<![CDATA[The rapid growth of digital news volume has produced information overload, while exact keyword-matching retrieval remains vulnerable to synonymy and polysemy, causing relevant documents to be missed. Prior studies generally compare only two document representation methods on small-scale datasets, leaving a gap in controlled evaluations that jointly compare statistical, lexical-hybrid, and neural approaches on a large-scale news domain. This study compares three document representation methods, namely Term Frequency-Inverse Document Frequency (TF-IDF), TF-IDF with WordNet-based query expansion, and Sentence-BERT (all-MiniLM-L6-v2), for news document retrieval using Cosine Similarity on the BBC News Dataset (14,305 documents with a hierarchical Ground Truth of 5 Topics and 51 Subtopics). Ten queries were evaluated using Precision@K, Recall@K, F1-Score@K (K=5, 10, 20), and execution time. The results show that Sentence-BERT consistently outperforms the other methods with a Precision@5 of 0.84, compared to TF-IDF (0.56) and TF-IDF+WordNet (0.52), while TF-IDF remains the fastest at online query time (23.86 ms per query). WordNet expansion actually reduces precision and increases execution time without a proportional accuracy gain. These findings confirm that transformer-based semantic representations are superior for news domains with high lexical variation, while TF-IDF remains relevant for computationally constrained real-time systems]]>
Copyrights © 2026
