Parallel sequence tagging for concept recognition

Area of Science:

• Biomedical Natural Language Processing
• Computational Biology
• Bioinformatics

Background:

• Named Entity Recognition (NER) and Normalisation (NEN) are crucial for biomedical text mining.
• Traditional serial pipelines for NER and NEN are susceptible to error propagation.
• A parallel architecture offers a potential solution to overcome limitations of serial processing.

Purpose of the Study:

• To propose and evaluate a parallel architecture for joint Named Entity Recognition and Normalisation (NER/NEN).
• To model NER and NEN as sequence-labeling tasks operating directly on source text.
• To investigate various harmonisation strategies for merging predictions from parallel classifiers.

Main Methods:

• Developed a parallel architecture for NER and NEN, treating both as sequence-labeling tasks.
• Examined different strategies for harmonising predictions from NER and NEN classifiers.
• Evaluated the approach on the CRAFT corpus (Version 4), including 20 annotation sets.

Main Results:

• The proposed parallel system significantly outperformed the baseline pipeline system on all 20 annotation sets of the CRAFT corpus.
• Optimising harmonisation strategies for each annotation set further refined system performance.
• Demonstrated the effectiveness of combining strengths from independent NER and NEN classifiers.

Conclusions:

• The parallel architecture successfully combines the strengths of NER and NEN classifiers.
• Effective prediction harmonisation requires individual calibration for each annotation set.
• This approach balances existing knowledge with the identification of novel concepts.