Visual Summary Identification From Scientific Publications via Self-Supervised Learning.

Area of Science:

• Computer Science
• Scientific Literature Analysis

Background:

• Scientific literature is growing exponentially, necessitating tools for efficient analysis.
• Graphical abstracts (visual summaries) can aid in understanding research.
• Previous methods for automatic graphical abstract identification were limited to the biomedical domain due to reliance on labeled data.

Purpose of the Study:

• To develop a novel benchmark dataset for visual summary identification across computer science domains.
• To propose a new self-supervised learning approach for identifying central figures in scientific publications.
• To facilitate domain transfer for visual summary identification tasks.

Main Methods:

• Creation of a new benchmark dataset comprising computer science conference papers.
• Development of a self-supervised learning approach to match in-text figure references with captions.
• Pre-training on a large unlabeled corpus of scientific publications.

Main Results:

• The self-supervised pre-training approach effectively reduces the need for large annotated datasets.
• The method demonstrates successful domain transfer, performing well on both computer science and biomedical datasets.
• The proposed method outperforms previous state-of-the-art methods without requiring task-specific annotations.

Conclusions:

• Self-supervised learning offers a viable solution for visual summary identification, overcoming domain limitations.
• The developed benchmark dataset and method advance the field of automated scientific literature analysis.
• This work paves the way for more accessible and efficient understanding of scientific research through visual summaries.