Abstract
In document-level relation extraction, entities may appear multiple times in a document, and their relationships can shift from one context to another. Accurate prediction of the relationship between two entities across an entire document requires building a global context spanning all relevant sentences. Previous approaches have focused only on the sentences where entities are mentioned, which fails to capture the complete document context necessary for accurate relation extraction. Therefore, this paper introduces a novel input embedding approach to capture the positions of mentioned entities throughout the document rather than focusing solely on the span where they appear. The proposed input encoding approach leverages global relationships and multi-sentence reasoning by representing entities as standalone segments, independent of their positions within the document. The performance of the proposed method has been tested on three benchmark relation extraction datasets, namely DocRED, Re-DocRED, and REBEL. The experimental results demonstrated that the proposed method accurately predicts relationships between entities in a document-level setting. The proposed research also has theoretical and practical implications. Theoretically, it advances global context modeling and multi-sentence reasoning in document-level relation extraction. Practically, it enhances relationship detection, enabling improved performance in real-world NLP applications requiring comprehensive entity-level insights and interpretability.
Abstract (translated)
在文档级关系抽取中，实体可能在一个文档中多次出现，并且它们之间的关系会随着上下文的变化而变化。准确预测整个文档中两个实体之间的关系需要构建一个覆盖所有相关句子的全局背景。以往的方法仅关注提到这些实体的那些句子，未能捕捉到进行精确关系提取所需的完整文档背景信息。因此，本文介绍了一种新颖的输入嵌入方法，用于在整个文档范围内捕获提及实体的位置，而不仅仅是它们出现的具体范围。所提出的输入编码方法通过将实体表示为独立于其在文档中的位置存在的单独段落来利用全局关系和跨句子推理。 该方法已在三个基准关系抽取数据集（即DocRED、Re-DocRED和REBEL）上进行了测试。实验结果表明，提出的方法能够准确预测文档级设置中实体之间的关系。此外，这项研究还具有理论和实践上的意义。理论上，它推动了全局背景建模和跨句子推理在文档级关系抽取中的应用。实践中，它可以增强关系检测能力，在需要全面实体级别洞察力和可解释性的现实世界自然语言处理（NLP）应用中表现出更好的性能。
URL
https://arxiv.org/abs/2507.22926
