Noise Reduction Learning Based on XLNet-CRF for Biomedical Named Entity Recognition.

Area of Science:

• Computational biology
• Bioinformatics
• Natural Language Processing

Background:

• Biomedical Named Entity Recognition (BioNER) systems are crucial for extracting information from vast biomedical literature.
• Deep neural networks, particularly transformer-based models, have shown success in BioNER.
• Existing models face challenges with noise interference during training and lack effective label dependency decoding.

Purpose of the Study:

• To propose and evaluate novel noise reduction models for improving BioNER performance.
• To address limitations in current BioNER systems, including noise interference and label dependency issues.
• To enhance the accuracy and efficiency of information extraction from biomedical texts.

Main Methods:

• Developed two noise reduction models: Shared Labels and Dynamic Splicing.
• Utilized XLNet encoding, a permutation language pre-training model.
• Employed Conditional Random Field (CRF) for decoding to handle label dependencies.
• Tested models on 15 diverse biomedical named entity recognition datasets.

Main Results:

• Achieved significant improvements in average F1-score: 1.504 for Shared Labels and 1.48 for Dynamic Splicing.
• Attained state-of-the-art performance on 7 out of 15 tested BioNER datasets.
• Demonstrated the effectiveness of the proposed models and the enhanced recognition capabilities of CRF.

Conclusions:

• The proposed Shared Labels and Dynamic Splicing models effectively reduce noise and improve BioNER performance.
• XLNet encoding combined with CRF decoding offers a robust approach for biomedical text analysis.
• The study provides insights into the applicable scope of these models based on data characteristics, advancing BioNER research.