Scholarlink

<p>Accurate classification of local citation relationship is critical for enhancing recommendation quality in scholarly information systems. To advance local citation relationship classification, this paper proposed a novel SciBERT + DualFocusNet architecture comprising three synergistic components: SciBERT for domain-specific semantic representation, DualFocusNet for hierarchical extraction and fusion of multi-scale global semantic and local contextual features, and large language models for generating semantically compressed reference summaries. Three sets of experiments were conducted on ACL ARC/AAN and ACT2 datasets to systematically validate the proposed approach. In the first set, experimental results investigated that the SciBERT + DualFocusNet model outperformed all competing baselines across accuracy, precision, recall, and F1-score metrics, achieving 84.98% on ACL ARC/AAN and 93.31% on ACT2. The second set identified optimal text combinations, revealing that pairing citation context with extended reference text (title + abstract + GPT-5-processed full text) yielded marginally better performance than alternative text combinations. The third set of ablation experiments confirmed that the local–global fusion strategy achieved optimal performance and further decomposed the contribution of each architectural component to overall system efficacy.</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

Local citation relationship classification for recommendation: a SciBERT-based DualFocusNet approach

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Yonghe Lu,
  • Meilu Yuan,
  • Minghong Chen

摘要

Accurate classification of local citation relationship is critical for enhancing recommendation quality in scholarly information systems. To advance local citation relationship classification, this paper proposed a novel SciBERT + DualFocusNet architecture comprising three synergistic components: SciBERT for domain-specific semantic representation, DualFocusNet for hierarchical extraction and fusion of multi-scale global semantic and local contextual features, and large language models for generating semantically compressed reference summaries. Three sets of experiments were conducted on ACL ARC/AAN and ACT2 datasets to systematically validate the proposed approach. In the first set, experimental results investigated that the SciBERT + DualFocusNet model outperformed all competing baselines across accuracy, precision, recall, and F1-score metrics, achieving 84.98% on ACL ARC/AAN and 93.31% on ACT2. The second set identified optimal text combinations, revealing that pairing citation context with extended reference text (title + abstract + GPT-5-processed full text) yielded marginally better performance than alternative text combinations. The third set of ablation experiments confirmed that the local–global fusion strategy achieved optimal performance and further decomposed the contribution of each architectural component to overall system efficacy.