A shape composition method for named entity recognition.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Large language models (LLMs) struggle with named entity recognition (NER), particularly for nested structures, due to dense sentence vectorization.
• The pre-training process in LLMs can overwhelm the decoding stage with sentence-specific information, degrading NER performance.

Purpose of the Study:

• To develop a novel approach for named entity recognition that overcomes the limitations of traditional LLMs.
• To improve the recognition of complex and nested named entities by decomposing them into semantic components.

Main Methods:

• A discriminative language model maps sentences into a high-order semantic space, decomposing named entities into entity bodies and edges.
• A shape composition method utilizing a multi-objective learning neural architecture is proposed.
• The neural architecture simultaneously detects entity bodies and classifies entity edges through dual learning objectives.

Main Results:

• The proposed method demonstrates competitive performance across eight public datasets for named entity recognition.
• Analytical experiments confirm that allowing semantic expressions to develop naturally aligns with the entity recognition task.
• The approach yields finer-grained semantic representations, enhancing both NER and other Natural Language Processing (NLP) tasks.

Conclusions:

• The shape composition method effectively addresses the performance degeneration of LLMs in named entity recognition, especially for nested entities.
• Decomposing named entities into body and edge components in a high-order semantic space improves the decoding of complex semantic structures.
• This novel approach offers a promising direction for advancing NER and broader NLP applications through enhanced semantic understanding.