Predicting Personality from Text

Recent advances in natural language processing (NLP) have enabled new possibilities for automated personality assessment. Yet, the relationship between linguistic patterns and personality traits remains incompletely understood. With the specific aim of addressing the incomplete relationship we use the Big Five model and Schwartz’s value categories to predict personality traits from text data, to classify individuals into four different tribes: Nerds, Fatherlanders, Treehuggers, and Spiritualists. Using a dataset of YouTube video transcripts, we developed and evaluated a machine learning model that uses BERT-based embeddings to predict personality traits. Our model achieved an accuracy of 75%, a recovery rate of 80%, and an F1-score of 0.77 with a Hamming loss of 20%, indicating a balanced performance on false-negatives and false-positives. Analysis of linguistic similarities revealed significant overlap between Fatherlands and Treehuggers in terms of cosine and sequence similarity, as well as between Treehuggers and Spiritualists, as indicated by Manhattan and Minkowski distance measures. These results show that personality-based tribal boundaries are more fluid than previously thought, especially in cases where tribes share a common argumentative background, while the Nerd tribe has more distinct linguistic patterns that allow for clearer classification. Further research is needed to explore these nuanced interconnections.