Concept recognition as a machine translation problem.

Area of Science:

• Biomedical Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Automated ontology concept assignment is crucial for biomedical NLP.
• Current methods struggle with large ontologies and limited data.
• Machine translation offers a novel approach to concept recognition.

Purpose of the Study:

• To systematically evaluate sequence-to-sequence machine learning for biomedical concept recognition.
• To identify optimal methods and hyperparameters for accuracy and efficiency.
• To provide insights into system strengths, weaknesses, and future improvements.

Main Methods:

• Utilized sequence-to-sequence machine learning models, including Bidirectional Encoder Representations from Transformers for biomedical text mining (BioBERT) for span detection.
• Employed the Open-Source Toolkit for Neural Machine Translation (OpenNMT) for concept normalization.
• Conducted extensive studies on alternative methods and hyperparameter selections.

Main Results:

• Achieved state-of-the-art performance for most ontologies in the CRAFT Corpus using BioBERT and OpenNMT.
• Demonstrated significantly reduced computational resource requirements (hardware, memory, time) compared to alternative approaches.
• Identified factors contributing to accuracy and efficiency in sequence-to-sequence models.

Conclusions:

• Machine translation is a highly effective strategy for fully machine-learning-based concept recognition.
• The proposed approach yields state-of-the-art results on the CRAFT Corpus.
• Future work can explore alternative target concept representations for further advancements.