BGNER: A boundary guidance framework for enhanced named entity recognition

Named Entity Recognition (NER) is a fundamental task in natural language processing that involves identifying and classifying text segments corresponding to real-world entities. Span-based NER methods, which treat each potential text segment as a candidate entity, have produced strong results. However, these approaches often fail to fully leverage boundary information, as they classify spans independently without explicitly incorporating the context immediately outside the span boundaries. This limitation can result in suboptimal recognition, particularly for nested or ambiguous entities. To overcome this drawback, we introduce BGNER, a method that employs a joint representation of entities and their boundaries. In this approach, each candidate span is enriched not only with an entity type label but also with a boundary label that captures its relationship with surrounding spans. We design a novel neural architecture to exploit this representation: a boundary-aware global pointer module efficiently scores spans by incorporating boundary context, and an entity-guided 3D convolutional neural network captures local patterns between spans and their boundaries. Importantly, predicted boundary spans serve as contextual cues during the decoding stage to validate entity predictions. We evaluate BGNER on six diverse NER datasets spanning both English and Chinese, as well as flat and nested entity scenarios. The experimental results show that our model achieves state-of-the-art performance across all datasets. These findings demonstrate that enhancing span representations with boundary information leads to more accurate and robust entity recognition in various NER settings.