Benchmarking table recognition performance on biomedical literature on neurological disorders

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Data Extraction

Background:

• Table recognition systems extract quantitative data from documents.
• Existing systems struggle with complex biomedical tables due to limited benchmark data.
• Biomedical literature contains intricate tables requiring specialized extraction methods.

Purpose of the Study:

• To introduce a novel, curated benchmark dataset for evaluating table extraction in the biomedical domain.
• To address the scarcity of training and benchmark data for complex scientific tables.
• To facilitate the development and assessment of advanced table recognition applications.

Main Methods:

• Developed a highly curated benchmark dataset from a hand-curated literature corpus on neurological disorders.
• Evaluated state-of-the-art table extraction systems using the proposed benchmark.
• Introduced a new performance metric and improvements for evaluation procedures.

Main Results:

• The benchmark dataset enables tuning and evaluation of table extraction applications for complex biomedical tables.
• Evaluation of existing systems highlighted challenges in recognizing intricate table structures.
• The proposed evaluation metric and improvements enhance performance assessment.

Conclusions:

• The created benchmark dataset is crucial for advancing table recognition in the biomedical field.
• Openly accessible dataset and source code promote further research and development.
• Addressing challenges in complex table recognition is vital for scientific data extraction.