starling-labs/Starling-Entity-Tagger

Starling-Entity-Tagger: Biomedical Entity Extraction Model

Starling-Entity-Tagger is a 20 billion parameter model developed by starling-labs, specifically engineered for high-precision entity extraction from scientific and biomedical texts. It processes input in a paragraph-by-paragraph format, extracting a wide range of biomedical entities and outputting them as a structured JSON dictionary.

Key Capabilities

• Comprehensive Entity Types: Extracts entities across 20 distinct biomedical categories, including Anatomy, Antibody, Disease, Gene, Protein, SmallMolecule, and Pathway.
• Structured JSON Output: Generates a JSON object with an "entities" array, where each entity includes name, synonyms, ent_type, organism, surface_forms, alias, resolved_name, and paragraphs.
• Contextual Aggregation: Aggregates mentions of the same entity across multiple paragraphs, providing a consolidated view.
• Sophisticated Entity Handling: Manages composite entities (e.g., fusion proteins) and splits delimited entities (e.g., "A/B/C") into distinct items when appropriate.
• Detailed Entity Attributes: Provides specific attributes for entities like surface_forms (for molecular entities), alias, and resolved_name (for small molecules).

Good For

• Biomedical Information Extraction: Ideal for researchers and developers needing to extract structured data from scientific literature, clinical notes, or patents.
• Knowledge Graph Construction: Facilitates the creation and enrichment of biomedical knowledge graphs by providing normalized and categorized entities.
• Automated Data Curation: Useful for automating the process of identifying and cataloging key biological and chemical entities from large text corpora.