---
id: 20260304-T0-07
title: "从全局到局部：文档分类与摘要新方法"
title_en: "Document Classification: New Method from Local Context"
url: https://ai.daily.yangsir.net/daily/20260304-T0-07
issue_date: 2026-03-04
publish_date: 2026-03-03T05:00:00.000Z
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2603.00021
---

# 从全局到局部：文档分类与摘要新方法

一篇arXiv论文提出了一种数据驱动的文档表示方法，通过动态滑动窗口注意力模块构建文档图表示。该方法能有效捕捉文档中的局部上下文信息，提升分类和摘要任务的性能。论文指出，传统方法难以处理长文档的局部依赖关系，而新方法通过动态窗口机制解决了这一问题。实验显示，该方法在多个基准测试中优于现有技术。

## English Version

**Document Classification: New Method from Local Context**

An arXiv paper presents a data-driven document representation method using dynamic sliding window attention to build document graphs. This approach effectively captures local context information in documents, improving classification and summarization performance. Traditional methods struggle with long document dependencies, which the new method solves through dynamic window mechanisms, outperforming existing techniques in multiple benchmarks.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2603.00021)

**详情页**：https://ai.daily.yangsir.net/daily/20260304-T0-07

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*