Skip to content

metadata.md

Latest commit

 

History

History
130 lines (90 loc) · 5.28 KB

metadata.md

File metadata and controls

130 lines (90 loc) · 5.28 KB

Metadata Processing

The Posters.science platform uses advanced AI-powered processing to automatically extract, validate, and enrich metadata from scientific posters. This ensures FAIR compliance and optimal discoverability.

AI-Powered Extraction

Large Language Model

The extraction tool uses Llama 3.1 8B for structured information extraction from poster PDFs. The model is deployed on local GPU hardware using 4-bit quantization for optimized inference, targeting under 60 seconds per poster.

Model Performance

Through systematic evaluation on 400 posters from Zenodo and Figshare, Llama 3.1 8B significantly outperformed traditional NLP methods (Grobid) across all fields.

Extraction Strategy

The system uses adaptive prompting with specialized templates for different extraction tasks. Few-shot learning includes 3-5 example extractions selected based on similarity. Key strengths include:

  • Handling irregular layouts
  • Understanding scientific terminology
  • Robustness to typos
  • Extracting information from context

Content Analysis Pipeline

graph TD
    A[PDF Upload] --> B[OCR Processing]
    B --> C[Text Extraction]
    C --> D[Structure Recognition]
    D --> E[Entity Extraction]
    E --> F[Metadata Generation]
    F --> G[Validation]
    G --> H[User Review]

    I[Abstract Text] --> J[Natural Language Processing]
    J --> K[Keyword Extraction]
    K --> L[Topic Classification]
    L --> F
Loading

Features

  • High-Accuracy Text Recognition: Advanced OCR for poster text extraction
  • Multi-language Support: Recognition of various languages and scripts
  • Layout Preservation: Maintains document structure and formatting
  • Image Processing: Handles charts, graphs, and visual elements

Natural Language Processing

  • Entity Recognition: Automatic identification of authors, institutions, and dates
  • Topic Classification: AI-powered subject categorization
  • Keyword Extraction: Automatic keyword and phrase identification
  • Abstract Summarization: Content summarization and enhancement

Processing Workflow

1. Initial Extraction

  • File Analysis: PDF structure and content analysis
  • Text Recognition: OCR processing and text extraction
  • Layout Detection: Section and element identification
  • Quality Assessment: Content completeness evaluation

2. Entity Recognition

  • Author Extraction: Name and affiliation identification
  • Institution Detection: Organization and department recognition
  • Date Parsing: Publication and event date extraction
  • Location Identification: Geographic and venue information

3. Content Analysis

  • Topic Classification: Subject area determination
  • Keyword Generation: Automatic keyword extraction
  • Abstract Enhancement: Content summarization and improvement
  • Reference Parsing: Citation and bibliography analysis

4. Validation and Enrichment

  • Data Validation: Format and completeness checking
  • External Lookup: ORCID, DOI, and database integration
  • Standardization: Controlled vocabulary mapping
  • Quality Scoring: Metadata completeness assessment

Confidence Scoring

Each extracted field receives a confidence score (0-100%) based on:

  • Format Compliance: Adherence to expected data formats
  • Database Validation: Verification against external databases
  • Context Analysis: Consistency with surrounding information

Fields below 70% confidence trigger user review during the submission process.

External Database Integration

The system validates and enriches metadata through integration with authoritative databases:

  • ORCID: Author identification and profile linking
  • ROR (Research Organization Registry): Institution validation and standardization
  • Crossref Funder Registry: Funding agency validation

For U.S. federal funding, the system cross-references:

  • NIH Reporter: National Institutes of Health grant information
  • NSF Award Search: National Science Foundation award data

Content Structure

Poster content is stored in a posterContent object with the following structure:

  • posterTitle: The main title text extracted from the poster itself
  • sections: Array of content sections, each with:
    • sectionTitle: Heading as it appears on the poster (e.g., "Introduction", "Methods", "Results & Discussion", "Conclusions")
    • sectionContent: Text content of that section

This flexible structure accommodates the diverse, unstructured nature of scientific posters without forcing content into predefined rigid sections.

Image and Table Captions

Captions are stored in separate arrays:

  • imageCaption: Array of objects containing multi-line image captions (caption1, caption2, etc.)
  • tableCaption: Array of objects containing multi-line table captions (caption1, caption2, etc.)

Technical Implementation

The metadata extraction tool is containerized in Docker with Python 3.10+, exposing RESTful API endpoints. Key features include:

  • Redis-based job queuing: Asynchronous processing of extraction requests
  • Rate limiting: Prevention of abuse and resource exhaustion
  • Horizontal scaling: Support for multiple GPU instances
  • API endpoints: RESTful interface for submission integration