Controlled Attention-Based Prototype Representation Learning for Named Entity Recognition

Prototype-based named entity recognition (NER) methods typically construct entity representations using characters or spans. However, due to overlapping contextual information in nested entities, such representations often become ambiguous, making it difficult to accurately measure the distances between entity instances and their corresponding prototypes. To address this issue, we propose a sentence-level prototype NER framework based on controlled attention. Our method introduces boundary-aware attention cues that enable the model to perceive entity boundaries and establish global semantic dependencies between entities and the entire sentence. The framework consists of three key components: (1) attention cue generation for boundary-aware sentence encoding, (2) abstraction of entity context semantics to obtain discriminative entity representations, and (3) prototype representation learning via contrastive learning. Specifically, we first generate attention cues to build sentence-level entity instances, then extract contextual features to form abstract representations, and finally learn class-specific prototypes through a contrastive objective. A series of experimental results demonstrate the effectiveness of our approach. In particular, on the CoNLL2003 dataset, our method achieves a 6.37% higher F1 score compared to previous prototype-based methods, with a smaller standard deviation, indicating both superior performance and improved stability.