Bayesian analysis of ConceptNet relations on PubMed dataset

Abstract

ConceptNet is a popular knowledge graph built using crowd sourcing. To construct a knowledge graph from plain text relationship identification between words is a critical task. Downstream tasks like finding similarity is sensitive to these relationships. From PubMed abstracts, words are extracted and stop words removed. Using Glove (word vector) “Nearest neighbor” words are identified as candidate words to this root PubMed word. Relationship between these words is identified via numberbatch vectors of ConceptNet. Similarity for each word pair is calculated. Bayesian Random Effects Model (REM) is used to study this relationship strata. Analysis shows that there is heterogeneity among the relationships.