Boosting drug named entity recognition using an aggregate classifier.

Area of Science:

• Biomedical Natural Language Processing (NLP)
• Computational Linguistics
• Pharmacogenomics

Background:

• Drug named entity recognition (NER) is crucial for extracting key pharmacological information.
• Supervised machine learning for drug NER typically requires large, manually annotated datasets, which are labor-intensive to create and maintain.
• This limitation hinders the development of advanced biomedical NLP applications.

Purpose of the Study:

• To enhance drug NER performance without exclusive reliance on manual annotations.
• To investigate methods for improving drug NER using limited or no gold-standard data.
• To develop a robust drug NER system by combining heterogeneous data sources and models.

Main Methods:

• A voting system was developed to integrate diverse models, including dictionary knowledge (DrugBank), a small gold-standard corpus (pharmacokinetic corpus), and silver annotations.
• Genetic programming was employed to evolve 11 regular-expression patterns for recognizing common drug suffixes, enhancing recall.
• Maximum entropy and multinomial logistic regression classifiers were trained using gold-standard and silver annotated data.

Main Results:

• Aggregating models based on dictionary knowledge, patterns, and gold-standard annotations improved performance, achieving a maximum F-score of 95%.
• Combining models trained on silver annotations, dictionary knowledge, and patterns yielded performance comparable to models trained solely on gold-standard data.
• Morphological similarities among drug names are suggested as a key factor for this comparable performance.

Conclusions:

• Gold-standard data are not an absolute requirement for achieving high-performance drug NER.
• Combining heterogeneous models, particularly those leveraging dictionary knowledge, can achieve classification performance similar or comparable to models trained exclusively on extensive gold-standard annotations.
• This approach offers a more efficient pathway for developing effective drug NER systems.