Attention-based interactive multi-level feature fusion for named entity recognition.

Area of Science:

• Natural Language Processing (NLP)
• Deep Learning
• Machine Learning

Background:

• Named Entity Recognition (NER) is crucial for NLP tasks, identifying entities like persons, locations, and organizations.
• Deep Neural Networks (DNNs) are widely used for NER, but often overlook multi-level entity features and their dependencies.
• Existing DNN models struggle to fully leverage diverse features such as lexical phrases, capitalization, and suffixes.

Purpose of the Study:

• To propose a novel attention-based interactive multi-level feature fusion (AIMFF) framework for enhanced NER.
• To address the limitations of current models in utilizing multi-level entity features and inter-feature dependencies.
• To improve the accuracy and robustness of Named Entity Recognition systems.

Main Methods:

• The AIMFF framework integrates input, feature extraction, feature fusion, and sequence labeling layers.
• It generates word- and character-level embeddings and captures global/local word and character features.
• Cross-attention mechanisms are employed to interactively fuse word- and character-level features for enriched representations.

Main Results:

• Comparative experiments on three datasets demonstrated superior performance of the AIMFF model.
• The proposed framework achieved better results than several state-of-the-art NER models.
• The interactive fusion of multi-level features significantly boosted NER performance.

Conclusions:

• The AIMFF framework effectively captures and fuses multi-level entity features for improved NER.
• Attention-based interactive fusion is a promising direction for advancing NLP tasks.
• The model offers a significant improvement over existing methods for Named Entity Recognition.