Khmer News Headline Generation: A Comparative Evaluation of Extractive, Neural Network, and Transformer-Based Approaches

Abstract

Automatic headline generation remains a challenging task for low-resource languages such as Khmer due to the limited availability of annotated datasets and language-specific resources. This study investigates Khmer headline generation by comparing traditional extractive approaches, neural sequence-to-sequence, multilingual Transformer-based models. A Khmer news dataset of 65,393 article–headline pairs was collected from official Cambodian government websites and used to train and evaluate five models: Lead-1, TextRank, RNN-Attention, mT5, and mBART-50. Performance was evaluated using ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore metrics. Results demonstrate that zero-shot Transformer models fail catastrophically on Khmer (mT5 ROUGE-1: 22.99; mBART-50 ROUGE-1: 12.55), while LoRA fine-tuned Transformers significantly outperform all extractive and recurrent approaches. Fine-tuned mT5 achieved the highest ROUGE-2 (65.17), ROUGE-L (66.68), and BERTScore (92.29), while fine-tuned mBART-50 obtained the highest ROUGE-1 (79.28), representing superior performance across all metrics. These results indicate that multilingual pre-trained language models are effective for low-resource headline generation and suggest that task-specific adaptation is essential to activate latent multilingual knowledge in low-resource settings. This approach can be applied to other low-resource headline generation tasks, and the benchmark is relevant for future Khmer NLP research.