From local to global: Leveraging document graph for named entity recognition

Named Entity Recognition (NER) is a fundamental task in Natural Language Processing (NLP) that aims to identify the span and category of entities within text. Recent advancements have demonstrated significant improvements in NER performance by incorporating document-level context. However, due to input length limitations, these models only consider the context of nearby sentences, failing to capture global long-range dependencies within the entire document. To address this issue, we propose a novel span-based two-stage method that formulates the document as a span graph, enabling the capture of global long-range dependencies at both token and span levels. Specifically, (1) we first train a binary classifier without considering entity types to extract candidate spans from each sentence. (2) Then, we leverage the robust contextual understanding and structural reasoning capabilities of Large Language Models (LLMs) like GPT to incrementally integrate these spans into the document-level span graph. By utilizing this span graph as a guide, we retrieve relevant contextual sentences for each target sentence and jointly encode them using BERT to capture token-level dependencies. Furthermore, by employing a Graph Transformer with well-designed position encoding to incorporate graph structure, our model effectively exploits span-level dependencies throughout the document. Extensive experiments on resource-rich nested and flat NER datasets, as well as low-resource distantly supervised NER datasets, demonstrate that our proposed model outperforms previous state-of-the-art models, showcasing its effectiveness in capturing long-range dependencies and enhancing NER accuracy.