Introduction
The transformation of scholarly papers into presentation slides is a cognitively demanding task for researchers. While automated text summarization has advanced significantly, Automatic Slide Generation (ASG) remains a complex challenge due to the specific structural requirements of slides. This research by Peking University addresses two primary bottlenecks in ASG: the lack of fine-grained, aligned datasets and the failure of existing models to account for the hierarchical nature of slide content.
Dataset Construction
To facilitate progress in ASG, the authors introduced a new benchmark dataset consisting of 12,267 samples. Unlike previous datasets that paired entire papers with slide decks, this work focuses on fine-grained alignment. Each slide is mapped to a specific region of the paper (typically up to three paragraphs) based on similarity scores. The dataset specifically targets the first two levels of bullet points (L1 and L2), which constitute over 92% of slide content, ensuring the models learn the relationship between primary ideas and supporting details.
Methodology
The study evaluates several architectures to determine the most effective approach for generating structured slide content:
- Two-Stage Model: An abstractive approach that first generates L1 bullet points (the "skeleton") and then uses these points along with the source text to generate L2 sub-points (the "flesh").
- Transformer (Seq2Seq): A standard sequence-to-sequence model that attempts to generate the entire slide hierarchy in one pass using special structural tags.
- Extractive Baselines: These include TextRank, which selects key sentences, and phrase-based methods that extract and organize phrases into a hierarchy.
- Oracle: A theoretical upper bound representing the maximum performance achievable by an extractive approach.
Experimental Results
The models were evaluated using ROUGE-1, ROUGE-2, and ROUGE-L metrics across three evaluation frameworks: Hierarchy, Align, and Penalty.
The results demonstrate the clear superiority of the Two-Stage model. In the Hierarchy metric, the Two-Stage model achieved a ROUGE-1 score of 13.45, more than doubling the performance of the standard Transformer (5.78) and significantly outperforming the best extractive baseline, TextRank (9.89). This confirms that explicitly modeling the hierarchy of slides is more effective than treating them as flat text.
However, a significant gap remains between the Two-Stage model and the Oracle score (20.53). This disparity highlights the difficulty of the ASG task, particularly in capturing the concise, "telegraphic" language characteristic of high-quality presentation slides.
Conclusion and Future Work
This research underscores that hierarchy is a fundamental component of effective slide generation. While the Two-Stage model represents a significant advancement, several challenges remain. Future work should address the inclusion of deeper nesting levels, the integration of visual elements like figures and tables, and the refinement of language models to better mimic the punchy, abstractive style of human-authored slides.
Summary updated