BERT and ULMFiT Ensemble for Personality Prediction from Indonesian Social Media Text
Noptovius Halimawan, Derwin Suhartono, Aryo Pradipta Gema, Rezki Yunanda · 2022
Predicting personality is a growing topic in the field of natural language processing. The study of personality prediction has been proven to benefit the development of recommender systems and automated personality assessments by previous studies. Additionally, the widespread usage of social media in Indonesia such as Twitter has served as a potential source of data for developing such models. Existing personality prediction models has explored the implementation of both traditional machine learning models and deep learning models, with the latter proven to perform better with more data. Despite so, there is not much development of deep learning-based personality prediction models in the domain of Indonesian text. We propose a deep learning-based ensemble model based on Universal Language Model Fine Tuning (ULMFiT) and Bidirectional Encoder Representations from Transformers (BERT), the state-of-the-art pre-trained models for text classification. The model is trained with Indonesian tweets labelled with Five Factor Model (FFM) personality traits along with the users' metadata. The model is tested with five predetermined scenarios to prove the robustness of its performance. The final evaluation of the model proves that it has successfully achieved optimal performance of 76.41 % macro-average accuracy and 75.55% macro-average F-1 score, exceeding the state-of-the-art BERT and ULMFiT