Neural Content Extraction for Poster Generation of Scientific Papers

• The task matters because posters summarize papers through panelized text and graphics, requiring systems to identify important points and arrange them clearly.Automatic generation could reduce manual effort to minor revisions and save time.Source: §1 · p. 1
• Previous poster-generation work mainly addresses layout and panel composition, while content extraction and graphical elements receive insufficient attention [Qiang et al., 2016].Earlier systems may use naive extraction, ignore graphs, or assume each paper section maps directly to one panel.Source: §1 · p. 1
• A publicly unavailable benchmark dataset has hindered further research on automatic poster generation.Source: §1 · p. 1
• The paper creates a dataset for scientific-paper poster generation and proposes a three-step method spanning section filtering, content extraction, and poster composition.The study particularly focuses on extracting important sentences and graphs to form poster panels.Source: §1 · pp. 1–2
• The proposed neural extractive model extracts important sentences and graphs simultaneously, with automatic and human evaluations plus ablations supporting its efficacy.Source: §1 · pp. 1–2

• Paramita et al. classify paper sentences with hand-crafted features and templates, but ignore graphical elements and produce text-only posters.Source: §2 · p. 2
• Qiang et al. focus on panel attributes and layout, but use unsupervised TextRank extraction, require manual graph selection, and rely on unrealistic section-to-panel assumptions [Qiang et al., 2016].Their dataset contains only 85 paper-poster pairs and is not publicly accessible.Source: §2 · p. 2
• Slide generation differs from poster generation because slides use multiple pages, hierarchical point structures, and sometimes animations, whereas posters organize content into static panels [Fu et al., 2021].Source: §2 · p. 2
• Poster generation resembles multimodal summarization because both combine important document content with text and images.Existing multimodal summarization commonly targets news or Wikipedia and uses natural images that are easier to interpret than scholarly figures and tables.Source: §2 · p. 2

• Paper and poster files are collected from public conference webpages, with paper text manually extracted and figures, tables, and captions automatically extracted then manually corrected.Source: §3 · p. 2
• Human aligners map poster panels to paper sections and label whether poster figures and tables originate from the paper, using a third adjudicator when two aligners disagree.Source: §3 · p. 2
• The dataset contains 260 paper-poster pairs and 846 section-panel pairs, with an 82.15% agreement rate for section-panel alignment.Sentence labels optimize ROUGE-2 F1 against panel text, while graph labels come from human aligners.Source: §3 · p. 2
• The dataset represents each sample with section sentences and labels, section figures or tables with captions and labels, and the panel's text, figures, and tables.Source: §3 · pp. 2–3
• The dataset uses 10-fold cross-validation with samples from the same paper kept in the same split to prevent overlap between training, validation, and test sets.Source: §3 · p. 3
• About 55.34% of paper graphs appear in posters, while the overall text compression ratio is 16.87, indicating strong content reduction.Source: §3 · p. 3

4 Methodology

• The three-step pipeline filters unimportant sections, extracts sentences and graphs from the rest, and composes the resulting panels into a poster.Content extraction is the study’s main focus, while section filtering and poster composition are described more briefly.Source: §4 · p. 3

4.1 Section Filter

• Section filtering encodes section text with RoBERTa, applies a transformer for paper-level representations, and classifies each section’s importance.The classifier selects source sections for poster panels.Source: §4.1 · p. 3
• Section classification accuracy is 88.01% for the full model, versus 74.52% using only titles and 84.08% without paper-level representations.These ablations indicate that section content and global paper information improve importance prediction.Source: §4.1 · p. 3

4.2 Content Extraction

• The content extractor jointly predicts importance scores for section sentences and graph captions, using captions instead of raw figures and tables.Captions provide natural-language descriptions that are easier for models to learn than pictorial or numerical contents.Source: §4.2 · p. 3
• RoBERTa encodes sentences and graph captions, while transformer layers learn section-aware representations from their concatenated inputs.The model assumes section context and cross-modal relevance help determine the importance of both text and graphs.Source: §4.2 · p. 4
• Additional attention weights emphasize links between graphs and referring sentences, as well as among sentences referring to the same graph.The weights are normalized by the number of referring sentences and are intended to stabilize and improve performance.Source: §4.2 · p. 4
• Balanced cross-entropy losses handle class imbalance, and the model jointly optimizes sentence and graph losses with a weighted graph-loss term.The graph-loss weight is larger because graph captions are fewer than section sentences.Source: §4.2 · p. 4

4.3 Poster Composition

• Poster composition lets users choose portrait or landscape orientation and selects predefined templates using panel, text, and graph attributes.The system outputs a LaTeX document using tikzposter, which users can further modify after compilation.Source: §4.3 · p. 4

5 Evaluation Setup

• The evaluation concentrates on content extraction and uses baseline comparisons plus ablation studies to assess model components.Separate comparisons are defined for sentence extraction and graph extraction.Source: §5 · p. 4

5.1 Implementation Details

• The encoder has three layers with d_k = 768 and d_ff = 1536, while h_1 and h_2 are set to 1e-2 and 1e-3.Training uses Adam with learning rate 8e-5 and loss settings α = 1.0, β = 0.5 for sentences, α = β = 1.0 for graphs, and γ = 3.0.Source: §5.1 · p. 4

5.2 Compared Models for Text Extraction

• Text-extraction comparisons include Lead-3, TextRank, BERTSum, RoBERTaSum, and GPT-2.The set spans simple extractive, unsupervised, pretrained encoder, and abstractive approaches.Source: §5.2 · pp. 4–5
• Text-extraction ablations remove additional attention weights, graph captions, or replace caption encoding with image features, while an oracle provides an extractive upper bound.The image variant uses ResNet for graph features, whereas the proposed model encodes graph captions with RoBERTa.Source: §5.2 · p. 5

5.3 Compared Models for Graph Extraction

• Graph-extraction comparisons include a caption–sentence similarity threshold model and a RoBERTa classifier that uses captions without section text.Additional ablations remove section sentences, attention weights, or use image features.Source: §5.3 · p. 5

6.1 Automatic Evaluation

• The proposed model outperforms all baseline models on text extraction, with jointly trained text and graph extraction benefiting both tasks.The authors attribute the improvement to the close connection between text and graph extraction during training.Source: §6.1 · pp. 5–6
• The full model outperforms the ablation without captions, while explicit sentence–graph attention weights provide slightly better and more stable performance.Encoding the whole section can cause truncation at BERT’s 512-token limit, whereas graph information also benefits text extraction.Source: §6.1 · p. 5
• RoBERTa achieves slightly higher graph-extraction accuracy than the similarity-based model, while image-based extraction performs relatively poorly.The results indicate that textual similarity helps identify important graphs, whereas image features alone are less effective.Source: §6.1 · pp. 5–6

6.2 Human Evaluation

• Human judges find that the model performs better than or as well as baselines on informativeness, fluency, and non-redundancy.The evaluation compares the model with TextRank, RoBERTaSum, and GPT-2 using blinded win, lose, or tie judgments.Source: §6.2 · p. 6

6.3 Case Study

• The generated panel text captures important factors such as batch size, multiple GPUs, and gradient accumulation, whereas RoBERTaSum extracts only a less important transition sentence.The model still misses some points because of length limitations.Source: §6.3 · p. 6
• The generated poster extracts important points with related graphs, and its panels are well placed with a reasonable layout.This example illustrates the output of the paper’s three-step generation method.Source: §6.3 · p. 6, Figure 3 · p. 6

7 Conclusion

• The study builds a dataset for automatic poster generation and designs a three-step method to address limitations of previous work.The framework particularly focuses on extracting sentences and graphs from paper sections.Source: §7 · p. 6
• The proposed neural model extracts sentences and graphs simultaneously, and evaluation results verify its effectiveness.Future work will enlarge the dataset and investigate merging the three steps into one end-to-end model.Source: §7 · p. 7