A boundary enhanced multi-task neural attention approach for Chinese named entity recognition

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Named Entity Recognition (NER) is crucial for Chinese information processing.
• The absence of explicit word boundaries in Chinese presents unique challenges for NER tasks.

Purpose of the Study:

• To propose a novel approach for Chinese NER by framing it as a joint task.
• To enhance NER performance by integrating boundary detection with entity identification.

Main Methods:

• An encoder-decoder architecture is employed, combining sequence labeling for NER and binary classification for boundary prediction.
• Hybrid embeddings and a convolutional attention network are used for enhanced word representations and contextual information extraction.
• Two Bi-GRU networks predict entity start/end, and a feature fusion layer integrates main and auxiliary tasks.

Main Results:

• The proposed multi-task learning framework significantly improves Chinese NER performance.
• Experimental results on Weibo and Ontonotes5.0 datasets demonstrate superior performance compared to existing models.

Conclusions:

• Jointly learning boundary detection and entity identification effectively addresses challenges in Chinese NER.
• The integrated approach enhances the accuracy and robustness of Chinese NER systems.