SKIMMR

SKIMMR generates graph-based representations of entities extracted from biomedical text to support skim-reading and knowledge discovery in life sciences.

Key Features:

• Dynamic Graph-Based Entity Extraction: Automatically extracts entities using shallow parsing, co-occurrence analysis, and semantic similarity computation to produce dynamic graphs of entities.
• High-Level Network Overview: Constructs networks of articles from weighted binary statements categorized as co-occurrence and similarity and organizes them within a vector-space model.
• Co-Occurrence and Similarity Analysis: Computes co-occurrence weights using point-wise mutual information and similarity statement weights via cosine distance on co-occurrence vectors.
• Fuzzy Indexing: Builds fuzzy indices of terms, statements, and provenance article identifiers to support fuzzy querying and result ranking.
• Automated Experimental Evaluation: Evaluates generated graphs against gold standards derived from PubMed, TREC challenge, and MeSH data.

Scientific Applications:

• Literature skim-reading and knowledge discovery: Enables rapid assimilation and networked exploration of large biomedical literature corpora.
• Spinal Muscular Atrophy research: Demonstrated applicability for extracting and navigating entity networks in Spinal Muscular Atrophy literature.
• Parkinson's Disease research: Demonstrated applicability for extracting and navigating entity networks in Parkinson's Disease literature.

Methodology:

Extracts entities via shallow parsing; generates weighted binary statements from text and organizes them in a vector-space model; computes co-occurrence weights with point-wise mutual information and similarity weights via cosine distance on co-occurrence vectors; constructs co-occurrence vectors; builds fuzzy indices for terms, statements, and provenance article identifiers; and performs automated evaluations against PubMed-, TREC challenge-, and MeSH-derived gold standards.