.png){kind=icon}
.png){kind=icon}
Return
STaR: a soft-labeling and triplet-aware retriever for efficient retrieval-augmented QA
.png){kind=icon}
PRE
AI
J
C
Y
T
D
DOI：10.1007/s00530-026-02559-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Retrieval-augmented generation (RAG) has emerged as a powerful paradigm for improving the factuality and grounding of large language models (LLMs). However, existing RAG systems suffer from several persistent limitations, including retrieval inefficiency, high computational costs, and limited semantic understanding due to the reliance on hard labels and recursive retrieval methods. To address these issues, this study proposes STaR, a novel retriever fine-tuning framework that integrates BM25 similarity graph-based soft labeling with a triplet similarity learning strategy based on Sentence-BERT (SBERT). Our approach constructs graded semantic labels through BM25 and similarity graph traversal, overcoming the binary constraints of traditional supervision and enriching the learning signal for more precise vector alignment. Furthermore, this study introduces a triplet-aware SBERT training architecture that explicitly models relative semantic distances between queries and candidate passages, significantly enhancing retrieval ranking precision and semantic robustness. This study also adopts a dual-level evaluation strategy that combines lexical level and semantic level accuracy, leveraging their complementary strengths to better reflect retrieval quality and enhance downstream generation. Experimental results on a proprietary dataset of 98,037 QA pairs demonstrate that our method outperforms strong baselines—achieving HR@5 = 0.93 with consistent gains in MRR and favorable ARP. Compared with traditional and hybrid retrieval methods, STaR not only improves semantic precision but also reduces unnecessary computation and redundancy, making it well-suited for resource-constrained and real-world QA deployments.
Keywords:
Retrieval-augmented generation
Large language models
Question-answering systems
Soft labeling
Retriever fine-tuning
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
Cited Papers