面向科研信息需求的语义要素候选发现与互补融合论文推荐方法
Semantic-Element-Aware Candidate Discovery and Complementary Fusion for Scientific Paper Recommendation
摘要
面向科研任务的论文推荐不仅要找到主题相近的文献，还应发现能够为背景梳理、方法迁移、机制解释和实验复用提供支持的工作。现有方法多依赖整体文本相似性或既有学术关系生成候选，难以捕捉论文之间在研究问题、方法、理论、数据和实验条件等方面的细粒度联系，且候选阶段遗漏的相关论文通常难以由后续排序弥补。为此，本文提出一种语义要素驱动的候选发现与互补融合方法，并构建从候选发现到价值优化的多阶段推荐框架。首先，以研究问题为锚点，沿方法、理论、数据、指标、软件和仪器等语义要素扩展候选，并与文本检索形成互补召回；随后，对合并候选进行初步排序与筛选，并在统一候选池上进行精细排序；最后，依据论文侧学术价值对列表前部进行有限调整，并通过独立的相关性风险控制机制约束潜在的相关性损失。实验结果表明，语义要素候选能够稳定补充强文本检索遗漏的相关论文，新增候选可被学习排序、专用重排器和大语言模型等不同精排机制有效利用；价值重排在不引入新候选的条件下进一步提升了列表前部的学术价值，并满足预设的相关性安全约束。总体而言，本文方法通过将细粒度候选发现、相关性排序与学术价值优化分阶段建模，为面向具体科研任务的论文推荐提供了新的方法框架。
Abstract
Paper recommendation for scientific research tasks should go beyond identifying topically similar papers to uncovering studies that can support background investigation, method transfer, mechanism interpretation, and experimental reuse. Existing approaches typically generate candidates based on holistic text similarity or established scholarly relations, making it difficult to capture fine-grained connections between papers in terms of research questions, methods, theories, data, and experimental conditions. Moreover, relevant papers missed during candidate generation are often difficult to recover in subsequent ranking stages. To address these limitations, we propose a semantic-element-driven approach to candidate discovery and complementary fusion, together with a multi-stage recommendation framework spanning candidate discovery, relevance ranking, and academic-value optimization. Specifically, we first anchor candidate discovery on research questions and expand the candidate space through semantic elements including methods, theories, data, metrics, software, and instruments, which are further combined with text-based retrieval to achieve complementary recall. The merged candidates are then preliminarily ranked and filtered, followed by fine-grained ranking over a unified candidate pool. Finally, the top portion of the ranked list is selectively adjusted according to paper-level academic value, while an independent relevance-risk control mechanism constrains potential losses in relevance. Experimental results show that semantic-element-based candidate discovery consistently complements strong text retrieval by recovering relevant papers that would otherwise be missed, and that these additional candidates can be effectively exploited by different fine-ranking mechanisms, including learning-to-rank models, specialized rerankers, and large language models. Academic-value reranking further improves the scholarly value of highly ranked results without introducing new candidates, while satisfying the predefined relevance-safety constraints. Overall, by explicitly separating fine-grained candidate discovery, relevance ranking, and academic-value optimization, the proposed approach provides a new framework for paper recommendation tailored to specific scientific research tasks.关键词
学术论文推荐/科研信息需求/科研语义要素/候选发现/多源候选融合/学术价值Key words
academic paper recommendation/scientific information needs/scientific semantic elements/candidate discovery/multi-source candidate fusion/academic value引用本文复制引用
张彧.面向科研信息需求的语义要素候选发现与互补融合论文推荐方法[EB/OL].(2026-09-28)[2026-10-01].https://chinaxiv.org/abs/202609.00507.学科分类
科学、科学研究
