SSAM: a span spatial attention model for recognizing named entities.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Named entity recognition (NER) traditionally classifies spans independently, overlooking spatial relationships.
• Existing methods fail to capture multi-granular span dependencies effectively.

Purpose of the Study:

• To propose a novel Span Spatial Attention Model (SSAM) for enhanced named entity recognition.
• To address limitations in existing NER approaches by incorporating spatial structures.

Main Methods:

• Developed a Span Spatial Attention Model (SSAM) with a token encoder, span generation module, and 2D spatial attention network.
• Employed a two-channel span generation strategy for multi-granular feature capture.
• Applied spatial attention to a 2D sentence representation, unlike sequential attention.

Main Results:

• Achieved state-of-the-art performance on GENIA, ACE2005, and ACE2004 datasets.
• Reported F1-scores of 81.82% (GENIA), 89.04% (ACE2005), and 89.24% (ACE2004).
• Demonstrated the model's ability to adaptively encode important features and suppress irrelevant information.

Conclusions:

• The proposed SSAM effectively captures spatial structures and multi-granular span dependencies in NER.
• The 2D spatial attention mechanism offers a significant advancement over traditional sequential attention for NER tasks.
• The model's superior performance highlights the importance of spatial information in named entity recognition.