Deep learning-based personality recognition from text posts of online social networks

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 48; no. 11; pp. 4232 - 4246
• Main Authors: Xue, Di, Wu, Lifa, Hong, Zheng, Guo, Shize, Gao, Liang, Wu, Zhiyong, Zhong, Xiaofeng, Sun, Jianshan
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.11.2018; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Artificial neural networks; Computer Science; Deep learning; Machines; Manufacturing; Mechanical Engineering; Neural networks; Personality; Processes; Recognition; Semantics; Social networks; Statistical analysis; Texts; Deep learning; Personality recognition; Big Five personality; Online social networks
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-018-1212-4

Personality is an important psychological construct accounting for individual differences in people. Computational personality recognition from online social networks is gaining increased research attention in recent years. However, the majority of existing methodologies mainly focused on human-designed shallow statistical features and didn’t make full use of the rich semantic information in user-generated texts, while those texts are exactly the most direct way for people to translate their internal thoughts and emotions into a form that others can understand. This paper proposes a deep learning-based approach for personality recognition from text posts of online social network users. We first utilize a hierarchical deep neural network composed of our newly designed AttRCNN structure and a variant of the Inception structure to learn the deep semantic features of each user’s text posts. Then we concatenate the deep semantic features with the statistical linguistic features obtained directly from the text posts, and feed them into traditional regression algorithms to predict the real-valued Big Five personality scores. Experimental results show that the deep semantic feature vectors learned from our proposed neural network are more effective than the other four kinds of non-trivial baseline features; the approach that utilizes the concatenation of our deep semantic features and the statistical linguistic features as the input of the gradient boosting regression algorithm achieves the lowest average prediction error among all the approaches tested by us.