Semi-supervised learning of causal relations in biomedical scientific discourse

Area of Science:

• Biomedical Natural Language Processing
• Computational Biology
• Scientific Discourse Analysis

Background:

• The exponential growth of biomedical literature necessitates automated text analysis.
• Bio-text mining tools extract entities and events but struggle with factual connections.
• Understanding causal relationships in scientific text remains a significant challenge.

Purpose of the Study:

• To develop and evaluate a method for recognizing causal triggers and their arguments in biomedical scientific discourse.
• To enhance the accuracy of automated causal relation extraction.

Main Methods:

• A novel self-learning approach is introduced for identifying causal triggers.
• New features are incorporated to improve machine learning performance.
• The method is evaluated against supervised and rule-based approaches.

Main Results:

• The self-learning approach achieved 83.47% performance in recognizing causal triggers.
• The method demonstrated improved accuracy in recognizing spans of causal arguments compared to existing methods.

Conclusions:

• Leveraging unlabelled biomedical data can significantly enhance causal discourse relation recognition.
• Improved causal relation extraction supports tasks like hypothesis generation and contradiction detection.
• This work contributes to building more comprehensive causal networks from scientific literature.