Automatic discourse connective detection in biomedical text.

Area of Science:

• Biomedical text mining
• Computational linguistics
• Natural language processing

Background:

• Biomedical text mining systems traditionally focus on clause-level relations.
• Recognizing discourse-level relations is essential for advanced text analysis.
• Discourse connectives are key indicators of discourse relations.

Purpose of the Study:

• Develop and evaluate supervised machine-learning models for identifying discourse connectives in biomedical text.
• Explore domain adaptation techniques to improve connective detection.
• Enhance the understanding of discourse structure in scientific literature.

Main Methods:

• Utilized support vector machines and conditional random fields for classification.
• Trained models on the Biomedical Discourse Relation Bank (24 articles).
• Applied novel domain adaptation techniques using the Penn Discourse Treebank.

Main Results:

• Supervised machine learning successfully identified discourse connectives.
• Domain adaptation techniques achieved the highest performance.
• The best model obtained an F1 score of 0.761.

Conclusions:

• Supervised machine learning is a viable approach for discourse connective detection in biomedical text.
• Domain adaptation significantly boosts the performance of these models.
• The BioConn classifier demonstrates the practical application of these methods.