| Resumen |
This study addresses sentiment preservation in multilingual text summarization by proposing a reinforcement learning framework with a composite reward function. The approach combines sentiment consistency, content retention, fluency, and diversity metrics, leveraging an adapted mT5 model with language-specific processing. Evaluation across five languages shows 12–15% improvement in sentiment preservation, with particularly strong results (34% reduction in errors) for morphologically complex languages like Finnish. The framework demonstrates practical value in e-commerce and mental health applications, achieving 22 higher customer satisfaction and 18% better emotional tone detection. While effective, computational overhead and challenges with low-resource languages remain for future optimization. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026. |
