A1390
Title: Application of latent semantic scaling to high-dimensional text data for personality assessment
Authors: Yoshito Tan - University of Tokyo (Japan) [presenting]
Keishi Nomura - Toyo University (Japan)
Kensuke Okada - The University of Tokyo (Japan)
Abstract: Interest in the forced-choice assessment format in high-stakes contexts, such as personality assessment in personnel selection, has been increasing because it can mitigate social desirability bias by matching the social desirability levels of the personality trait words being compared. However, obtaining social desirability ratings beforehand is time- and cost-intensive. To address this issue, leveraging the strong correlation between emotional valence and social desirability of trait words and using latent semantic scaling (LSS) to scale unidimensional valences as a proxy is proposed. As a semi-supervised sentiment analysis technique, LSS is interpretable and cost-efficient because it combines positive and negative seed words for weak supervision and unsupervised learning of word embeddings from high-dimensional text data. Given pre-trained embeddings, LSS can be seen as a simple linear model that applies the cosine similarity matrix of embeddings as a linear operator and maps the initial valences of seed words onto the scaled valences of other words. To evaluate the usefulness of the LSS-based valences, the extent to which they predicted the social desirability scores of trait words in the Big-Five personality assessment is examined. Medium-to-strong correlations are observed between them. These results imply that the proposed method contributes considerably to efficient forced-choice format construction.
