Presentations are not always linear! GNN meets LLM for Document-to-Presentation Transformation with Attribution

Abstract: Automatically generating a presentation from the text of a long document is a challenging and useful problem. In contrast to a flat summary, a presentation needs to have a better and non-linear narrative, i.e., the content of a slide can come from different and non-contiguous parts of the given document. However, it is difficult to incorporate such non-linear mapping of content to slides and ensure that the content is faithful to the document. LLMs are prone to hallucination and their performance degrades with the length of the input document. Towards this, we propose a novel graph based solution where we learn a graph from the input document and use a combination of graph neural network and LLM to generate a presentation with attribution of content for each slide. We conduct thorough experiments to show the merit of our approach compared to directly using LLMs for this task.

• The paper introduces GDP, a novel method integrating Graph Neural Networks and Large Language Models to convert documents into coherent, non-linear presentations.
• It leverages graph construction, GNN-based paragraph embeddings, and spectral clustering to effectively cluster content before generating slides with iterative LLM prompts.
• Evaluation on the SciDuet dataset shows improved narrative structure and content fidelity over baseline approaches, validating the method's practical impact.

Document-to-Presentation Transformation with GNN and LLM

Introduction

The transformation of long documents into presentations poses significant challenges due to the need for a non-linear narrative structure that effectively captures the document's essence. The paper "Presentations are not always linear! GNN meets LLM for Document-to-Presentation Transformation with Attribution" introduces a novel methodology that leverages Graph Neural Networks (GNN) and LLMs to address these challenges. The aim is to generate presentations that attribute content accurately and maintain coherence despite the non-linear mapping from the source document.

Methodology Overview

The proposed method, GDP (Graph-based Document to Presentation), constructs a graph from the input document where nodes represent paragraphs. GNNs are employed to learn latent semantic relationships between these paragraphs, allowing for effective clustering into coherent slide content. The approach addresses the inherent limitations of LLMs, such as hallucinations and the challenges posed by long input contexts.

Figure 1: A presentation (right) with non-linear narrative and attribution to the source document (left).

Graph Construction and Neural Network Integration

1. Graph Construction: Each paragraph in the document is represented as a node. Edges between nodes are formed based on semantic similarity, quantified using a fine-tuned RoBERTa-based classifier. The threshold for edge creation is determined experimentally to balance graph sparsity and connectivity.
2. Graph Neural Network (GNN) Training: A two-layer Graph Convolutional Network (GCN) processes this graph structure to embed paragraphs into a semantically rich vector space. The unsupervised training objective is to minimize a binary cross-entropy loss over the graph's edges, promoting similar embeddings for semantically linked paragraphs.
3. Clustering via Spectral Clustering: The node embeddings obtained from the GNN are clustered into groups representing slides. Spectral clustering is utilized owing to its ability to handle complex, non-convex data distributions, which are anticipated in the latent paragraph representations.

   Figure 2: Architectural Diagram.

Slide Generation with LLM

Once clusters are established, an LLM, specifically GPT-3.5, enhances the cluster-to-slide transformation. The model is prompted iteratively to generate slides by feeding it text from clustered paragraphs along with summaries of preceding slides to maintain narrative coherence. This step combines the narrative structuring capabilities of neural networks with the natural language generation strengths of LLMs.

Experimental Setup and Evaluation

The proposed method is evaluated on the SciDuet dataset, comprising academic documents and their corresponding presentations. The authors compare GDP against baseline methods, including direct LLM applications:

• Baseline Comparisons: Standard GPT-based approaches, such as GPT-Flat and GPT-COT, perform poorly in terms of content fidelity and narrative flow primarily due to the linear nature of their context processing.
• Performance Metrics: The evaluation utilizes ROUGE scores for lexical matching, Coverage metrics for content completeness, Perplexity for fluency, and a custom non-linearity metric to assess narrative structure.

Non-Linearity and Content Attribution

The GDP methodology demonstrates significant improvements in generating presentations that reflect non-linear narratives akin to human-generated presentations. The non-linearity metric for human-created presentations is approximately 38.6%, while GDP achieves 24.9%, indicating a more narrative-centric arrangement of slides without linear constraints.

Figure 3: Qualitative example to compare the slides generated by a baseline GPT-Flat and our approach GDP from the input document.

Implications and Future Work

This research presents a well-rounded approach that melds GNN's structure learning with LLM's language generation capabilities, effectively addressing both narrative coherence and attribution accuracy. Future directions could explore incorporating multimodal inputs, adapting the framework for diverse document types, and enhancing template selection to enrich slide presentation aesthetics.

Conclusion

The GDP approach marks significant progress in document-to-presentation transformations, effectively managing non-linear narratives and maintaining content fidelity. By leveraging the advanced capabilities of GNNs and LLMs, the methodology overcomes traditional summarization limitations, offering a sophisticated tool for automating presentation generation.