S-NER: A Concise and Efficient Span-Based Model for Named Entity Recognition.

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning

Background:

• Named Entity Recognition (NER) is crucial for downstream NLP tasks.
• Traditional NER models using sequence labeling and Conditional Random Fields (CRFs) suffer from cascading misclassifications due to label dependencies.

Purpose of the Study:

• To propose S-NER, a novel span-based NER model.
• To overcome the limitations of traditional sequence labeling approaches in NER.
• To enhance the accuracy and robustness of entity recognition in raw texts.

Main Methods:

• S-NER segments raw texts into candidate entity spans.
• Entity types are determined by direct classification of span semantic representations, eliminating label dependency.
• A concise neural architecture utilizing BERT as an encoder and a feed-forward network as a decoder is employed.

Main Results:

• S-NER consistently outperforms state-of-the-art baselines across multiple benchmark datasets and domains.
• Experimental results show significant improvements in F1-score compared to traditional methods.
• Extensive analyses validate the effectiveness and efficiency of the S-NER model.

Conclusions:

• The proposed S-NER model offers a more effective approach to Named Entity Recognition.
• Span-based classification eliminates error propagation inherent in sequence labeling models.
• S-NER demonstrates superior performance and robustness for diverse NLP applications.