Scholarlink

<p>To address the issue of semantic distortion in translation caused by structural differences between Urdu and English, a linguistically informed large-model-based Urdu-English neural machine translation method is proposed. Given that current general-purpose large models are predominantly trained on high-resource languages like English, they struggle to capture the specific grammar, complex morphological variations, and idiomatic expressions of Urdu. In this paper, first designs an adaptive multi-layer linguistic injection method that integrates lexical, syntactic, and semantic features into the large language model. These features include key linguistic markers such as tense, gender, and politeness; syntactic transformations between Subject-Verb-Object and Subject-Object-Verb structures; as well as idiomatic expressions and cultural differences. Next, a multi-knowledge integration and prompting technique is employed to dynamically adjust translations based on sentence complexity. Finally, Low-Rank Adaptation is used for efficient parameter fine-tuning, further enhancing translation performance. Experimental results demonstrate that this method significantly outperforms traditional neural machine translation systems, achieving a notable improvement in + 4.7 BLEU scores.</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

Linguistic knowledge injected into large language model for Urdu-English neural machine translation

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Muhammad Naeem Ul Hassan,
  • Zhengtao Yu,
  • Khalil Ullah,
  • Jian Wang,
  • Ying Li,
  • Shengxiang Gao,
  • Shuwan Yang,
  • Cunli Mao

摘要

To address the issue of semantic distortion in translation caused by structural differences between Urdu and English, a linguistically informed large-model-based Urdu-English neural machine translation method is proposed. Given that current general-purpose large models are predominantly trained on high-resource languages like English, they struggle to capture the specific grammar, complex morphological variations, and idiomatic expressions of Urdu. In this paper, first designs an adaptive multi-layer linguistic injection method that integrates lexical, syntactic, and semantic features into the large language model. These features include key linguistic markers such as tense, gender, and politeness; syntactic transformations between Subject-Verb-Object and Subject-Object-Verb structures; as well as idiomatic expressions and cultural differences. Next, a multi-knowledge integration and prompting technique is employed to dynamically adjust translations based on sentence complexity. Finally, Low-Rank Adaptation is used for efficient parameter fine-tuning, further enhancing translation performance. Experimental results demonstrate that this method significantly outperforms traditional neural machine translation systems, achieving a notable improvement in + 4.7 BLEU scores.