Extracting experimental parameter entities from scientific articles.

Area of Science:

• Environmental Health Sciences
• Toxicology
• Computational Biology

Background:

• Systematic reviews synthesize research evidence but are time-consuming.
• Automating the extraction of experimental details is crucial for efficient evidence synthesis.

Purpose of the Study:

• To evaluate state-of-the-art supervised multi-label sequence classification systems for identifying experimental design factors in environmental health articles.
• To analyze model performance based on lexical diversity, tokenization, and entity dependencies.

Main Methods:

• Three supervised multi-label sequence classification systems were assessed.
• Systems identified 24 experimental design factors across Animal, Dose, Exposure, and Endpoint categories.
• Analysis included lexical diversity, tokenization impact, and inter-entity dependencies.

Main Results:

• Algorithms using embedded representations generally outperformed statistical algorithms.
• Performance was impacted by the lexical diversity of design parameters.
• Challenges remain in handling non-contiguous mentions and diverse terminology.

Conclusions:

• Automated systems show promise for extracting experimental design factors in environmental health.
• Further development is needed to address challenges posed by lexical diversity and complex entity relationships.