Related Experiment Video
Updated: Aug 29, 2025
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
4.0K
Research on Named Entity Recognition Based on Multi-Task Learning and Biaffine Mechanism
Wenchao Gao1, Yu Li1, Xiaole Guan1
1School of Mechanical Electronic & Information Engineering, China University of Mining & Technology, Beijing 100083, China.
Computational Intelligence and Neuroscience
|September 5, 2022
Summary
This study introduces a novel named entity recognition (NER) method using multi-task learning and a biaffine mechanism to improve nested entity recognition. The approach enhances recall by adding boundary detection, significantly boosting performance on benchmark datasets.
Area of Science:
- Natural Language Processing
- Machine Learning
- Information Extraction
Background:
- Traditional span-based nested entity recognition (NER) methods struggle with boundary supervision, leading to high precision but low recall.
- Existing approaches often fail to accurately predict entities with overlapping or nested structures.
Purpose of the Study:
- To propose a novel NER method that addresses the limitations of span-based methods in nested entity recognition.
- To enhance the recall of nested entity recognition by incorporating explicit boundary supervision.
Main Methods:
- A multi-task learning framework is employed, dividing the NER task into two subtasks: entity span classification and boundary detection.
- A biaffine mechanism is utilized for entity span classification to score potential spans and assign entity types.
- Boundary detection captures relationships between adjacent words to identify entity boundaries, providing additional supervision.
Main Results:
- The proposed method significantly improves F1 scores on nested NER tasks, achieving gains of up to 7.05% on GENIA, 12.63% on ACE2004, and 14.68% on ACE2005.
- Experimental results demonstrate superior performance compared to existing nested entity recognition methods.
Conclusions:
- The integration of multi-task learning and a biaffine mechanism effectively enhances nested entity recognition.
- The boundary detection subtask is crucial for overcoming the low recall issues inherent in traditional span-based NER approaches.
