Does Latent Semantic Analysis Reflect Human Associations?

In the past decade, Latent Semantic Analysis (LSA) was used in many NLP approaches with sometimes remarkable success. However, its abilities to express semantic relatedness have been not yet systematically investigated. In this work, the semantic similarity measures as provided by LSA (based on a term-by-term matrix) are compared with human free associations. Three tasks have been performed: (i) correlation with human association norms, (ii) discrimination of associated and unassociated pairs and (iii) prediction of the first human response. After a presentation of the results a closer look is taken to the statistical behavior of the data, and a qualitative (example-based) analysis of the LSA similarity values is given as well.