Scholarlink

<p>Relation prediction is a key task in knowledge graph completion that involves inferring the correct relation between a given head–tail entity pair. While transformer-based models such as KG-BERT leverage semantic representations effectively, they often overfit and fail to capitalize on misclassified or low-confidence instances. To address these limitations, we propose <b>wiqiRBERT</b>, a confidence-aware focused retraining framework that selectively fine-tunes the model using two types of challenging examples: misclassified triples and those with low prediction confidence. This post hoc optimization strategy, integrated with mixed-precision training and robust checkpoint control, enhances performance without architectural modifications. We evaluate wiqiRBERT on three benchmark datasets: <b>WN18RR</b>, <b>WN18</b>, and <b>FB15k-237</b>. Experimental results demonstrate consistent improvements over strong baselines. Notably, wiqiRBERT achieves state-of-the-art results across all metrics on WN18RR, superior MRR and Hits@3 on WN18, and outperforms prior methods on Hits@3 in the more challenging FB15k-237 dataset. These findings confirm the efficacy of confidence-driven retraining as a lightweight yet impactful enhancement for BERT-based relation prediction models.</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

wiqiRBERT: lightweight relation prediction via confidence-aware focused retraining

  • 期刊论文

  • 发表日期：

    2025年12月18日

查看全文
  • Waqas Ali,
  • Bin Yao

摘要

Relation prediction is a key task in knowledge graph completion that involves inferring the correct relation between a given head–tail entity pair. While transformer-based models such as KG-BERT leverage semantic representations effectively, they often overfit and fail to capitalize on misclassified or low-confidence instances. To address these limitations, we propose wiqiRBERT, a confidence-aware focused retraining framework that selectively fine-tunes the model using two types of challenging examples: misclassified triples and those with low prediction confidence. This post hoc optimization strategy, integrated with mixed-precision training and robust checkpoint control, enhances performance without architectural modifications. We evaluate wiqiRBERT on three benchmark datasets: WN18RR, WN18, and FB15k-237. Experimental results demonstrate consistent improvements over strong baselines. Notably, wiqiRBERT achieves state-of-the-art results across all metrics on WN18RR, superior MRR and Hits@3 on WN18, and outperforms prior methods on Hits@3 in the more challenging FB15k-237 dataset. These findings confirm the efficacy of confidence-driven retraining as a lightweight yet impactful enhancement for BERT-based relation prediction models.