Enhanced Self-Esteem Classification: Leveraging Data Augmentation and Transformer-Based Sentence Embeddings

Abstract

This study investigates automated self-esteem assessment from self-descriptive text using transformer-based sentence embeddings. Although prior research has explored text, behavioral, and multimodal signals, the combined effects of data augmentation, embedding choice, and classifier complexity in text-only self-esteem classification remain insufficiently understood.
Accordingly, this study aims to systematically evaluate embedding–classifier combinations under both low-resource and augmented data conditions. Textual self-descriptions were collected from 298 undergraduate students at UIN Salatiga and labeled using the Indonesian version of the Rosenberg Self-Esteem Scale, yielding three self-esteem categories. To address data scarcity, a controlled translation-based augmentation pipeline with expert psychological validation was applied exclusively to the training set. Seven multilingual sentence embedding models were paired with eight classification algorithms, and performance was evaluated using macro-averaged metrics, along with training and inference time. Results reveal a two-regime pattern: (1) in limited-data settings, strong embeddings with simple classifiers perform best, (2) whereas in augmented settings, representation quality dominates and classifier choice has a marginal effect. The findings suggest that prioritizing high-quality embeddings and carefully validated data augmentation enables accurate, scalable, and cost-effective text-based self-esteem assessment for real-world psychological applications.