Unsupervised cross-domain named entity recognition using entity-aware adversarial training

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Neural network models for named entity recognition (NER) typically require extensive manually labeled data.
• These models are ineffective in new domains with entirely unlabeled data.
• A significant challenge is the distribution divergence between source and target domains during knowledge transfer.

Purpose of the Study:

• To develop an unsupervised cross-domain model for NER.
• To leverage labeled data from a source domain to predict entities in an unlabeled target domain.
• To mitigate distribution divergence and improve cross-domain knowledge transfer.

Main Methods:

• Proposed an unsupervised cross-domain model for NER.
• Employed adversarial training to reduce distribution divergence between source and target domains.
• Designed an entity-aware attention module to guide adversarial training and minimize entity feature discrepancies.

Main Results:

• The proposed model demonstrated superior performance compared to existing methods.
• Achieved state-of-the-art results in unsupervised cross-domain NER.
• The entity-aware attention module effectively reduced feature discrepancies between domains.

Conclusions:

• The unsupervised cross-domain model successfully addresses the challenge of NER in unlabeled domains.
• Adversarial training combined with an entity-aware attention module is effective for cross-domain knowledge transfer.
• The approach offers a viable solution for NER tasks with limited labeled data.