Convolutional Bi-LSTM for Automatic Personality Recognition From Social Media Texts

Automatic Personality Recognition (APR) has received much attention in recent years due to its wide range of important applications across various fields. The growing use of online social networks provides valuable opportunities for APR, as a strong correlation has been found between what users post on these platforms and their personality traits. Consequently, various APR models have been developed to infer the Big Five personality traits from social media user-generated texts. However, most of these models heavily relied on hand-crafted features, which are unable to capture deep contextual information and learn complex patterns from texts. More importantly, the performance of text-based APR is still unsatisfactory, especially at the level of each personality dimension. To tackle this issue, we propose a new model, called APR_ConvLSTM, that aims to improve text-based APR performance by integrating two robust deep learning architectures: CNN and Bi-LSTM. Unlike existing APR models, the APR_ConvLSTM is a unified end-to-end model where all personality traits are predicted simultaneously and effectively without a need for laborious feature engineering. We also developed a new labeled Big Five personality dataset, called X-Big5, which has been in need for a long time in the APR field. Extensive experiments on the X-Big5 and a publicly available benchmark dataset (PAN-2015 Author Profiling) demonstrate the promising performance of our model over its contenders. Overall, the proposed model achieved the highest Accuracy and F-1 score of 79.51% and 86.54% on the PAN-2015 dataset and 87.95% and 81.35%, respectively, on the X-Big5 dataset. Moreover, it shows promising performance over its competitors, with the highest average Accuracy and F-1 score of 79.01% and 80.56%, respectively, on the combined dataset. The model reached competitive results in predicting Openness, Extraversion, Agreeableness, and Neuroticism traits with the highest F1 scores of 88.60%, 77.35%, 76.16%, and 74.52%, respectively, on the combined dataset. The proposed model can positively impact the analysis of social media text generated by different users and help identify their personality traits.