Distantly supervised biomedical relation extraction using piecewise attentive convolutional neural network and...

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Distantly supervised relation extraction (RE) faces challenges with erroneous training data.
• Existing methods for handling noisy data in biomedical RE are insufficient.
• Accurate biomedical relation prediction is crucial for knowledge acquisition.

Purpose of the Study:

• To address the insufficient modeling of instance-label correlations in distantly supervised biomedical RE.
• To develop a novel deep learning and reinforcement learning model for improved biomedical relation prediction.
• To enhance the accuracy of identifying relationships in biomedical texts.

Main Methods:

• Proposed a piecewise attentive convolutional neural network and reinforcement learning (PACNN+RL) model.
• Utilized PACNN to encode semantic information from biomedical text.
• Employed a reinforcement learning method with memory backtracking to mitigate erroneous data issues.
• Evaluated the model on four biomedical RE tasks using distantly supervised data.

Main Results:

• The PACNN+RL model achieved competitive performance across eight biomedical corpora.
• Outperformed baseline systems, achieving an F1-score of 0.5592 on the may-prevent dataset and 0.6666 on the may-treat dataset.
• Obtained new state-of-the-art results on four out of five benchmark datasets for protein-protein interaction RE.
• Demonstrated substantial performance improvements due to the model's denoising capabilities.

Conclusions:

• The PACNN+RL model significantly improves performance on distantly supervised biomedical RE tasks.
• The model's denoising effect is key to its enhanced performance.
• PACNN+RL is expected to be a valuable tool for large-scale RE and biomedical knowledge discovery.
• Source code and a demonstration program are publicly available.