Measuring associational thinking through word embeddings

Abstract The development of a model to quantify semantic similarity and relatedness between words has been the major focus many studies in various fields, e.g. psychology, linguistics, natural language processing. Unlike measures proposed by most previous research, this article is aimed at estimating automatically strength associative that can be semantically related or not. We demonstrate performance depends not only on combination independently constructed word embeddings (namely, corpus- network-based embeddings) but also way these vectors interact. research concludes weighted average cosine-similarity coefficients derived from independent double vector space tends yield high correlations with human judgements. Moreover, we evaluating associations through measure relies rank ordering pairs reveal some findings go unnoticed traditional such as Spearman’s Pearson’s correlation coefficients.