Big five personality prediction based on pre-training language model and sentiment knowledge base

Abstract

Machine learning-based personality prediction methods have become the mainstream prediction method. But existing Big Five personality prediction methods based on user-generated text rely too much on pre-trained language models to ignore the sentiment information. To tackle these issues, we propose a novel Big Five personality prediction method based on pre-trained language models and SenticNet6, which is a new sentiment knowledge base. The correlation analysis and ablation experiment shows that SenticNet6 can be used for Big Five personality prediction tasks. The experiment results indicate that our proposed method improves the state-of-art results for each Big Five personality trait by 9.89%, 9.39%, 2.99%, 3.01%, and 4.28% on Essays dataset, respectively.