Language Style Matters: Personality Prediction from Textual Styles Learning

Personality prediction is a fundamental task for many scientific disciplines and behavioral economics. The main-stream studies predict users' personality traits through their texts on social media, and focus on understanding the semantic information in the textual language. Few previous research has addressed the language styles are effective in personality prediction. The language styles, which had been widely studied in many psycholinguistic literatures, can reveal some personality traits of users, such as social connections, and psychological state. Based on this consideration, we make full use of the semantic and language styles of the textual to make a prediction of user personality traits. However, two crucial issues should be addressed: (1) how to characterize personality-related language styles; (2) how to leverage the language styles to enhance textual semantic understanding. To characterize the language styles, we make a priority attempt to construct a personality graph with the language style information extracted from influential literature in psycholinguistics. To use it, in this work, we propose a novel hierarchical graph attention network for personality prediction (PerHGAT), to aggregate the language style information into the semantic learning of words in user posts, and then aggregate the semantic information of words into user representation, enables PerHGAT to leverage the personality knowledge of language styles, which is helpful for personality traits prediction. We conduct extensive experiments on three real-world datasets (Youtube, PAN and MyPersonality), and the proposed method achieves state-of-the-art performances. The outperforming results demonstrate the effectiveness of our model in predicting the personality traits of users.