Word Relationships are not Created Equally

Construction of knowledge graph from plain text entails identifying relationships among the words. Downstream tasks like finding similarity is sensitive to these relationships. From PubMed abstracts, words are extracted. “Nearest neighbor” words are identified as candidate words using Word2Vec, Glove and FastText. Conceptnet is a popular knowledge graph using which we find relationship between these words. Similarity for each word pair is calculated. Random Effects Model (REM) is applied to study this relationship strata using the similarity scores. Analysis shows that there is heterogeneity among the relationships independent of the base similarity metrics used.