<![CDATA[Over the past few years, researches on Big Five personality prediction model focusing on Indonesian tweets has mostly faced data scarcity and imbalance problem. This study aims to build Big Five personality prediction models for the low-resource language setting using Indonesian social media texts in platform X, and conduct a comparative performance analysis using deep learning model (hybrid BiLSTM-BiGRU) and transferred learning models (IndoBERT and IndoXLNet) with the optimization using various data augmentation and feature expansion techniques to overcome data scarcity and imbalance. We carried out data collection of 11746 tweets from hybrid sources (BFI questionnaire and keyword-based crawling), preprocessing using tokenization and stemming, extraction using Word2Vec and FastText, data augmentation (using SYN-REP, BACK-TRANS, RAND-INS, and SMOTE variants), Top-N feature expansion (with N = 1, 5, and 10), hyperparameter tuning using Optuna (for learning rate, batch size, dropout rate, and weight decay), and evaluating the model's accuracy, precision, recall, and F1-score. The result shows that IndoBERTweet+BACK-TRANS+Top-10 achieved 84% accuracy and an F1-score of 81.5%, surpassing IndoXLNet+BACKTRANS+Top-5 (82%) and LSTM+GRU+FastText+KMeansSMOTE+Top-5 (78%), with the BACK-TRANS technique improves overall F1 result by 13%. These findings bolster computational psychology informatics for low-resource NLP focusing on Indonesian language, as well as aiding mental health screening with 18% better generalization in imbalanced data, and enabling scalable trait inference in human resource and mental health applications.]]>
Copyrights © 2026
