Global Span Semantic Dependency Awareness and Filtering Network for nested named entity recognition

Abstract

Span-based methods for nested named entity recognition (NER) are effective in handling the complexities of nested entities with hierarchical structures. However, these methods often overlook valid semantic dependencies among global spans, resulting in a partial loss of semantic information. To address this issue, we propose the Global Span Semantic Dependency Awareness and Filtering Network (GSSDAF). Our model begins with BERT for initial sentence encoding. Following this, a span semantic representation matrix is generated using a multi-head biaffine attention mechanism. We introduce the Global Span Dependency Awareness (GSDA) module to capture valid semantic dependencies among all spans, and the Local Span Dependency Enhancement (LSDE) module to selectively enhance key local dependencies. The enhanced span semantic representation matrix is then decoded to classify the spans. We evaluated our model on seven public datasets. Experimental results demonstrate that our model effectively handles nested NER, achieving higher F1 scores compared to baselines. Ablation experiments confirm the effectiveness of each module. Further analysis indicates that our model can learn valid semantic dependencies between global spans, significantly improving the accuracy of nested entity recognition. Our code is available at https://github.com/Shaun-Wong/GSSDAF.