Scholarlink

<p>Machine translation into sign language (or Sign Language Machine Translation—SLMT) presents a promising and emerging solution for overcoming barriers to information access and communication among Deaf individuals. However, the development of such systems is particularly challenging due to the low-resource nature of sign languages, including Brazilian Sign Language (Libras), which is the focus of this study. To address this challenge, we propose a hybrid machine translation approach for Brazilian Portuguese to Libras. Our solution combines text-to-gloss translation, integrating both rule-based and neural machine translation methods. The advantage of this strategy lies in using rule-based preprocessing for deterministic tasks, while a neural data-driven model handles more complex tasks, such as word-sense disambiguation, directional verbs, intensifier adverbs, and negative incorporation. This approach enhances both the fluency and naturalness of the translations. To support this approach, we constructed a new parallel corpus consisting of 70,000 sentence pairs (Brazilian Portuguese and corresponding Libras glosses). We also conducted computational experiments and user tests to compare our hybrid approach with the existing rule-based version of VLibras, which is currently deployed across thousands of Brazilian websites. The results indicate that our approach outperforms the current version, offering a promising direction for advancing SLMT solutions in low-resource contexts.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid translation for sign languages: combining rule-based and neural machine translation in a low-resource scenario

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Diego R. B. da Silva,
  • Manuella A. C. B. Lima,
  • Samuel de M. Moreira,
  • Virginia P. Campos,
  • Renan P. O. Costa,
  • Tiago M. U. de Araújo,
  • Rostand E. O. Costa,
  • Daniel F. L. de Souza,
  • Dilainne D. de Albuquerque

摘要

Machine translation into sign language (or Sign Language Machine Translation—SLMT) presents a promising and emerging solution for overcoming barriers to information access and communication among Deaf individuals. However, the development of such systems is particularly challenging due to the low-resource nature of sign languages, including Brazilian Sign Language (Libras), which is the focus of this study. To address this challenge, we propose a hybrid machine translation approach for Brazilian Portuguese to Libras. Our solution combines text-to-gloss translation, integrating both rule-based and neural machine translation methods. The advantage of this strategy lies in using rule-based preprocessing for deterministic tasks, while a neural data-driven model handles more complex tasks, such as word-sense disambiguation, directional verbs, intensifier adverbs, and negative incorporation. This approach enhances both the fluency and naturalness of the translations. To support this approach, we constructed a new parallel corpus consisting of 70,000 sentence pairs (Brazilian Portuguese and corresponding Libras glosses). We also conducted computational experiments and user tests to compare our hybrid approach with the existing rule-based version of VLibras, which is currently deployed across thousands of Brazilian websites. The results indicate that our approach outperforms the current version, offering a promising direction for advancing SLMT solutions in low-resource contexts.