Scholarlink

<p>The performance of retrieval-augmented generation (RAG) systems is heavily dependent on the quality of embedding models, which serve as the interface between knowledge repositories and generative models. Current fine-tuning approaches for embedding models rely predominantly on synthetic data generation and simple contrastive learning frameworks, which often lack semantic depth and fail to capture complex domain-specific relationships, particularly in morphologically rich languages such as Greek and in specialized domains such as legal text processing. In this paper, we introduce KG-Tune, a knowledge graph-driven fine-tuning methodology, which is specifically tailored for RAG applications. KG-Tune addresses the limitations of existing methodologies by leveraging structured knowledge extraction through a systematic two-phase training framework that incorporates (i) systematic extraction of entities, relationships and atomic facts from domain-specific knowledge graphs to generate semantically grounded training instances, and (ii) hard negative mining to enhance discriminative capabilities. Experimental evaluation on four heterogeneous Greek datasets demonstrates that the proposed methodology achieves considerable improvements, with top-3 accuracy gains ranging from 5.4% to 18.0% and an average performance improvement of 11.1% compared to traditional fine-tuning approaches. An ablation study further confirms the significance of these improvements.</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

KG-Tune: Improving embedding models for domain-specific RAG via knowledge graph-driven fine-tuning

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Athanasios Koursaris,
  • Ioannis E. Livieris,
  • Alexandra Apostolopoulou,
  • Konstantinos Kanaris,
  • Dimitris Tsakalidis,
  • George Domalis,
  • Nikos Karacapilidis

摘要

The performance of retrieval-augmented generation (RAG) systems is heavily dependent on the quality of embedding models, which serve as the interface between knowledge repositories and generative models. Current fine-tuning approaches for embedding models rely predominantly on synthetic data generation and simple contrastive learning frameworks, which often lack semantic depth and fail to capture complex domain-specific relationships, particularly in morphologically rich languages such as Greek and in specialized domains such as legal text processing. In this paper, we introduce KG-Tune, a knowledge graph-driven fine-tuning methodology, which is specifically tailored for RAG applications. KG-Tune addresses the limitations of existing methodologies by leveraging structured knowledge extraction through a systematic two-phase training framework that incorporates (i) systematic extraction of entities, relationships and atomic facts from domain-specific knowledge graphs to generate semantically grounded training instances, and (ii) hard negative mining to enhance discriminative capabilities. Experimental evaluation on four heterogeneous Greek datasets demonstrates that the proposed methodology achieves considerable improvements, with top-3 accuracy gains ranging from 5.4% to 18.0% and an average performance improvement of 11.1% compared to traditional fine-tuning approaches. An ablation study further confirms the significance of these improvements.