Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SciGA: A Comprehensive Dataset for Designing Graphical Abstracts in Academic Papers

Takuro Kawada, Shunsuke Kitada, Sota Nemoto, Hitoshi Iyatomi

Abstract

Graphical Abstracts (GAs) play a crucial role in visually conveying the key findings of scientific papers. Although recent research increasingly incorporates visual materials such as Figure 1 as de facto GAs, their potential to enhance scientific communication remains largely unexplored. Designing effective GAs requires advanced visualization skills, hindering their widespread adoption. To tackle these challenges, we introduce SciGA-145k, a large-scale dataset comprising approximately 145,000 scientific papers and 1.14 million figures, specifically designed to support GA selection and recommendation, and to facilitate research in automated GA generation. As a preliminary step toward GA design support, we define two tasks: 1) Intra-GA Recommendation, identifying figures within a given paper well-suited as GAs, and 2) Inter-GA Recommendation, retrieving GAs from other papers to inspire new GA designs. Furthermore, we propose Confidence Adjusted top-1 ground truth Ratio (CAR), a novel recommendation metric for fine-grained analysis of model behavior. CAR addresses limitations of traditional rank-based metrics by considering that not only an explicitly labeled GA but also other in-paper figures may plausibly serve as GAs. Benchmark results demonstrate the viability of our tasks and the effectiveness of CAR. Collectively, these establish a foundation for advancing scientific communication within AI for Science.

SciGA: A Comprehensive Dataset for Designing Graphical Abstracts in Academic Papers

Abstract

Graphical Abstracts (GAs) play a crucial role in visually conveying the key findings of scientific papers. Although recent research increasingly incorporates visual materials such as Figure 1 as de facto GAs, their potential to enhance scientific communication remains largely unexplored. Designing effective GAs requires advanced visualization skills, hindering their widespread adoption. To tackle these challenges, we introduce SciGA-145k, a large-scale dataset comprising approximately 145,000 scientific papers and 1.14 million figures, specifically designed to support GA selection and recommendation, and to facilitate research in automated GA generation. As a preliminary step toward GA design support, we define two tasks: 1) Intra-GA Recommendation, identifying figures within a given paper well-suited as GAs, and 2) Inter-GA Recommendation, retrieving GAs from other papers to inspire new GA designs. Furthermore, we propose Confidence Adjusted top-1 ground truth Ratio (CAR), a novel recommendation metric for fine-grained analysis of model behavior. CAR addresses limitations of traditional rank-based metrics by considering that not only an explicitly labeled GA but also other in-paper figures may plausibly serve as GAs. Benchmark results demonstrate the viability of our tasks and the effectiveness of CAR. Collectively, these establish a foundation for advancing scientific communication within AI for Science.

Paper Structure

This paper contains 23 sections, 5 equations, 45 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Dataset, Tasks, and Metric
  4. SciGA-145k Dataset
  5. Task Definition for GA Design Support
  6. Confidence Adjusted top-1 GT Ratio (CAR)
  7. Experiments
  8. Experimental Setup
  9. Intra-GA Recommendation
  10. Inter-GA Recommendation
  11. Benchmark Methods
  12. Results and Discussion
  13. Intra-GA Recommendation
  14. Inter-GA Recommendation
  15. Conclusion
  16. ...and 8 more sections

Figures (45)

  • Figure 1: Example GAs or teasers in our SciGA-145k. These visual summaries highlight key contributions of scientific papers. In addition, SciGA-145k provides access to structured full texts, rich metadata, and non-GA figures, enabling comprehensive visual-language analysis in AI for Science.
  • Figure 2: Intra-GA Recommendation
  • Figure 3: Inter-GA Recommendation
  • Figure 5: Illustration of CAR, our proposed recommendation metric. Each column shows predicted top-5 scores, with the GT highlighted in yellow. CAR assigns partial credit to understandable errors, similar to cases where the outcome is uncertain but still successful (red box), and evaluates clearly correct or incorrect predictions appropriately (blue box). Unlike R@$k$ or nDCG, CAR assigns instance-level continuous scores without graded labels, based on the full score distribution, not just GT rank.
  • Figure 6: GAs / Teasers
  • ...and 40 more figures