Loekito, Jimmy Agustian (2026) Pengembangan Chatbot Berbasis Indobert Dengan Algoritma Ransformer Yang Ditingkatkan Untuk Mendukung Layanan Konseling Kesehatan Mental Berbasis Alkitab. Doctoral thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 7026231002-Doctoral.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
7026231002-Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy |
Abstract
Penelitian ini mengembangkan model klasifikasi sembilan emosi dan prototipe chatbot konseling kesehatan mental berbasis Alkitab yang mengintegrasikan IndoBERT, representasi Valence-Arousal-Dominance (VAD), intent-emotion mapping, Retrieval-Augmented Generation (RAG), safety referral, dan triage safety layer. Sistem diposisikan sebagai alat bantu awal berbasis teks untuk membantu mengenali kecenderungan emosi, memilih jalur respons, menyediakan dukungan spiritual yang bersumber, dan mengarahkan pengguna kepada bantuan yang lebih sesuai. Chatbot tidak dirancang sebagai alat diagnosis atau pengganti konselor, psikolog, psikiater, dokter, pembimbing rohani, maupun layanan darurat. Penelitian menggunakan pendekatan Design Science Research. Dataset terdiri atas 6.955 teks konseling berbahasa Indonesia yang diberi label cemas, netral, sedih, harapan, takut, tenang, marah, malu, dan senang. Data asli dibagi menjadi 5.564 data pelatihan dan 1.391 data pengujian. Augmentasi melalui translation dan back-translation hanya diterapkan pada data pelatihan sehingga jumlahnya meningkat menjadi 7.859, sedangkan data pengujian tetap tidak berubah. Pengembangan model mencakup baseline IndoBERT, augmentasi, 5-Fold Cross Validation, pembobotan loss heuristik VAD-Focus, Layer-wise Learning Rate Decay, structured head pruning, ensemble, dan ablation study. Baseline tanpa augmentasi mencapai akurasi 0,8327. Eksperimen dengan augmentasi tanpa 5-Fold Cross Validation menghasilkan akurasi 0,8531 dan Macro-F1 0,8478. Pada rangkaian ablation study, baseline menghasilkan akurasi 0,861 dan Macro-F1 0,855. Penambahan VAD-Focus meningkatkan hasil menjadi 0,873 dan 0,866; penambahan structured head pruning menghasilkan 0,888 dan 0,882; sedangkan ensemble beberapa seed mencapai akurasi 0,902 dan Macro-F1 0,895. Hasil terakhir merupakan performa gabungan model, bukan performa satu model tunggal. Nilai VAD digunakan sebagai metadata interpretatif untuk membantu membedakan emosi yang berdekatan dan tidak diperlakukan sebagai skor klinis. Pengujian fungsional menunjukkan bahwa klasifikasi emosi, VAD, intent, RAG, safety referral, dan triage dapat dijalankan dalam satu alur yang dapat ditelusuri. Triage membedakan jalur Psychological Dominant, Somatic Stress- Related, Medical Red Flag, Crisis atau Self-Harm Risk, dan Unclear. Pada jalur Medical Red Flag dan Crisis, aturan keselamatan dapat melakukan override terhadap keluaran emosi atau intent, memprioritaskan rujukan, dan menempatkan RAG serta ayat Alkitab sebagai dukungan sekunder. Sistem juga mencatat emosi, tingkat keyakinan, VAD, intent, triage, stage percakapan, sumber retrieval, dan tindakan keselamatan untuk mendukung audit proses. Evaluasi awal prototipe dilakukan melalui subject matter expert review secara formatif oleh seorang ahli psikologi berkualifikasi doktoral dengan pengalaman 10 tahun dalam konseling keluarga. Evaluasi mencakup telaah terhadap instrumen 18 butir dan pemeriksaan 25 screenshot respons dari lima sistem pada lima prompt identik. Hasil telaah menunjukkan bahwa 16 butir memenuhi kriteria penggunaan instrumen, sedangkan EM4 dan R2 memerlukan penyempurnaan redaksi untuk memperjelas kondisi penerapannya. Temuan evaluasi digunakan untuk mengidentifikasi kebutuhan penyempurnaan pada keamanan respons, empati, kejelasan, rujukan, dukungan spiritual, dan keterlacakan sumber. Evaluasi ini tidak diperlakukan sebagai pengesahan klinis atau konsensus profesi; evaluasi oleh beberapa ahli dan pengujian pengguna yang lebih luas masih diperlukan. Orisinalitas penelitian terdiri atas tiga bagian. Pertama, operasionalisasi klasifikasi sembilan emosi berbasis IndoBERT dengan representasi VAD pada teks konseling berbahasa Indonesia. Kedua, prosedur penyusunan bobot loss heuristic VAD-Focus berdasarkan profil afektif, distribusi kelas, kedekatan emosi, dan pola kesalahan model yang diuji secara iteratif melalui ablation study. Ketiga, metode integratif yang menghubungkan klasifikasi emosi, VAD, intent-emotion mapping, RAG, safety referral, triage, stage percakapan, dan log keputusan ke dalam chatbot yang emotion-aware, safety-aware, source-aware, dan dapat ditelusuri.
===============================================================================================================================
This study develops a nine-emotion classification model and a prototype Bible-based mental health counseling chatbot by integrating IndoBERT, Valence- Arousal-Dominance (VAD) representation, intent-emotion mapping, Retrieval- Augmented Generation (RAG), safety referral, and a triage safety layer. The system is positioned as a text-based early-support tool for identifying emotional tendencies, selecting response pathways, providing source-grounded spiritual support, and directing users to more appropriate assistance. It is not designed as a diagnostic instrument or as a replacement for counselors, psychologists, psychiatrists, physicians, spiritual advisers, or emergency services. The study adopts a Design Science Research approach. The dataset contains 6,955 Indonesian counseling texts labeled as anxious, neutral, sad, hopeful, fearful, calm, angry, ashamed, and happy. The original data were divided into 5,564 training instances and 1,391 test instances. Translation and back- translation augmentation were applied only to the training set, increasing it to 7,859 instances, while the test set remained unchanged. Model development
included an IndoBERT baseline, data augmentation, 5-Fold Cross Validation, heuristic VAD-Focus loss weighting, Layer-wise Learning Rate Decay, structured head pruning, ensemble learning, and an ablation study. The baseline without augmentation achieved an Accuracy of 0.8327. The augmented experiment without 5-Fold Cross Validation achieved an Accuracy of 0.8531 and a Macro-F1 of 0.8478. In the ablation sequence, the baseline achieved an Accuracy of 0.861 and a Macro-F1 of 0.855. Adding VAD-Focus increased the scores to 0.873 and 0.866; adding structured head pruning produced 0.888 and 0.882; and an ensemble across several seeds achieved an Accuracy of 0.902 and a Macro-F1 of 0.895. The final score represents combined model performance rather than a single model. VAD values are used as interpretive metadata to distinguish closely related emotions and are not treated as clinical scores. Functional testing shows that emotion classification, VAD, intent, RAG, safety referral, and triage can operate within a traceable workflow. The triage layer distinguishes Psychological Dominant, Somatic Stress-Related, Medical Red Flag, Crisis or Self-Harm Risk, and Unclear pathways. In Medical Red Flag and Crisis pathways, safety rules can override emotion or intent outputs, prioritize referral, and place RAG and Bible verses as secondary support. The system records emotion labels, confidence, VAD, intent, triage, conversation stage, retrieved sources, and safety actions to support process auditing. The prototype was initially evaluated through a formative subject matter expert review conducted by a doctoral-level psychology expert with 10 years of
experience in family counseling. The evaluation included an 18-item instrument and 25 response screenshots from five systems, all using five identical prompts. The
review found that 16 items met the instrument-use criteria, while EM4 and R2 required wording refinement to clarify their conditions of application. The findings
were used to identify improvements in response safety, empathy, clarity, referral, spiritual support, and source traceability. The evaluation was not treated as a
clinical endorsement or professional consensus; review by multiple experts and broader user testing remain necessary. The study offers three forms of originality. First, it operationalizes a nine- emotion classification based on IndoBERT, using VAD representation for Indonesian counseling texts. Second, it proposes a heuristic VAD-Focus loss- weighting procedure based on affective profiles, class distribution, emotion proximity, and baseline error patterns, evaluated iteratively through an ablation study. Third, it introduces an integrative method that links emotion classification, VAD, intent-emotion mapping, RAG, safety referral, triage, conversation stage, and decision logs into an emotion-aware, safety-aware, source-aware, and traceable chatbot.
| Item Type: | Thesis (Doctoral) |
|---|---|
| Uncontrolled Keywords: | IndoBERT; Valence-Arousal-Dominance; VAD-Focus; emotion classification; Retrieval-Augmented Generation; triage safety layer; Bible-based counseling chatbot IndoBERT; Valence-Arousal-Dominance; VAD-Focus; emotion classification; Retrieval-Augmented Generation; triage safety layer; Bible-based counseling chatbot |
| Subjects: | T Technology > T Technology (General) > T56.8 Project Management T Technology > T Technology (General) > T57.5 Data Processing |
| Divisions: | Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 55003-(S3) PhD Thesis |
| Depositing User: | Jimmy Agustian Loekito |
| Date Deposited: | 28 Jul 2026 08:11 |
| Last Modified: | 28 Jul 2026 08:11 |
| URI: | http://repository.its.ac.id/id/eprint/139838 |
Actions (login required)
 |
View Item |