Fusion of Channel Spatial Attention Enhances Span-Boundary Awareness for Nested Named Entity Recognition

The named entity recognition (NER) task is a fundamental task in the field of information extraction, aiming to locate the boundaries of entities within a sentence and classify them accordingly. To address the issue of blurred entity boundaries in span classification models, we propose a nested named entity recognition method that fuses channel and spatial attention to enhance span boundary perception. Firstly, we employ a double affine attention mechanism to capture the semantic correlations between word units, generating a 3D feature matrix. Secondly, the 3D feature matrix is enhanced by designing the Channel Local Enhancement Module (CLEM) and the Spatial Local Enhancement Module (SLEM), which respectively enhance local channel information interaction and local spatial information interaction to mitigate the problem of entity boundary blurring. Additionally, inspired by the Channel Attention and Spatial Attention mechanism (CBAM), we propose a fusion strategy where CLEM is applied in series with SLEM to improve the information enhancement effects of both modules, thereby increasing the accuracy of nested named entity recognition. To verify the effectiveness of the model, we tested it on three benchmark datasets: ACE04, ACE05, and Genia. Experimental results show that the proposed model outperforms existing methods in terms of entity recognition accuracy. Furthermore, ablation experiments were conducted to analyze and demonstrate the effectiveness of the tandem fusion strategy of CLEM and SLEM.