[Paper Review] Leveraging Linguistic Characteristics for Bipolar Disorder Recognition with Gender Differences

Yen-Hao Huang, Yi‐Hsin Chen+5 more|arXiv (Cornell University)|Jul 17, 2019

Mental Health via Writing18 references4 citations

TL;DR

This study proposes a gender-enriched, purely linguistic approach for bipolar disorder (BD) recognition using syntactic patterns derived from Twitter posts, leveraging graph-based pattern construction and attention mechanisms. It achieves over 91% F1 score, outperforming TF-IDF, LIWC, and pre-trained models like BERT and ELMo, while capturing gender-specific linguistic differences in first-person pronouns, tenses, and emotional expressions.

ABSTRACT

Most previous studies on automatic recognition model for bipolar disorder (BD) were based on both social media and linguistic features. The present study investigates the possibility of adopting only language-based features, namely the syntax and morpheme collocation. We also examine the effect of gender on the results considering gender has long been recognized as an important modulating factor for mental disorders, yet it received little attention in previous linguistic models. The present study collects Twitter posts 3 months prior to the self-disclosure by 349 BD users (231 female, 118 male). We construct a set of syntactic patterns in terms of the word usage based on graph pattern construction and pattern attention mechanism. The factors examined are gender differences, syntactic patterns, and bipolar recognition performance. The performance indicates our F1 scores reach over 91% and outperform several baselines, including those using TF-IDF, LIWC and pre-trained language models (ELMO and BERT). 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.

Motivation & Objective

• To develop a purely linguistic, context-aware method for bipolar disorder recognition without relying on social behavior or pre-defined dictionaries.
• To investigate how gender differences in language use affect BD recognition performance.
• To construct gender-specific syntactic patterns that improve model interpretability and clinical relevance.
• To reduce reliance on keyword-based features that may misclassify domain experts as patients.
• To provide transferable, domain-agnostic linguistic patterns for clinical use in mental health screening.

Proposed method

• Constructs syntactic patterns using graph pattern construction based on word co-occurrence and dependency structures in user posts.
• Applies a pattern attention mechanism to learn salient linguistic patterns from sequences of text.
• Integrates gender-specific linguistic patterns by training separate models on male and female BD user data.
• Uses a gender-enriched feature representation that combines male- and female-tuned syntactic patterns.
• Employs a data-driven approach to learn n-gram and skip-gram style patterns without relying on keyword dictionaries.
• Trains and evaluates models on 349 BD users' Twitter data collected 3 months before self-disclosure, with gender labels manually verified.

Experimental results

Research questions

• RQ1 Can purely linguistic syntactic patterns improve bipolar disorder recognition performance compared to traditional methods?
• RQ2 How do gender differences in language use affect the performance and interpretability of BD recognition models?
• RQ3 Do gender-specific syntactic patterns enhance recognition accuracy and capture clinically relevant linguistic markers?
• RQ4 Can graph-based pattern learning outperform keyword-based (e.g., TF-IDF, LIWC) and pre-trained language models (e.g., BERT, ELMo) in BD detection?
• RQ5 Are first-person pronouns, absolute words, and tense usage significant markers of BD, and do they differ by gender?

Key findings

• The proposed model achieves an F1 score above 91%, significantly outperforming TF-IDF, LIWC, ELMo, and BERT baselines.
• The use of gender-enriched syntactic patterns improves recognition performance by capturing distinct linguistic markers between male and female BD users.
• Female BD users show a higher frequency of present-tense constructions (e.g., "I am"), while male BD users use more past-tense forms (e.g., "I was"), with statistical significance (p < 0.001 for present tense).
• The model captures clinically relevant linguistic features such as increased use of first-person pronouns (1.7× more common) and absolute words like "always" and "never".
• The syntactic patterns are more interpretable than deep learning embeddings, as they highlight contextually meaningful n-gram and skip-gram patterns.
• The model reduces the risk of misclassifying mental health professionals as patients by avoiding keyword-based reliance, instead focusing on structural linguistic patterns.