Abstract
Social media platforms have become popular places for online self-disclosure. The reuse of such rich social media data drives researchers to study mental disorders online. Most studies that investigated the digital diagnosis of bipolar disorder (BD) combine dual features: Social media and linguistics, to train patient diagnostic models. However, when social media features are not available, numerous scenarios can impede the performances of these models. This study focuses on the textual contexts of BD patients who self-report on social media, and proposes a linguistic BD recognition approach (LiBRA) based on the syntactic patterns of word usage from the 3-month period prior to diagnosis. The performance indicates the F1 scores of the proposed LiBRA outperform several state-of-the-art linguistic baselines, including those using TF-IDF, LIWC and pre-trained language models. The analyses show that LiBRA can offer explainability and linguistic representations that help distinguish emotional behavior. The contributions of the present study are: (1) The features are contextualized, domain-agnostic, and purely linguistic. (2) The performance of BD recognition is improved by gender-enriched linguistic pattern features, which are constructed with gender differences in language usage.