PosterSentry: A Lightweight Multimodal Classifier for Scientific Poster Quality Assurance in Open Repositories

Description

Abstract

Scientific posters are one of the most common ways researchers communicate their results. It is estimated that millions are presented annually at conferences. They contain rich, early-stage knowledge that could be extremely valuable. Yet, posters are rarely shared and reused. To address this, we began developing Posters.science, a platform to facilitate the sharing of FAIR (Findable, Accessible, Interoperable, Reusable) posters. The platform is also building an index of FAIR scientific posters by extracting machine-actionable metadata from PDFs stored in open-access repositories. After downloading and parsing more than 30,000 PDFs tagged in the Zenodo and Figshare repositories with labels indicating that they were posters, we identified an enormous quality issue. Many of those tagged PDFs contained non-poster documents such as multi-page research papers, conference programs, abstract books, presentation slides, and other documents that were not posters. This mislabeling likely occurred due to a lack of standardization in how researchers label their work. Figure 1. PosterSentry architecture. Text, visual, and structural features are extracted independently and concatenated into a 542-d vector for classification. Because curating thousands of individual PDFs for accurate classification was impractically time-consuming, we built PosterSentry as a simple, multichannel system capable of determining if a given PDF is a valid scientific poster. We incorporated three separate modalities in PosterSentry to determine if a given PDF contains a scientific poster (Figure 1): 1) static text features, where first-page text is extracted via PyMuPDF and embedded with Potion-base-32m model2vec static embeddings (512 dimensions); 2) visual features consisting of color statistics, edge density, spatial complexity via fast fourier transform (FFT), white space ratio, and color diversity of the rendered page image (15 dimensions); 3) structural features such as page count, dimensions, number of fonts used in the document, number of text blocks in the document and text density of the pdf (also 15 dimensions). Features were normalized using StandardScaler from scikit-learn prior to being inputted into a logistic regression model. This normalization is important because otherwise the massive (512-dimension) static text embedding would dominate the smaller structural/visual features that provided strong discriminatory power when predicting whether or not a given PDF is a poster. For training, we used 3,606 examples (1,803 per class). Positive samples were single-page PDFs tagged as posters in Zenodo or Figshare, using a heuristic assumption of PDF page count and repository metadata. On a stratified 85/15 held-out split, PosterSentry achieved an overall accuracy of 87.3%, with strong performance in both classes (F1: 0.874 for non-posters, 0.871 for posters). Because the test set shares the same heuristic labeling as training, these figures reflect internal consistency rather than ground-truth accuracy. The top structural predictors were page count (-5.49), file size per page (+7.65), and total file size (-5.44); visual predictors included color diversity (+0.95) and edge density (+0.79). When applied to the entire 30,205-pdf database, PosterSentry correctly classified 80.2% of the documents as posters and 19.8% as non-posters. PosterSentry is deliberately simple. The model head is 10 KB (numpy .npz), with a ~60 MB embedding backbone downloaded once. Inference requires no GPU and no deep learning framework: only numpy, scikit-learn, PyMuPDF, and Pillow. In batch mode with multiprocessing, the entire 30K corpus is classified in under 30 minutes. This lightweight design is intentional: PosterSentry sits at the front of the Posters.science pipeline, screening PDFs before expensive LLM-based extraction of metadata . PosterSentry is fully open source (MIT). The model, code, and training data are published on HuggingFace. A limitation of the current training dataset is the lack of manual curation. To address that, we are planning to collaborate next with the Collaborative Metadata (COMET) initiative to create a manually curated benchmark set for rigorous external validation. Future work also includes active learning on borderline cases and integration with repository metadata APIs for multi-signal classification. In this presentation, we will present the development of PosterSentry, results across 30,205 poster PDFs from open repositories, and discuss collaboration with COMET. Posters are among the most common early-communication artifacts in biomedical and bioinformatics research, yet they remain chronically underrepresented in open science infrastructure. PosterSentry directly addresses this gap: by ensuring that records labeled as posters in open repositories actually contain posters, it enables downstream FAIR metadata tools like poster2json to operate reliably at scale. We believe this is directly relevant to the BOSC community.