RECEEM

RECEEM reconstructs entities omitted by coordination ellipsis in medical text to improve concept normalization across biomedical corpora.

Key Features:

• Graph-Based Representation Model: Represents coordinated elliptical expressions as graphs to enable identification and analysis of candidate conjunct pairs.
• Unsupervised Approach: Operates as an unsupervised, graph-based method for reconstruction of entities from coordination ellipsis.
• Identification Module: Systematically identifies all potential candidate conjunct pairs from coordinated elliptical expressions.
• Coefficient Calculation: Computes coefficients for each candidate conjunct using an embedding model to assess relevance and accuracy.
• Global Optimization: Selects the most appropriate decompositions through a process of global optimization.
• Pathfinding Algorithm: Rebuilds concepts by navigating the graph structure with a pathfinding algorithm.

Scientific Applications:

• Evaluation on Biomedical Corpora: Evaluated on biomedical literature, clinical narratives, and eligibility criteria from clinical trials with F1 scores of 0.862, 0.721, and 0.870 respectively.
• Enhancement of NLP Tools: Integration into two existing NLP tools improved F1 scores from 0.248 to 0.460 and from 0.287 to 0.630 on a dataset of 1,125 coordination ellipses.

Methodology:

Uses an unsupervised graph-based model to represent coordinated elliptical expressions, identifies candidate conjunct pairs, computes coefficients with an embedding model, selects decompositions via global optimization, and reconstructs concepts using a pathfinding algorithm.