Big Five Personality Prediction from Essays through Vocabulary Graphs and Contextual Transformer Representations

This study develops an approach to predict personality traits based on the Big Five model through essay text by combining global and contextual information using a Vocabulary Graph Convolutional Network (VGCN) and Bidirectional Encoder Representations from Transformers (BERT). VGCN is used to capture global relationships between words through a vocabulary graph representation built based on NPMI weights, while BERT plays a role in producing richer and more contextual text representations. Experimental results show that the integration of structural and contextual features can provide good performance in multi-label classification tasks. Based on the grid search process and 5-fold cross-validation, the best model configuration obtained a macro F1 score of 0.6827 and a weighted F1 score of 0.6829 in out-of-fold prediction. The evaluation results for each feature show that the recall value is close to 1.0, while the precision value ranges from 0.50 to 0.56. This condition indicates a tendency for the model to produce more positive predictions. Furthermore, the results of the study highlight the importance of optimizing the prediction threshold value to obtain a more optimal compromise between precision and recall. Overall, the integration of graph-based semantic information with contextual representations from BERT shows that the VGCN-BERT approach has good potential for application in multi-label personality classification