Long-Range Named Entity Recognition: A Comprehensive Survey

Abstract

The exponential growth of unstructured digital text has created a pressing need for sophisticated Natural Language Processing (NLP) methods to extract meaningful information. Named Entity Recognition (NER), the task of identifying and classifying named entities in text, is a cornerstone of this effort. While traditional NER has achieved remarkable success on short, self-contained texts, its application to long-form docu-ments—such as legal contracts, clinical records, and scientific literature—presents formidable challenges. This survey provides a comprehensive analysis of the state-of-the-art in Long-Range Named Entity Recognition. We trace the evolution from classical statistical models to the rise of Transformers, detailing the inher-ent quadratic complexity of models like BERT that limits their scalability. We conduct an in-depth exploration of the primary architectural paradigms designed to overcome this bottleneck: efficient Transformers that employ sparse attention mechanisms, and graph-based approaches that model explicit relational struc-tures within documents. Furthermore, we investigate critical challenges, including the data scarcity problem in specialized domains and unique linguistic complexities in multilingual con-texts. Drawing from recent analyses, we synthesize persistent open problems in document-level information extraction, focusing on long-distance coreference resolution and the need for robust, multi-step reasoning. Finally, we chart a course for future research, postulating that the next generation of solutions will be found in hybrid architectures that synergistically combine the strengths of deep sequential encoders with structured reasoning frameworks