PACNN+RL

PACNN+RL integrates a Piecewise Attentive Convolutional Neural Network and reinforcement learning to denoise distantly supervised biomedical relation extraction and improve relation prediction accuracy.

Key Features:

• Piecewise Attentive Convolutional Neural Network (PACNN): Encodes semantic information from biomedical text using piecewise attention to focus on relevant text segments.
• Reinforcement Learning with Memory Backtracking: Incorporates a reinforcement learning module with memory backtracking to revisit and refine instance-level predictions.
• Instance-label Correlation Modeling: Models correlations between instances and noisy distant supervision labels to improve label quality.
• Distant Supervision from UMLS and MEDLINE: Trains on distantly supervised data derived from the Unified Medical Language System (UMLS) combined with MEDLINE abstracts for large-scale learning.
• Denoising Capability: Reduces noise in distantly supervised training data to enhance relation extraction performance.

Scientific Applications:

• Distantly supervised biomedical relation extraction: Applies denoising and relation prediction methods to distantly supervised biomedical corpora.
• Drug–disease relation extraction (may-prevent, may-treat): Evaluated on may-prevent (F1-score 0.5592) and may-treat (F1-score 0.6666) datasets.
• Drug–drug interaction extraction (DDI 2011): Evaluated on the DDI corpus from 2011 with an F1-score of 0.3838.
• Protein–protein interaction relation extraction: Achieved state-of-the-art performance on four out of five protein–protein interaction datasets.

Methodology:

PACNN+RL combines a piecewise attentive convolutional neural network with reinforcement learning incorporating a memory backtracking mechanism, trained on distantly supervised data derived from UMLS and MEDLINE abstracts to model instance-label correlations.