Psycholinguistic Tripartite Graph Network for Personality Detection

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4229–4239 August 1–6, 2021. ©2021 Association for Computational Linguistics 4229 Psycholinguistic Tripartite Graph Network for Personality Detection Tao Yang, Feifan Yang, Haolan Ouyang, Xiaojun Quan∗ School of Computer Science and Engineering, Sun Yat-sen University, China {yangt225,yangff6,ouyhlan}@mail2.sysu.edu.cn quanxj3@mail.sysu.edu.cn Abstract Most of the recent work on personality de- tection from online posts adopts multifarious deep neural networks to represent the posts and builds predictive models in a data-driven manner, without the exploitation of psycholin- guistic knowledge that may unveil the connec- tions between one’s language usage and his psychological traits. In this paper, we propose a psycholinguistic knowledge-based tripartite graph network, TrigNet, which consists of a tri- partite graph network and a BERT-based graph initializer. The graph network injects struc- tural psycholinguistic knowledge from LIWC, a computerized instrument for psycholinguis- tic analysis, by constructing a heterogeneous tripartite graph. The graph initializer is em- ployed to provide initial embeddings for the graph nodes. To reduce the computational cost in graph learning, we further propose a novel flow graph attention network (GAT) that only transmits messages between neigh- boring parties in the tripartite graph. Benefit- ing from the tripartite graph, TrigNet can ag- gregate post information from a psychological perspective, which is a novel way of exploit- ing domain knowledge. Extensive experiments on two datasets show that TrigNet outperforms the existing state-of-art model by 3.47 and 2.10 points in average F1. Moreover, the flow GAT reduces the FLOPS and Memory mea- sures by 38% and 32%, respectively, in com- parison to the original GAT in our setting. 1 Introduction Personality detection from online posts aims to identify one’s personality traits from the online texts he creates. This emerging task has attracted great interest from researchers in computational psycholinguistics and natural language processing due to the extensive application scenarios such as ∗Corresponding author. Post-1 Post-2 Function Quant Affect Social Drives Post Node Word Node Category Node of me thanks love sharing it for a lot good advice Figure 1: An example of our tripartite graph. The con- tent of Post-1 and Post-2 are “A lot of good advise for me.” and “Love it! Thanks for sharing!”, respectively. personalized recommendation systems (Yang and Huang, 2019; Jeong et al., 2020), job screening (Hiemstra et al., 2019) and psychological studies (Goreis and Voracek, 2019). Psychological research shows that the words peo- ple use in daily life reflect their cognition, emotion, and personality (Gottschalk, 1997; Golbeck, 2016). As a major psycholinguistic instrument, Linguistic Inquiry and Word Count (LIWC) (Tausczik and Pennebaker, 2010) divides words into psycholog- ically relevant categories (e.g., Function, Affect, and Social as shown in Figure 1) and is commonly used to extract psycholinguistic features in con- ventional methods (Golbeck et al., 2011; Sumner et al., 2012). Nevertheless, most recent works (Her- nandez and Knight, 2017; Jiang et al., 2020; Keh et al., 2019; Lynn et al., 2020; Gjurkovi´c et al., 2020) tend to …

Content truncated to 500 words