Shallow semantic parsing of randomized controlled trial reports

Area of Science:

• Natural Language Processing
• Machine Learning
• Biomedical Informatics

Background:

• Propbank, a semantic role labeling resource, is typically trained on Wall Street Journal (WSJ) text.
• Applying Propbank to specialized domains like medical abstracts presents challenges due to linguistic differences.
• Automated annotation of Randomized Controlled Trials (RCTs) is crucial for information extraction.

Purpose of the Study:

• To measure the performance of Propbank-based Machine Learning (ML) for annotating RCT abstracts.
• To assess the impact of domain transfer (WSJ to medical) on ML annotation accuracy.
• To evaluate the feasibility of using existing Propbank resources in the biomedical domain.

Main Methods:

• Utilized Propbank, a semantic role annotation corpus.
• Trained and tested ML models on both intra-domain (WSJ/WSJ) and cross-domain (WSJ/medical abstract) datasets.
• Compared annotation performance metrics between the two domain settings.

Main Results:

• Intra-domain performance (WSJ/WSJ) achieved superior accuracy.
• Cross-domain performance (WSJ/medical abstract) demonstrated reasonable effectiveness.
• The study identified performance variations when applying a general linguistic resource to specialized medical text.

Conclusions:

• Propbank-based ML models can achieve acceptable performance for annotating medical abstracts, even with cross-domain transfer.
• Domain adaptation strategies may further enhance the accuracy of ML annotation in biomedical NLP.
• This research highlights the potential and limitations of leveraging general-domain semantic resources for specialized scientific text processing.