FigSum: automatically generating structured text summaries for figures in biomedical literature.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Scientific Communication

Background:

• Figures in biomedical articles are crucial but often hard to understand from legends alone.
• Key information supporting figures is frequently dispersed throughout the full-text article.
• Existing methods lack a structured approach to integrate figure-related information.

Purpose of the Study:

• To develop and evaluate FigSum, an automated system for summarizing biomedical figures.
• To enhance the comprehension of figures by aggregating scattered information.
• To generate structured summaries based on rhetorical categories (IMRaD).

Main Methods:

• FigSum processes figures within biomedical articles.
• It extracts and aggregates information scattered across the full-text.
• An automated machine learning classifier predicts sentence categories (Introduction, Methods, Results, Discussion - IMRaD).

Main Results:

• FigSum generates structured text summaries for each figure.
• The system captures 53% of sentences in scientist-annotated gold standard summaries.
• FigSum achieves a ROUGE-1 score of 0.70, outperforming a baseline system.

Conclusions:

• FigSum effectively summarizes biomedical figures by integrating information from full-text articles.
• The system improves figure comprehension through structured, IMRaD-based summaries.
• Automated figure summarization shows promise for enhancing scientific literature accessibility.