Scholarlink

<p>To address entity-alignment challenges in aquaculture arising from cross-lingual naming divergence and vernacular ambiguity, while avoiding catastrophic forgetting in conventional fine-tuning of large language models, we proposed MoE-Aquatic, a mixture-of-experts framework for cross-lingual entity alignment. An Internal Prompt mechanism encodes taxonomic path knowledge, while language and domain experts cooperate to perform cross-lingual alignment. On ASFIS_2025 and DBP15K (ZH-EN/JA-EN), Hits@1 reaches 0.881, 0.821, and 0.827. Ablation experiments demonstrate the necessity of each module, with removal of backbone or expert components leading to significant performance degradation. MoE-Aquatic effectively mitigates naming ambiguity and hierarchical deficiencies, providing a novel approach for accurate cross-lingual entity alignment in aquaculture.</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

MoE-Aquatic: research on a mixture-of-experts for cross-lingual entity alignment method in aquaculture

  • 期刊论文

  • 发表日期：

    2025年12月11日

查看全文
  • Chun Zhang,
  • Sijia Zhang,
  • Zhenglin Li

摘要

To address entity-alignment challenges in aquaculture arising from cross-lingual naming divergence and vernacular ambiguity, while avoiding catastrophic forgetting in conventional fine-tuning of large language models, we proposed MoE-Aquatic, a mixture-of-experts framework for cross-lingual entity alignment. An Internal Prompt mechanism encodes taxonomic path knowledge, while language and domain experts cooperate to perform cross-lingual alignment. On ASFIS_2025 and DBP15K (ZH-EN/JA-EN), Hits@1 reaches 0.881, 0.821, and 0.827. Ablation experiments demonstrate the necessity of each module, with removal of backbone or expert components leading to significant performance degradation. MoE-Aquatic effectively mitigates naming ambiguity and hierarchical deficiencies, providing a novel approach for accurate cross-lingual entity alignment in aquaculture.