"Data for Differentiable Semantic Interaction: Decision-Focused Learning for Narrative Maps"
收藏DataCite Commons2026-04-26 更新2026-05-03 收录
下载链接：
https://ieee-dataport.org/documents/data-differentiable-semantic-interaction-decision-focused-learning-narrative-maps
下载链接
链接失效反馈官方服务：
资源简介：
"Narrative maps summarize news corpora as directed acyclic graphs of articles. Existing systems handle analyst feedback locally: each correction adds a constraint or retrains a projection, but the coherence function never changes. We reformulate extraction as a differentiable quadratic program and replace the fixed coherence function with a parametric model, enabling analyst interactions to propagate as gradient updates that globally reshape the coherence surface. We map all five interaction types to differentiable loss terms and evaluate across 22 scenarios on two news corpora, converging on at least 75% of seeds per scenario with clearest advantages on temporal and topic-dominance violations. This data shows the results of a comparison between the traditional 3MSI pipeline from Keith et al. (2023) and our proposal."
提供机构：
IEEE DataPort
创建时间：
2026-04-26
