Depression remains a severe global mental health crisis, affecting over 264 million people worldwide. Recent research has demonstrated that linguistic patterns in social media posts can serve as reliable markers for depression detection. Titla-Tlatelpa et al. [1] introduced a profile-based sentiment-aware approach that specialized classifiers according to users’ demographic traits and incorporated sentiment polarity through a novel Bag of Polarities (BoP) representation. However, their work relied on simple lexicon-based methods for profiling, which limited its potential. This paper addresses these limitations by proposing an enhanced profiling system that leverages state-of-the-art author profiling models and introduces fine-grained age categorization. Our approach replaces lexicon-based profiling with transformer-based models for more accurate gender and age prediction, and introduces a multi-class age classification system (teen, young adult, adult, middle-aged, senior) instead of binary young/senior categorization. Experimental results on Reddit and Twitter datasets show significant improvements, with up to 9.9% enhancement in F1-score on the Reddit dataset compared to the original approach, while maintaining interpretability. Our work demonstrates that accurate demographic profiling is crucial for the profile-based paradigm in mental health assessment.
Depression remains a severe global mental health crisis, affecting over 264 million people worldwide. Recent research has demonstrated that linguistic patterns in social media posts can serve as reliable markers for depression detection. Titla-Tlatelpa et al. [1] introduced a profile-based sentiment-aware approach that specialized classifiers according to users’ demographic traits and incorporated sentiment polarity through a novel Bag of Polarities (BoP) representation. However, their work relied on simple lexicon-based methods for profiling, which limited its potential. This paper addresses these limitations by proposing an enhanced profiling system that leverages state-of-the-art author profiling models and introduces fine-grained age categorization. Our approach replaces lexicon-based profiling with transformer-based models for more accurate gender and age prediction, and introduces a multi-class age classification system (teen, young adult, adult, middle-aged, senior) instead of binary young/senior categorization. Experimental results on Reddit and Twitter datasets show significant improvements, with up to 9.9% enhancement in F1-score on the Reddit dataset compared to the original approach, while maintaining interpretability. Our work demonstrates that accurate demographic profiling is crucial for the profile-based paradigm in mental health assessment.