%D 2025
%I Universitas Pendidikan Indonesia
%L repoupi137818
%O https://scholar.google.com/citations?hl=en&user=V5sB0ioAAAAJ

ID SINTA Dosen Pembimbing:
Rani Megasari: 5992674
Rosa Ariani Sukamto: 5974496
%K Beasiswa, BERT, Chatbot, Fine-Tuning, Natural Language Processing (NLP)

BERT, Chatbot, Fine-Tuning, Natural Language Processing (NLP), Scholarship
%A - Raisyad Jullfikar
%A - Rani Megasari
%A - Rosa Ariani Sukamto
%X Komunikasi efektif antara alumni dan organisasi penting untuk penyampaian informasi salah satunya terkait beasiswa. Di IKA UPI, Sumber Daya Manusia (SDM) terbatas membuat respons lambat dan inefesiensi komunikasi. Penelitian ini membuat chatbot menggunakan model IndoBERT yang merupakan model berbasis BERT (Bidirectional Encoder Representations from Transformers) untuk menjawab terkait beasiswa secara otomatis. Model diadaptasi melalui lima fine-tuning, yaitu Masked Language Modeling (MLM), Next Sentence Prediction (NSP), Intent Classification, Extractive QA (SQuAD), dan Semantic Retrieval. Evaluasi memakai metrik accuracy, recall, f1-score, exact match (EM), top-k accuracy, mean reciprocal rank (MRR), dan latency. Hasil uji menunjukkan kinerja kuat di tugas dasar seperti f1-score 93,33% untuk NSP dan 86,74% untuk intent. Tantangan tampak pada pemahaman konteks lebih dalam seperti pada extractive QA meraih f1-score 43,57% dan retrieval MRR 0,34. Temuan ini menegaskan kelayakan penerapan chatbot menggunakan model IndoBERT untuk pengelolaan komunikasi beasiswa IKA UPI, sekaligus memetakan tantangan dalam tugas-tugas pemahaman konteks yang lebih dalam.

Effective communication between alumni and organizations is crucial for conveying information, including scholarships. At IKA UPI, limited human resources (HR) make communication responses slow and inefficient. This study created a chatbot using the IndoBERT model, a BERT-based model (Bidirectional Encoder Representations from Transformers) to automatically answer questions about scholarships. The model was modified through five fine-tunings: Masked Language Modeling (MLM), Next Sentence Prediction (NSP), Intent Classification, Extractive QA (SQuAD), and Semantic Retrieval. The evaluation used metrics of accuracy, recall, f1-score, exact match (EM), top-k accuracy, mean reciprocal rank (MRR), and latency. Test results showed strong performance in basic tasks, such as an f1-score of 93.33% for NSP and 86.74% for intent. Challenges arose in deeper understanding, such as in extractive QA, which achieved an f1-score of 43.57% and a retrieval MRR of 0.34. This finding of the feasibility of implementing a chatbot using the IndoBERT model for managing IKA UPI scholarship communications, reflects both the challenges in the tasks of understanding the context more deeply.
%T PENGEMBANGAN MODEL INDOBERT DALAM CHATBOT PENGELOLAAN KOMUNIKASI BEASISWA IKATAN ALUMNI UPI