Personality Identification from Social Media Using Ensemble Bert and Roberta

Abstract

Social media growth was fast because many people use it to express their feelings, shared information, and interact with others. With this growth of social media, many researchers are interested in using social media data to conduct research about personality identification. The identification result can be used as a parameter to screen candidate attitudes in the company recruitment process. Some approaches were used for research about personality; one of the most used is Big Five Personality. In this research, an ensemble model between BERT and RoBERTa was introduced for personality prediction from the Twitter and Youtube datasets. Data augmentation method also introduce for handling the imbalance class for each dataset. Pre-trained model BERT and RoBERTa was used as the feature extraction method and modeling process. To predict each trait in Big Five Personality, the voting ensemble from BERT and RoBERTa achieved an average f1 score 0,730 for Twitter dataset and 0,741 for Youtube dataset. Using the proposed model, we conclude that data augmentation can increase average performance compared to the model without data augmentation process.