information-extraction

Information Extraction

Purpose

Extract structured information (entities, relations, events, clauses) from unstructured scientific and domain-specific text.

Key Datasets

• ChemProt (bigbio/chemprot): Chemical-protein interaction extraction from BioCreative VI; 10 relation types (CPR:3-CPR:9) between chemicals and proteins
• CUAD (atticus-project/cuad): Contract Understanding Atticus Dataset; 41 clause types from 510 legal contracts (CC-BY licensed)
• JNLPBA: Biomedical named entity recognition (protein, DNA, RNA, cell line, cell type)
• SciERC: Scientific entity and relation extraction from AI paper abstracts

Protocol

1. Schema definition — Define target entity types, relation types, and attributes
2. Preprocessing — Sentence segmentation, tokenization, abbreviation expansion
3. Entity recognition — Identify and classify named entities (NER)
4. Relation extraction — Detect relationships between entity pairs (RE)
5. Normalization — Map entities to standard ontologies (MeSH, ChEBI, UniProt)
6. Output structuring — Format as structured JSON, RDF triples, or knowledge graph

Extraction Types

• Chemical-protein interactions: Substrate, inhibitor, agonist, antagonist, activator
• Legal clause extraction: Termination, IP rights, non-compete, indemnification, limitation of liability
• Gene-disease associations: Causal, biomarker, therapeutic target
• Drug-drug interactions: Synergistic, antagonistic, pharmacokinetic

Rules

• Report extraction confidence scores for each entity/relation
• Provide span offsets for traceability back to source text
• Normalize entities to standard identifiers (CAS, UniProt ID, etc.)
• Handle nested entities and overlapping relations
• Validate extracted facts against known databases when possible