Matan P, Velvizhy P
Department of Computer Science and Engineering, College of Engineering Guindy, Anna University, Chennai, Tamil Nadu, 600025, India.
Sci Rep. 2025 Jun 30;15(1):20348. doi: 10.1038/s41598-025-03290-3.
Machine translation plays a critical role in expanding access to information across diverse languages and cultures. For children's literature, there is a need for translation models that can preserve both linguistic accuracy and emotional sensitivity. However, existing automated systems often struggle with the adaptations required for young readers. This study addresses this gap by developing a novel English-to-Tamil translation model for children's stories, combining the Universal Networking Language (UNL) for semantic representation with emotional paraphrasing techniques. Our approach uses a neuro-symbolic AI framework, specifically integrating the T5 transformer and few-shot learning, allowing effective model adaptation with minimal data. Evaluation with BiLingual Evaluation Understudy (BLEU), Translation Error Rate (TER), and Metric for Evaluation of Translation with Explicit Ordering (METEOR) scores (0.8978, 0.15, and 0.8869 respectively) highlights the model's high performance in maintaining both accuracy and contextual sensitivity. These metrics underscore the system's capability to deliver culturally relevant and child-appropriate translations. This research contributes to machine translation by bridging neural and symbolic methods, providing an adaptable, low-resource solution that supports cross-cultural understanding and accessible content for young readers.
机器翻译在跨越不同语言和文化扩大信息获取方面发挥着关键作用。对于儿童文学来说，需要能够兼顾语言准确性和情感敏感度的翻译模型。然而，现有的自动化系统往往难以满足年轻读者所需的适应性要求。本研究通过开发一种用于儿童故事的新型英语到泰米尔语翻译模型来填补这一空白，该模型将用于语义表示的通用网络语言（UNL）与情感释义技术相结合。我们的方法使用了一种神经符号人工智能框架，具体整合了T5变换器和少样本学习，能够以最少的数据实现有效的模型适应性调整。通过双语评估辅助工具（BLEU）、翻译错误率（TER）和显式排序翻译评估指标（METEOR）得分（分别为0.8978、0.15和0.8869）进行的评估突出了该模型在保持准确性和上下文敏感度方面的高性能。这些指标强调了该系统能够提供与文化相关且适合儿童的翻译。这项研究通过弥合神经方法和符号方法之间的差距，为机器翻译做出了贡献，提供了一种适应性强、资源需求少的解决方案，支持跨文化理解并为年轻读者提供易于获取的内容。
