Automatic Personality Evaluation from Transliterations of YouTube Vlogs Using Classical and State-of-the-Art Word Embeddings

Abstract

The study of automatic personality recognition has gained attention in the last decade thanks to a variety of applications deriving from this field. The Big Five model (also known as OCEAN) constitutes a well-known method to label different personality traits. This work considered transliterations of video recordings collected from YouTube (originally provided by the Idiap research institute) and automatically generated scores for the Big Five personality traits, which were also in the database. The transliterations were modeled with three different word embedding approaches (Word2Vec, GloVe, and BERT) and three different levels of analysis, namely a regression to predict the score of each personality trait, a binary classification between the strong vs. weak presence of each trait, and a tri-class classification according to three different levels of manifestations in each trait (low, medium, and high). According to our findings, the proposed approach provides similar results to others reported in the specialized literature. We believe that further research is required to find better results. Our results, as well as others reported in the literature, suggest that there is a big gap in the study of personality traits based on linguistic patterns, which highlights the need to work on collecting and labeling data considering the knowledge of expert psychologists and psycholinguists.