Unveiling Hidden Themes of Gender Specific Social Anxiety through Linguistic Exploratory Analysis Using LLM and Topic Modeling Techniques

Abstract

This article aims to unveil hidden themes related to social anxiety in gender-specific-manner. For this purpose, dataset of over 12,000 Reddit posts related to social anxiety were used. Traditional preprocessing steps including lemmatization were applied to clean the data. Initially, Llama 3 was employed for zero-shot gender classification using an appropriate prompt to label the posts by gender. The zero-shot classification was then evaluated against human judgment and baseline algorithms. Top2Vec was fine-tuned to identify prevalent linguistic traits and topics within the female and male groups. Various embedding methods were experimented with, and coherence scores were used as evaluation metric for searching best embedding for topic modeling with high coherence score. Doc2vec gives the best coherence score. The optimal settings generated topic vectors with relevant keywords for each gender, highlighting key social anxiety themes. A method was devised to identify the most similar and dissimilar topics for both genders. The analysis revealed significant similarities in male social anxiety posts with female posts in themes of social interaction, mental health, daily activities, dating, and professional communication. Conversely, the least similar topics in female social anxiety posts compared to male posts centered around issues like appearance, facial expressions, school interactions, and strategies for overcoming social anxiety. This analysis underscores the diverse contexts of social anxiety experiences across genders.