Efficient Named Entity Recognition with Overlapping and Nested Mentions Using Hypergraphs and Neural Network

Named Entity Recognition (NER) remains a pivotal task in Natural Language Processing (NLP), crucial for information extraction and document classification applications. Traditional models like linear-chain Conditional Random Fields (CRFs) struggle to handle overlapping and nested mentions, which frequently occur in specialized domains such as biomedical and legal texts. This study addresses these limitations by proposing an advanced model that integrates mention hypergraphs, mention separators, and neural networks to efficiently recognize complex entity structures while maintaining linear time complexity. The model’s performance is tested across multiple datasets, including ACE2004, ACE2005, and GENIA, demonstrating significant improvements in F1 scores and scalability compared to baseline models. The proposed approach also introduces a dynamic feature composition mechanism that enhances the recognition of rare and unseen words, making it adaptable to diverse text genres, including noisy and informal data. The results suggest that the model excels in traditional NER tasks and extends its applicability to low-resource settings and real-world datasets. Future work could further optimize computational efficiency and explore domain-specific adaptations to refine performance in specialized fields.