PCN-WSD

Academic paper and open-source implementation of a fixed-dimensional hierarchical Predictive Coding Network for Turkish Word Sense Disambiguation, trained with purely local Hebbian rules and zero backpropagation. On clean embeddings it doesn't beat k-NN; its contribution is graceful degradation under Gaussian noise — a reproducible crossover at σ=0.7 (54% vs 35% accuracy) where iterative inference reconstructs corrupted inputs from learned priors. Includes a documented negative result (dimensional-reduction hybrid at ~10%) and bridges classical cognitive theory (Tolman, Spearman) with modern Free-Energy / PC theory.