Predicting Spanish lexical-affective values by distributional word vectors

YERRO, M.; VIVAS, J.; GONZÁLEZ, M. A.; PASSONI, I.

Chicago

Encuentro; Psychonomic 61 St. Annual Meeting; 2020

Psychonomic Society

The study of variables associated with semantic concepts is an arduous process. This fact is usually translated into databases that are limited in their extension. Based on Recchia and Louwerse (2015), this work explores the possibility of inferring two highly studied variables in lexical-affective norms, valence and activation (arousal), from word vectors enriched with subword information (Bojanowski et al., 2017) trained on Spanish corpora.Using lexical-affective values obtained from the Spanish adaptation of ANEW (Redondo et al., 2007), and word vectors from fastText?s Spanish pre-trained model (Grave et al., 2018), we trained a neural network with 85% of the words available, and held the remaining 15% as a test set. The correlation values found between the predicted and the real values of the test set (valence r ~ .85; arousal r ~.66) are equivalent to those found between different norms, whose residual variance is commonly associated with sampling differences. The use of this method is proposed to approximate affective values when no data norms are available for the treated set. Finally, we provide the predicted values for the 400 nouns of the Spanish semantic feature production norms (Vivas et al., 2017).