Identifikasi Kepribadian Big Five menggunakan Arsitektur Hierarchical Transformer dengan Mekanisme Label Attention pada User-Generated Content

Jovanka, Tamara (2026) Identifikasi Kepribadian Big Five menggunakan Arsitektur Hierarchical Transformer dengan Mekanisme Label Attention pada User-Generated Content. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221213-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Penelitian prediksi kepribadian berbasis teks masih menghadapi beberapa tantangan, yaitu keterbatasan model transformer standar dalam menangkap konteks naratif yang tersebar lintas banyak dokumen, serta dominasi studi pada model MBTI dibandingkan Big Five yang lebih umum digunakan dalam psikologi. Untuk mengatasi tantangan tersebut, penelitian ini mengadaptasi arsitektur Hierarchical Transformer with Label Attention (HTLA) yang sebelumnya dikembangkan untuk klasifikasi MBTI agar dapat memprediksi kelima trait Big Five (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) pada data teks media sosial. Dataset yang digunakan adalah PANDORA, korpus komentar Reddit dengan label kepribadian hasil pelaporan mandiri, di mana 1.568 pengguna dengan label Big Five lengkap diekstrak dari total 10.295 pengguna dan masing-masing direpresentasikan oleh maksimum 50 komentar yang ditokenisasi menggunakan RoBERTa. Arsitektur HTLA terdiri dari Word-Level Encoder berbasis RoBERTa, Document-Level Transformer Encoder, mekanisme Label Attention untuk representasi spesifik per trait, dan lima Per-Trait Classifier independen. Pelatihan menerapkan strategi two-phase training, differential learning rate, serta penanganan class imbalance melalui pos_weight dinamis. Model mencapai rata-rata Balanced Accuracy sebesar 0,5961, dengan performa terbaik pada trait Neuroticism (0,6229) dan Agreeableness (0,6137). Studi ablasi mengonfirmasi bahwa Document-Level Encoder merupakan komponen paling krusial, yang penghapusannya memicu penurunan sebesar 0,0869 poin dan minority class collapse, membuktikan bahwa sinyal kepribadian pada teks media sosial bersifat naratif dan kumulatif sehingga membutuhkan pemodelan konteks hierarkis.
==================================================================================================================================
Text-based personality prediction research still faces several challenges: the limitations of standard transformer models in capturing narrative context scattered across multiple documents, and the predominance of studies on the MBTI model compared to the Big Five, which is more widely used in psychology. To address these challenges, this study adapts the Hierarchical Transformer with Label Attention (HTLA) architecture, previously developed for MBTI classification, to predict the five Big Five personality traits (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) from social media text data. The dataset used is PANDORA, a corpus of Reddit comments with self-reported personality labels, from which 1,568 users with complete Big Five labels were extracted from a total of 10,295 users, each represented by a maximum of 50 comments tokenized using RoBERTa. The HTLA architecture consists of a RoBERTa-based Word-Level Encoder, a Document-Level Transformer Encoder, a Label Attention mechanism for trait-specific representations, and five independent Per-Trait Classifiers. Training employs a two-phase strategy, differential learning rates, and class imbalance handling through dynamic pos_weight. The model achieves an average Balanced Accuracy of 0.5961, with the best performance on Neuroticism (0.6229) and Agreeableness (0.6137). The ablation study confirms that the Document-Level Encoder is the most crucial component, whose removal triggers a 0.0869-point drop and minority class collapse, proving that personality signals in social media texts are narrative and cumulative, thus requiring hierarchical context modeling.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Big Five Personality Traits, Hierarchical Transformer, Label Attention, Natural Language Processing, Klasifikasi Kepribadian.
Subjects:	T Technology > T Technology (General) T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Tamara Jovanka
Date Deposited:	30 Jul 2026 02:13
Last Modified:	30 Jul 2026 02:13
URI:	http://repository.its.ac.id/id/eprint/138189

Actions (login required)

View Item