<![CDATA[News classification is critical for organizing digital content and enhancing user engagement. Static embeddings such as GloVe, however, often fail to capture dynamic contextual relationships, limiting classifier performance. This study investigates the integration of frozen DistilBERT embeddings used strictly as feature extractors into a BiLSTM news‐classification pipeline to harness context‐aware representations while preserving computational efficiency. Experiments were conduccted on the Fancyzhx/ag_news dataset, which contains 127,700 samples, using a controlled 70/30 train–test split, comparing three BiLSTM variants: baseline (no pretrained embeddings), GloVe‐BiLSTM, and DistilBERT‐BiLSTM. Model architectures and hyperparameters are held constant to ensure a fair evaluation. Experimental results demonstrate that DistilBERT‐BiLSTM achieves 93.2 % accuracy, outperforming GloVe‐BiLSTM by 1.3% and the baseline by 2.5%. UMAP visualizations reveal more distinct semantic clusters with DistilBERT embeddings, and token‐level heatmaps confirm sharper intra‐sentence focus on domain‐specific terms. These findings contribute to informatics research by demonstrating the value of frozen transformer embeddings for lightweight yet high-performing text classification, with practical applications in information retrieval, content moderation, and real-time news analytics.]]>
Copyrights © 2026
