On the Usefulness of Personality Traits in Opinion-oriented Tasks

Proceedings of Recent Advances in Natural Language Processing, pages 547–556 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_062 547 On the Usefulness of Personality Traits in Opinion-oriented Tasks Marjan Hosseinia University of Houston {ma.hosseinia}∗ Eduard Dragut Temple University edragut@temple.edu Dainis Boumber University of Houston {dboumber}† Arjun Mukherjee University of Houston arjun@cs.uh.edu Abstract We use a deep bidirectional transformer to ex- tract the Myers-Briggs personality type from user-generated data in a multi-label and multi- class classification setting. Our dataset is large and made up of three available personality datasets of various social media platforms in- cluding Reddit, Twitter, and Personality Cafe forum. We induce personality embeddings from our transformer-based model and investi- gate if they can be used for downstream text classification tasks. Experimental evidence shows that personality embeddings are effec- tive in three classification tasks including au- thorship verification, stance, and hyperpartisan detection. We also provide novel and inter- pretable analysis for the third task: hyperpar- tisan news classification. 1 Introduction The vocabulary we use in everyday language is a rich source of information about our beliefs, thoughts, and personalities (Pennebaker et al., 2015). Many efforts in text analysis provide com- pelling evidence that our everyday language car- ries psychological cues (Gottschalk and Gleser, 1979; Stone et al., 1966; Weintraub, 1989; Pen- nebaker et al., 2015). With this study, we seek to determine the personality of a given text’s au- thor as defined by the Myers-Briggs Type Indica- tors or MBTI (Myers and Myers, 1995). Myers- Briggs uses four binary dimensions to classify peo- ple (Introvert–Extrovert, Intuitive–Sensing, Think- ing–Feeling, Judging–Perceiving), which gives 16 different types, such as INTJ and ENTJ. This work uncovers novel insights into the personality space of authors from their online writings. The personality signal carries the fingerprint of the individual’s psyche and, even though noisy, can be useful (as shown in this work) for a variety of ∗@gmail.com †@uh.edu downstream NLP tasks, such as authorship verifi- cation, stance, and hyperpartisan detection. Person- ality prediction does not only benefit commercial applications and psychology but also is advanta- geous in health care. Recent works link personality types and social media behavior with depression and posttraumatic stress disorder (Preot¸iuc-Pietro et al., 2015). This is significant because it opens new avenues for prevention care as certain per- sonality types can anticipate mental illness and schizophrenia. (Mitchell et al., 2015). This problem poses a non-trivial challenge be- cause a good solution must be capable of capturing the complexity and the depth of the human psyche as expressed through text. Anything short of that will result in task-specific pattern-matching. It fol- lows that the main technical difficulty presented by the task at hand is the discrepancy between the cor- pora concerning their distributions, which results in the domain shift. This is where our work becomes relevant as it aims to bridge the gap by transfer learning and universal language understanding. The problem is also challenging because the hu- man psyche is complex in its nature. The labels are fuzzy as the label distribution changes from population …

Content truncated to 500 words