anon-muses-neurips/citeroots

CiteRoots是一个双层测量框架，通过将“任何未来引用”收紧为在局部文本中发挥生成作用的引用以及引用论文作者自己认定为智力根源的引用，来补充MUSES前瞻检索基准。CiteRoots可独立使用：任何基于S2ORC论文对边操作的基准或分析都可以加入这些标签，以添加修辞或作者认可轴。它包括两个层次：CiteRoots-Rhetoric（修辞层），基于引用上下文的局部修辞角色分类为六个类别，聚合为论文级二元标签；CiteRoots-Endorsement（认可层），通过作者响应裁决工作台收集的作者认可生成灵感对。数据集包含多个文件，如修辞标签、人类黄金审计、认可对等，用于支持基准评估和分析。

CiteRoots is a two-layer measurement framework that complements the MUSES prospective retrieval benchmark by tightening any future citation toward citations that play generative roles in local text and citations the citing-paper authors themselves identify as intellectual roots. CiteRoots is independently usable: any benchmark or analysis that operates over S2ORC paper-pair edges can join with these labels to add a rhetorical or author-endorsement axis. It consists of two layers: CiteRoots-Rhetoric (passage-level → paper-level aggregated), which classifies the local rhetorical role of citation contexts into six categories aggregated to paper-level binary labels; and CiteRoots-Endorsement (paper-level), which collects author-attested generative-inspiration pairs through an author-response adjudication workbench. The dataset includes files such as rhetoric labels, human gold audit, endorsement pairs, and others to support benchmark evaluation and analysis.