Embedding-Based Machine Learning Approach for Automatic Classification of Turkish News Articles

Abstract

In this study, an automatic text classification approach for Turkish news articles is presented. The savasy/ttc4900 dataset from HuggingFace, consisting of seven news categories, was used. News texts were converted into 768-dimensional vector representations using the embeddinggemma model on the Ollama framework. These embeddings were then used to evaluate the performance of several machine learning algorithms. Seven models were tested: Support Vector Classifier (SVC), Logistic Regression, Multilayer Perceptron, K-Nearest Neighbors, Random Forest, Gaussian Naive Bayes, and Decision Tree. Model performance was assessed using accuracy, precision, recall, and F1-score metrics. Results showed that SVC and Logistic Regression achieved the highest accuracy in the high-dimensional embedding space. The findings demonstrate that embedding-based representations offer strong discriminative capability for Turkish news classification and that deep learning–derived vector embeddings can be effectively combined with traditional machine learning methods. These results emphasize the importance of vectorized text representations in natural language processing research.