Biomedical Entity Relation Mining via Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The rapid growth of biomedical literature makes manual collection and deep learning-based entity relation mining inefficient, relying heavily on costly large training datasets and complex machine learning features.
Innovation Solution
An entity relation mining method that preprocesses biomedical text data, uses regex matching and deep learning models for named entity recognition, and applies transfer and reinforcement learning to mine relations between entities, including constructing a bidirectional LSTM-CRF neural network model and fine-tuning BioBERT for relation extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are used for entity relation mining in biomedical literature, then classification accuracy is improved, but training time and computational resources increase significantly
Solution Approach 1:
The patent applies transfer learning by pre-training deep learning models on large-scale general biomedical text data before fine-tuning on specific entity relation tasks. This preliminary training on abundant data provides the model with foundational biomedical knowledge, reducing the need for extensive task-specific training and thereby decreasing training time while maintaining high accuracy
Solution Approach 2:
The patent extracts and utilizes only the essential features and representations learned from large-scale pre-training data that are most relevant to entity relation mining. By selecting and transferring only the critical learned patterns rather than retraining from scratch, the method achieves high accuracy with reduced training time and computational resources
2Measurement precision
If deep learning models are used for entity relation mining, then classification accuracy is improved, but server resources and computational power increase
Solution Approach 1:
The patent performs pre-training of deep learning models on large-scale biomedical data in advance, allowing the computationally intensive feature learning to be done once. The pre-trained models can then be deployed for entity relation mining with significantly reduced computational requirements during actual use, thereby lowering server resource consumption while maintaining high accuracy
Solution Approach 2:
The patent optimizes model parameters and architecture based on the specific requirements of entity relation mining tasks. By adjusting parameters such as model size, learning rate, and training epochs according to the task characteristics, the method achieves high classification accuracy with minimized computational resource usage
3Reliability
If manual collection methods are used for constructing biomedical databases, then data quality is maintained, but productivity decreases
Solution Approach 1:
The patent employs deep learning models that automatically perform entity recognition and relation extraction from biomedical literature. The system self-services by autonomously identifying entities, extracting relations, and constructing database entries without requiring manual curation for each record, thereby dramatically improving productivity while maintaining data quality through the models' learned understanding of biomedical concepts
4Measurement precision
If large training datasets are used for deep learning, then model performance is improved, but data construction cost increases
Solution Approach 1:
The patent leverages pre-trained models that have already been trained on large-scale general biomedical data. This preliminary training on abundant publicly available data provides a strong foundation that can be transferred to specific entity relation tasks with much smaller task-specific datasets, thereby achieving high model performance without incurring the high costs of constructing large annotated datasets from scratch
Solution Approach 2:
The patent uses pre-trained biomedical language models as an intermediary that bridges the gap between general biomedical text and specific entity relation tasks. These intermediary models have learned general biomedical knowledge from large corpora and can transfer this knowledge to downstream tasks, reducing the need for extensive task-specific training data while maintaining high performance
Data Source
AI summary
The present disclosure provides an entity relation mining method based on a biomedical literature, including the following steps: querying a disease-associated biomedical literature in a public database, and performing data preprocessing to obtain biomedical text data; performing biomedical named entity recognition on obtained biomedical text data in combination with a regex matching pattern and a deep learning model; and mining an entity relation with transfer learning and reinforcement learning based on an entity recognition result. By acquiring the disease-associated biomedical literature from a network, extracting an abstract and a title and performing entity recognition and relation mining, the present disclosure can effectively recognize biomedical noun entities in the literature and mine potential relations between various entities.