Abstract
PosterBot is an end-to-end automated system designed to transform dense scientific papers into visually organized academic posters. By using a three-stage neural pipeline—filtering sections, extracting key text and figures, and composing the final layout—it generates editable LaTeX posters that help researchers communicate their work more efficiently.
Introduction
Creating an academic poster is a time-consuming task that requires researchers to condense months of work into a single, readable sheet. While automatic slide generation has seen significant progress, the "poster problem" remains under-investigated. Previous attempts at automation often suffered from two major flaws: the "One-to-One" fallacy, where every paper section is incorrectly assumed to need a corresponding poster panel, and poor content extraction that fails to link text with supporting figures.
Methodology
PosterBot utilizes a three-step neural architecture to mimic the human design process:
1. Section Filtering
The system first determines which sections are "poster-worthy." It employs a RoBERTa model to encode individual section text, followed by a Transformer encoder to capture paper-level context. A classifier then scores each section's importance, allowing the system to omit less critical sections like "Related Work" when space is limited. Labels for this training were created via human evaluators reviewing existing papers and posters.
2. Content Extraction
For selected sections, the model identifies the most important sentences and graphics. A key innovation is the use of a reference-aware attention mechanism. Since raw images are often too abstract for current models to interpret in a scientific context, PosterBot treats figure captions as text. If a sentence explicitly mentions a figure (e.g., "as shown in Fig 2"), the model boosts the attention weight between that sentence and the corresponding figure's caption.
3. Poster Composition
Finally, the system fits the extracted content into pre-defined templates. It calculates the optimal layout based on panel count and word length, outputting a LaTeX document using the tikzposter package. This ensures the final product is a "first draft" that researchers can manually refine.
Key Insights and Results
The system was demonstrated at AAAI-22, showcasing its ability to condense complex papers into logical, five-panel layouts (Introduction, Models, Training Techniques, Experiments, and Conclusion).
A primary insight from this work is that figure captions are often more semantically useful than the pixels of the graphs themselves for determining content importance. Furthermore, the use of a Transformer encoder allows the model to understand a section's importance based on its hierarchical position within the entire paper rather than just its internal text.
Conclusion and Future Work
PosterBot represents a significant step toward automating academic design. Current limitations include the pipeline nature of the system, where errors in early stages can cascade. Future work aims to develop a single, unified model for extraction and composition and to integrate advanced computer vision models to better interpret pictorial information within scientific graphs.
Summary updated