<![CDATA[Personality prediction from social media text has become a critical topic in computational social science, as online posts frequently reflect individual behavioral and psychological characteristics. However, predicting personality from Indonesian social media material remains challenging due to informal language, slang, abbreviations, noisy expressions, and implicit personality-related cues. This study proposed a new OCEAN personality prediction framework that integrates pre-trained language models with an OCEAN-aware Query Expansion Ranking feature learning mechanism. Unlike conventional approaches that rely mainly on contextual embeddings, the proposed framework introduced trait-specific class-discriminative lexical features to strengthen personality-related representation. The prediction task was structured as five separate binary classification problems, with each personality attribute divided into High and Low groups. The experiments were conducted in two stages. First, numerous pre-trained language models, including IndoBERT, IndoBERTweet, Indonesian RoBERTa, and multilingual BERT, were fine-tuned and compared to identify the most suitable contextual representation model. The best baseline model, multilingual BERT, achieved an average accuracy of 75.48% and an average F1-score of 73.58%. Second, multilingual BERT was integrated with class-discriminative lexical features generated by the proposed OCEAN-aware Query Expansion Ranking mechanism. The trait-specific configuration improved the average accuracy to 78.76% and the average F1-score to 76.48%. These results demonstrated that integrating contextual semantic representations with OCEAN-aware lexical feature learning enhanced personality prediction from Indonesian social media material while also offering a clearer representation of trait-relevant language signals.]]>
