Nested Named Entity Recognition using Multilayer BERT-based Model: Notebook for the BioASQ Lab at CLEF 2024

Area of Science:

• Biomedical Natural Language Processing
• Computational Linguistics
• Bioinformatics

Background:

• Named Entity Recognition (NER) is vital in NLP for identifying and categorizing entities in text.
• Biomedical text presents unique challenges for NER due to complex language structures and nested entities.
• Existing NER models often struggle with the intricacies of nested and overlapping entities in specialized domains.

Purpose of the Study:

• To introduce an innovative multilayer BERT-based model for Nested NER in the biomedical domain.
• To effectively address the complexities of nested entities and overlapping mentions in biomedical literature.
• To enhance the accuracy and thoroughness of entity detection in specialized scientific texts.

Main Methods:

• Utilized a multilayer bidirectional encoder representation transformer (BERT) model, specifically pretrained PubMedBERT.
• Implemented a multilayer tagging process combined with robust contextual embeddings from PubMedBERT.
• Incorporated UMLS dictionaries to further refine biomedical entity recognition.

Main Results:

• The proposed Multilayer NER Model (MultilayerNERModel) demonstrated superior performance in identifying nested entities.
• Achieved the highest overall performance in the English-oriented track of the BioASQ competition's BioNNE dataset.
• Obtained an F1 score of 67.30% and a macro F1 score of 56.36% on the BioNNE English Dataset.

Conclusions:

• A multilayer approach significantly enhances the capability of NER models for nested entity recognition in biomedical texts.
• The MultilayerNERModel effectively manages overlapping entities, improving overall detection accuracy.
• Combining PubMedBERT embeddings with a multilayer strategy and UMLS dictionaries offers a powerful solution for biomedical NER.