GRAM-CNN: a deep learning approach with local context for named entity recognition in biomedical text

Area of Science:

• Computational biology
• Bioinformatics
• Natural Language Processing

Background:

• Traditional biomedical named entity recognition (NER) methods rely on costly hand-crafted features or task-specific rules, limiting their generalizability.
• While end-to-end neural networks excel in non-biomedical NER, they have not achieved competitive performance in the biomedical domain without domain-specific adaptations.

Purpose of the Study:

• To develop a novel end-to-end deep learning approach for biomedical NER that overcomes the limitations of existing methods.
• To create a model that leverages local contexts using n-gram character and word embeddings via Convolutional Neural Networks (CNNs).

Main Methods:

• Propose GRAM-CNN, a Convolutional Neural Network (CNN)-based deep learning model for biomedical NER.
• Utilize n-gram character and word embeddings to capture local contextual information around words for automatic labeling.
• The model requires no specific domain knowledge or manual feature engineering, enabling broad applicability.

Main Results:

• GRAM-CNN achieved state-of-the-art performance on three benchmark biomedical datasets.
• Achieved an F1-score of 87.26% on Biocreative II, 87.26% on NCBI, and 72.57% on JNLPBA.
• Demonstrated the effectiveness of CNN-based architectures for BioNER tasks.

Conclusions:

• GRAM-CNN represents a significant advancement in automated biomedical named entity recognition.
• The model's ability to generalize across different BioNER datasets highlights the power of deep learning without feature engineering.
• The source code, datasets, and pre-trained model are publicly available, facilitating further research and application.