Hybrid Statistical-Neural Framework for Context-Aware Word Sense Disambiguation Using Multi-Head Attention and TF-IDF Fusion

Understanding semantic nuances in natural language requires models that can bridge traditional statistical methods with modern deep learning. In this paper, we introduce the architecture AMBER (Attention-based Multi-head Bidirectional Enhanced Representations), a fusion of statistical embeddings (Word2Vec and TF-IDF) and neural contextualization. AMBER proposesulti-head attention, residual learning, and contextual enhancement, which enhances disambiguation in the case of polysemous words. Our end-to-end experiments on tasks of word sense assessment and contextual similarity show better separations of semantically different usages and better than single isolated and single context embedding strategies. This blended approach to statistical and neural semantics fills a gap between and provides stable and understandable embeddings to real-life language-related tasks.