Dynamic alpha factor optimization for hybrid semantic similarity in french word sense disambiguation

Abstract

Word sense disambiguation (WSD) remains critical for French, where polysemy complicates semantic interpretation. Hybrid approaches combining lexical and semantic methods typically rely on static weighting parameters that fail to adapt to varying contexts. A hybrid architecture integrating fuzzy Jaccard similarity with sentence-bidirectional encoder representations from transformers (SBERT) embeddings is proposed. A machine learning-based dynamic weighting mechanism replaces the fixed alpha=0.7. A Ridge regression model predicts the optimal alpha based on five features: entropy, Jaccard-SBERT disagreement, gloss length, context richness, and SBERT confidence. The model was trained on 20 sentences and validated on 13 sentences from a dataset of 33 ambiguous French phrases. Dynamic alpha achieves a 7.6% reduction in mean absolute error (MAE) (0.2397 to 0.2216) and a 7.3% reduction in root mean square error (RMSE) (0.2475 to 0.2294) compared to fixed alpha=0.7. Per-sentence gains reach 10.0%. Statistical analysis confirms significance (Wilcoxon, p=0.00195) with a small to medium effect size (Cohen's d=0.300). Feature coefficients reveal that disagreement (+0.41) and entropy (+0.32) are the most influential predictors.