JCBIE: a joint continual learning neural network for biomedical information extraction.

Area of Science:

• Biomedical Informatics
• Knowledge Representation
• Machine Learning

Background:

• Extracting knowledge from diverse biomedical data is crucial for building structured biomedical knowledge graphs (BKGs).
• Existing methods are limited by task-specific training sets, hindering large-scale BKG development and diverse applications.
• There is a need for methods that can learn and expand entity and relation types across different datasets.

Purpose of the Study:

• To propose a joint continual learning biomedical information extraction (JCBIE) network.
• To address the limitations of existing methods in handling heterogeneous biomedical data.
• To enable the extraction and expansion of diverse entities and relations from multiple datasets for improved BKG construction.

Main Methods:

• Developed a joint continual learning network (JCBIE) for biomedical information extraction.
• Employed distinct encoders for joint-feature extraction to prevent feature confusion.
• Integrated entity augmented inputs to link named entity recognition and relation extraction.
• Proposed a novel evaluation mechanism for cross-corpus generalization errors.

Main Results:

• The JCBIE network demonstrated the ability to learn and expand entity and relation types across different datasets.
• Separated encoders effectively mitigated feature confusion issues.
• The proposed evaluation mechanism accurately measured cross-corpus generalization errors.
• Empirical studies confirmed promising performance with the continual learning strategy on multiple corpora.

Conclusions:

• The JCBIE network offers a robust solution for extracting knowledge from heterogeneous biomedical data.
• This approach enhances the scalability and applicability of biomedical knowledge graphs.
• The proposed continual learning strategy and evaluation mechanism advance the field of biomedical information extraction.