Consistency enhancement of model prediction on document-level named entity recognition

Area of Science:

• Biomedical Natural Language Processing
• Bioinformatics
• Computational Biology

Background:

• Biomedical named entity recognition (NER) is vital for information extraction in biomedical applications.
• Document-level NER presents challenges beyond sentence-level analysis, often resulting in inconsistent predictions.
• Existing document NER models struggle to meet researcher expectations for label consistency.

Purpose of the Study:

• To investigate the causes of inconsistent predictions in document-level biomedical NER.
• To develop a method, ConNER, that enhances label consistency, particularly for modifiers like adjectives and prepositions.
• To improve the representation of biomedical entities by refining modifier labels.

Main Methods:

• Investigated underlying causes of inconsistent predictions in biomedical NER.
• Developed ConNER, a method to refine labels of modifiers (adjectives, prepositions) within entities.
• Evaluated ConNER on four popular biomedical NER datasets.

Main Results:

• ConNER achieved higher F1 scores than the state-of-the-art on three out of four datasets.
• Demonstrated absolute F1 score improvements of 7.5%-8.6% on two datasets.
• Qualitative analysis confirmed ConNER's ability to generate more consistent NER predictions.

Conclusions:

• ConNER effectively enhances label consistency in biomedical NER, especially for datasets with inherent low consistency.
• The method improves the representation of biomedical entities by refining modifier labels.
• ConNER offers a promising approach for more reliable information extraction in biomedical text.