An attentive joint model with transformer-based weighted graph convolutional network for extracting adverse drug...

Area of Science:

• Natural Language Processing
• Biomedical Informatics
• Computational Linguistics

Background:

• Adverse drug event (ADE) relation extraction is vital for drug safety surveillance.
• Current graph convolutional network (GCN) methods struggle with complex relations and underutilize syntactic information.
• Existing approaches often treat relations independently, ignoring inter-relational dependencies.

Purpose of the Study:

• To propose an attentive joint model, ADERel, for improved ADE relation extraction.
• To address limitations in exploiting syntactic information and handling complex, nested, or discontinuous ADE mentions.
• To enhance the modeling of interactions between multiple relations within medical texts.

Main Methods:

• Formulating ADE relation extraction as an N-level sequence labeling task.
• Developing a joint neural model that combines Bidirectional Encoder Representations from Transformers (BERT) with a weighted GCN (WGCN).
• Utilizing a weighted GCN to score dependency edges, capturing influential syntactic features.
• Employing multi-head attention for knowledge exchange across different relation levels.

Main Results:

• ADERel demonstrated superior performance over state-of-the-art methods on benchmark datasets (TAC 2017, n2c2 2018).
• The integration of transformer models with WGCN proved effective for diverse ADE relation types.
• Joint learning significantly improved the recognition of complex ADE relations.

Conclusions:

• ADERel offers a more effective approach to ADE relation extraction by leveraging contextual, structural, and syntactic information.
• The proposed model successfully handles complex relations and captures inter-relational dependencies.
• ADERel advances the field of drug safety surveillance through improved automated analysis of medical texts.