Calculating semantic relatedness of lists of nouns using WordNet path length

Lists of semantically related words are better recalled on immediate memory tests than otherwise equivalent lists of unrelated words. However, measuring the degree of relatedness is not straightforward. We report three experiments that assess the ability of various measures of semantic relatedness-including latent semantic analysis (LSA), GloVe, fastText, and a number of measures based on WordNet-to predict whether two lists of words will be differentially recalled. In Experiment 1, all measures except LSA correctly predicted the observed better recall of the related than the unrelated list. In Experiment 2, all measures except JCN predicted that abstract words would be recalled equally as well as concrete words because of their enhanced semantic relatedness. In Experiment 3, LSA, GLoVe, and fastText predicted an enhanced concreteness effect because the concrete words were more related; three WordNet measures predicted a small concreteness effect because the abstract and concrete words did not differ in semantic relatedness; and three other WordNet measures predicted no concreteness effect because the abstract words were more related than the concrete words. A small concreteness effect was observed. Over the three experiments, only two measures, both based on simple WordNet path length, predicted all three results. We suggest that the results are not unexpected because semantic processing in episodic memory experiments differs from that in reading, similarity judgment, and analogy tasks which are the most common way of assessing such measures.