Modernizing FAQ Search with NLP: A Comparative Performance Assessment of Lexical and Semantic Models

With the exponential growth of online support portals and digital helpdesks, Frequently Asked Questions (FAQ) systems have become essential to reducing customer support load and improving user experience. Traditional information-retrieval approaches such as TF-IDF and BM25 rely heavily on lexical overlap and therefore perform poorly when users phrase questions differently, use synonyms, or switch between languages. To overcome these limitations, this research explores a semantic FAQ retrieval model based on Sentence-BERT (SBERT), which generates dense sentence embeddings to capture contextual meaning. The study compares SBERT against a classical TF-IDF baseline using retrieval metrics such as Recall@K, Mean Average Precision (MAP@K), and Mean Reciprocal Rank (MRR). Anticipated results indicate a significant improvement in ranking quality, especially in paraphrased and user-style queries. The paper also discusses computational efficiency, memory requirements, and applicability in real-world deployments. This work highlights how semantic modelling can substantially enhance FAQ retrieval accuracy and provide faster, more reliable information access across institutional and enterprise environments.