Bi-directional context-aware network for the nested named entity recognition.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Span-based models are standard for Nested Named Entity Recognition (NER).
• Traditional models independently decode entity spans, neglecting inter-span semantic connections and positional data.
• This limitation hinders performance in complex Nested NER tasks.

Purpose of the Study:

• To propose a novel Bi-Directional Context-Aware Network (Bi-DCAN) for Nested NER.
• To enhance the understanding of semantic relationships between entity spans.
• To improve the accurate identification of entity positions within text.

Main Methods:

• Developed a new span-level semantic relation model.
• Implemented the Bi-DCAN to capture these semantic relationships.
• Integrated Rotary Position Embedding with a bi-affine mechanism for relative positional information capture.

Main Results:

• The Bi-DCAN model demonstrated superior performance on Nested NER tasks.
• Achieved improved F1 scores of 0.24 on ACE2005 and 0.09 on GENIA datasets.
• Reduced parameter count by 20 million compared to the Diffusion-NER model.

Conclusions:

• The Bi-DCAN effectively addresses limitations of traditional span-based models for Nested NER.
• The model accurately captures semantic and positional information for improved entity recognition.
• Bi-DCAN offers a more efficient and effective solution for Nested NER.