-Based Dynamic Task Weighting. This repository includes: model source code (PyTorch), training and evaluation scripts, ablation study configurations, hyperparameter settings, and publication-quality figures.