We compared the quality of prediction of word variables based on a Dutch word association and text corpus. Wederived estimates for: valence, arousal, dominance, concreteness and age of acquisition (AoA) for 2831 words. Based on thesimilarity between words we: (1) used projections on a dimension identified as the variable in question in a multidimensionalrepresentation, (2) used the k-nearest neighbors values, weighted according to their proximity. Estimates prevailed when basedon word associations. Differences between the predictions of the two methods were small. Based on the word association corpusit yielded correlations of .92, .85, and .85, for valence, arousal, and dominance, respectively. Its corresponding correlationsbased on the text corpus were .80, .74, and .67. For concreteness and AoA, both the association and the text corpus yieldedcorrelations of .88 and .73, respectively. This suggests word associations are better at capturing human ratings of affective wordvariables.

Obtaining norm scores for subjective properties of words can be quite cumbersome as it requires a considerable investment proportional to the size of the word set. We present a method to predict norm scores for large word sets from a word association corpus. We use similarities between word pairs, derived from this corpus, to construct a semantic space. Starting from norm scores for a subset of the words, we retrieve the direction in the space that optimally reflects the norm data associated with the words. This direction is used to orthogonally project all the other words in the semantic space on, providing predictions of the words on the variable of interest. In this study, we predict valence, arousal, dominance, age of acquisition, and concreteness and show that the predictions correlate strongly with the judgments of human raters. Furthermore, we show that our predictions are superior to those derived using other methods