Classification of Greek News Articles by Text Type Using Open Large Language Models

The automatic classification of Greek news articles presents unique challenges due to the relatively low availability of annotated datasets and the under-representation of the Greek language in large-scale language models. While previous studies have primarily focused on topic-based classification, this work investigates the classification of articles based on their text type (e.g. articles, letters, reviews, interviews, etc.).We evaluate the zero-shot classification capabilities of open large language models (LLMs), including LLaMA 3.1, Aya, and the Greek-LLM Meltemi, comparing their performance against a supervised approach that employs a fine-tuned BERT baseline for Greek. Our experiments explore various prompt engineering techniques that incorporate structured output and LLM-based reasoning, to improve classification performance.Results indicate that while generic LLMs show some promise, their classification errors are often skewed by prediction biases, with confusion matrices revealing difficulties in differentiating between certain categories. Moreover, the fine-tuned Greek BERT model significantly outperforms the zero-shot LLMs across all evaluation metrics, reinforcing the importance of domain-specific training for low-resource languages. Our findings highlight the need for further fine-tuning and dataset expansion to improve LLM classification reliability in Greek text-processing tasks.