Hybrid Semantic Retrieval: Augmenting Weighted TF-IDF with BERT for Enhanced Question Answering

This paper introduces a refined semantic search paradigm that significantly improves the precision and relevance of information retrieval, particularly within question-answering systems. Our novel approach integrates a meticulously designed weighted TF-IDF scheme with the contextual understanding capabilities of the BERT natural language model. By intuitively emphasizing "questionable spans" in documents via the weighted TF-IDF and simultaneously leveraging BERT to capture nuanced semantic meanings, our model effectively bridges the gap left by traditional lexical methods. We demonstrate through rigorous experiments on question-answering datasets that this hybrid strategy substantially outperforms existing semantic search techniques. The proposed model is designed for efficient scaling across large datasets, marking a considerable advancement in developing highly performant and semantically aware search engines for complex information landscapes.