Domain Analysis: Tirana-Aligned NLP Research

Generated July 8, 2026 from 1,000 real arXiv papers (20 queries x 50 results) Queries aligned to Tirana Noor Fatyanosa, S.Kom., M.Kom., Ph.D.’s research areas Based on supervised theses at Universitas Brawijaya repository.ub.ac.id

Source Queries

• tirana-bert-text-class
• tirana-bert2bert-chatbot
• tirana-cross-lingual-nlu
• tirana-cyber-log-cluster
• tirana-deep-clustering
• tirana-explainable-bert
• tirana-federated-bert
• tirana-hdbscan-anomaly
• tirana-indobert-sentiment
• tirana-indonesian-emotion
• tirana-limited-data-nlp
• tirana-log-anomaly-transformer
• tirana-multilingual-bert
• tirana-parameter-efficient
• tirana-roberta-emotion
• tirana-survey-bert-clinical
• tirana-survey-cyber-dl
• tirana-survey-emotion-dl
• tirana-survey-indo-llm
• tirana-survey-low-resource

Total papers scanned: 1000

UB Thesis Connections

The following UB Informatics theses supervised by/co-supervised by Tirana guided query selection:

1. Farid Muzaki (2026) — RoBERTa for multi-class emotion detection in English text
2. Brigitta Nilapaksi (2026) — IndoBERT sentiment analysis on Twitter/X for “Magang Berdampak 2025”
3. Naufal Akmal (2026) — HDBSCAN-based cyber attack investigation from web server logs
4. Rizky Dwi Purnomo (2024) — Preprocessing vs fine-tuning effects on BERT text classification
5. Femi Novia Lina (2024) — BERT2BERT for Indonesian chatbot dialogue response generation

Key Gaps Identified

1. No Indonesian Emotion Benchmark

Tirana’s RoBERTa thesis (Farid Muzaki, 2026) targets English emotion. Despite IndoBERT (Nilapaksi, 2026) for sentiment, no Indonesian multi-class emotion benchmark exists. RoBERTa/IndoBERT have never been systematically compared for Indonesian emotion tasks.

2. Preprocessing Studies are Shallow

Rizky Dwi Purnomo’s 2024 UB thesis compares preprocessing methods for BERT classification, but the study is limited to one model and one dataset. No comprehensive preprocessing ablation study exists across Indonesian BERT variants.

3. BERT2BERT Data Scarcity

Femi Novia Lina’s 2024 thesis uses BERT2BERT on limited Indonesian data. Data augmentation, back-translation, and self-training for low-resource Indonesian dialogue are unexplored.

4. HDBSCAN Parameter Sensitivity

Naufal Akmal’s 2026 thesis uses HDBSCAN for cyber logs but doesn’t study parameter sensitivity, deep embedding pre-clustering, or explainable clustering — all open gaps.

5. No Indonesian PEFT Study

All UB theses use full fine-tuning. LoRA, Adapters, and Prefix Tuning are untested for Indonesian BERT tasks despite being a major trend (arXiv 2501.19389, 2411.14961).

6. Explainability Gap

None of the Tirana-supervised theses incorporate explainable AI (attention visualization, SHAP, LIME) for their classification models. This is a clear gap given XAI requirements in academic publishing.

7. Federated / Privacy-Preserving Gap

Indonesian NLP research has zero studies on federated fine-tuning, differential privacy, or on-device deployment for BERT models. This is especially relevant for sensitive domains (emotion, sentiment).

Recommended Thesis Directions

1. Indonesian Emotion Detection: RoBERTa vs. IndoBERT vs. mBERT
2. PEFT Methods for Indonesian BERT: A Systematic Comparison
3. HDBSCAN with Deep Embedding Pre-Training for Cyber Log Clustering
4. Data Augmentation for Low-Resource Indonesian Dialogue Generation
5. Explainable Indonesian Sentiment Analysis with Attention Visualization
6. Federated BERT Fine-Tuning for Privacy-Preserving Indonesian NLP
7. Active Learning for Indonesian Text Classification with Limited Labels
8. Cross-Lingual Emotion Transfer from English RoBERTa to Indonesian
9. Continual Fine-Tuning of Indonesian BERT for Evolving Classification Tasks
10. Synthetic Data Generation with LLMs for Indonesian Low-Resource NLP

Key Papers Referenced

1. “How to Fine-Tune BERT for Text Classification?” – 1905.05583v3
2. “A Hybrid Classical-Quantum Fine Tuned BERT for Text Classification” – 2511.17677v1
3. “Cost-Aware Model Selection for Text Classification: Multi-Objective Trade-offs Between Fine-Tuned Encoders and LLM Prompting in Production” – 2602.06370v1
4. “Revisiting Few-sample BERT Fine-tuning” – 2006.05987v3
5. “Stage-wise Fine-tuning for Graph-to-Text Generation” – 2105.08021v2
6. “BERT Fine-tuning For Arabic Text Summarization” – 2004.14135v1
7. “Differentially Private Fine-tuning of Language Models” – 2110.06500v2
8. “PatentBERT: Patent Classification with Fine-Tuning a pre-trained BERT Model” – 1906.02124v2
9. “Fine-tuning with Very Large Dropout” – 2403.00946v3
10. “Partial Is Better Than All: Revisiting Fine-tuning Strategy for Few-shot Learning” – 2102.03983v1
11. “Legged Robots that Keep on Learning: Fine-Tuning Locomotion Policies in the Real World” – 2110.05457v1
12. “Improving BERT Fine-tuning with Embedding Normalization” – 1911.03918v2
13. “Evaluating Fine-Tuning Efficiency of Human-Inspired Learning Strategies in Medical Question Answering” – 2408.07888v2
14. “Sensi-BERT: Towards Sensitivity Driven Fine-Tuning for Parameter-Efficient BERT” – 2307.11764v2
15. “Improving BERT Fine-Tuning via Self-Ensemble and Self-Distillation” – 2002.10345v1