Adaptasi domain menggunakan Deep Coral pada Arsitektur BiLSTM-Attention untuk klasifikasi emosi teks journaling bahasa Indonesia

Lailatanzila, Nabila (2026) Adaptasi domain menggunakan Deep Coral pada Arsitektur BiLSTM-Attention untuk klasifikasi emosi teks journaling bahasa Indonesia. Sarjana thesis, UIN Sunan Gunung Djati Bandung.

Preview	Text 1_cover.pdf Download (212kB) | Preview
Preview	Text 2_abstrak.pdf Download (276kB) | Preview
Preview	Text 3_skbebasplagiarism.pdf Download (655kB) | Preview
Preview	Text 4_daftarisi.pdf Download (256kB) | Preview
Preview	Text 5_bab1.pdf Download (343kB) | Preview
	Text 6_bab2.pdf Restricted to Registered users only Download (687kB) | Request a copy
	Text 7_bab3.pdf Restricted to Registered users only Download (956kB) | Request a copy
	Text 8_bab4.pdf Restricted to Registered users only Download (1MB) | Request a copy
	Text 9_bab5.pdf Restricted to Registered users only Download (232kB) | Request a copy
	Text 10_daftarpustaka.pdf Restricted to Registered users only Download (261kB) | Request a copy
	Text 11_lampiran.pdf Restricted to Repository staff only Download (333kB) | Request a copy

Abstract

Mental health among adolescents and university students has become a global concern, leading to the increasing use of text-based mood journaling for self-monitoring of emotions. However, Indonesian emotion classification models are generally trained on data from the social media platform X, which consists of short and explicit texts. As a result, these models experience domain shift when applied to journaling texts that are longer, more narrative, and express emotions more implicitly. This study aims to integrate a Bidirectional Long Short-Term Memory with Attention (BiLSTM-Attention) model and Deep CORrelation ALignment (Deep CORAL) to address domain shift through an unsupervised domain adaptation approach from a labeled Source domain (X) to an unlabeled target domain (journaling). The research follows the CRISP-DM methodology and utilizes 6,804 labeled Source-domain samples covering the seven basic emotions, along with 5,368 journaling entries, including 376 labeled samples used as the test set. Deep CORAL is applied to the output of the attention layer by aligning the covariance matrices of the Source and target feature representations using an auto-scaling lambda mechanism. Experimental results show that integrating Deep CORAL improves the model's accuracy from 0.3856 to 0.4521, macro F1-score from 0.3421 to 0.3981, and weighted F1-score from 0.4437 to 0.5032 compared with the model without domain adaptation. Furthermore, the Centroid distance between the Source and target feature spaces is reduced by 95.3%, from 2.5067 to 0.1177, indicating that the learned feature representations become more domain-invariant. Although the performance improvement remains moderate and is not evenly distributed across all emotion classes, the findings demonstrate that the integration of Deep CORAL effectively enhances cross-domain emotion classification for Indonesian text under limited labeled target-domain data.

Actions (login required)

View Item