Domain Analysis: Natural Language Processing
Generated July 7, 2026 from 400 real arXiv papers (8 queries x 50 results)
Source Queries
nlp-ai-text-detectionnlp-clinical-ner-indonlp-indo-bert-emotionnlp-llm-alignmentnlp-lora-finetunenlp-low-resource-indonlp-peft-nernlp-rag-healthcare
Total papers scanned: 400
Category Distribution
| Category | Count |
|---|---|
| cs.CL | 233 |
| cs.AI | 162 |
| cs.LG | 117 |
| cs.CV | 55 |
| cs.IR | 39 |
| cs.HC | 19 |
| cs.CY | 19 |
| cs.CR | 14 |
| cs.SD | 12 |
| stat.ML | 10 |
| cs.MM | 7 |
| cs.SE | 7 |
| cs.DC | 6 |
| eess.AS | 6 |
| astro-ph.GA | 6 |
| eess.SP | 4 |
| astro-ph.IM | 4 |
| astro-ph.CO | 4 |
| cs.RO | 4 |
| cs.NI | 3 |
| physics.pop-ph | 3 |
| cs.DB | 3 |
| hep-ph | 3 |
| eess.IV | 3 |
| eess.SY | 2 |
| cs.IT | 2 |
| quant-ph | 2 |
| q-bio.QM | 2 |
| cs.ET | 2 |
| gr-qc | 2 |
| cs.GR | 2 |
| cs.CE | 2 |
| astro-ph | 2 |
| cs.OH | 1 |
| physics.ed-ph | 1 |
| cs.SI | 1 |
| cs.MA | 1 |
| math.OC | 1 |
| cs.AR | 1 |
| cs.PF | 1 |
| cond-mat.str-el | 1 |
| astro-ph.SR | 1 |
| physics.flu-dyn | 1 |
| astro-ph.EP | 1 |
| q-bio.GN | 1 |
| nlin.CG | 1 |
| cs.NE | 1 |
| cond-mat.dis-nn | 1 |
| cs.DL | 1 |
Key Findings & Gaps
Finding 1: Indonesian NLP Data Scarcity Persists
The NusaCrowd initiative (arXiv:2207.10524) confirms that “resources in Indonesian languages, especially the local ones, are extremely scarce and underrepresented.” Despite aggregating datasets, there is still no standardized NLU benchmark for Indonesian. Most benchmarks are translated from English or limited to single tasks like sentiment analysis.
Gap: An Indonesian NLU evaluation suite equivalent to GLUE/SuperGLUE does not exist.
Finding 2: LoRA and PEFT for Indonesian is Untested
LoRA-FAIR (arXiv:2411.14961), FSLoRA (arXiv:2501.19389), and other PEFT methods show strong results on English LLM fine-tuning but no study evaluates them on Indonesian language tasks. The computational benefits of LoRA are especially relevant for Indonesian researchers with limited GPU access.
Gap: No systematic comparison of PEFT methods exists for Indonesian NLP.
Finding 3: Clinical NER for Indonesian is Absent
Query-based NER (arXiv:1908.09138), morpho-syntactic NER (arXiv:1908.10261), and code-mixed NER (arXiv:2206.07318) all target English, Chinese, Bangla, or Hindi. No clinical NER system exists for Indonesian healthcare texts despite Indonesia having 270M+ speakers and growing electronic health records.
Gap: Indonesian clinical NER is entirely unexplored.
Finding 4: Emotion Analysis Lacks Indonesian Resources
While the sentiment analysis survey (arXiv:2005.11882) and aesthetic emotion research (arXiv:2003.07723) are mature for English, Indonesian emotion datasets are limited to basic polarity (positive/negative). Fine-grained emotion recognition (anger, sadness, fear, surprise, etc.) in Indonesian social media is absent.
Gap: No large-scale fine-grained emotion dataset exists for Indonesian.
Finding 5: RAG for Indonesian Healthcare is Unexplored
AR-RAG (arXiv:2506.06962) and related RAG papers demonstrate clinical relevance but operate exclusively on English data. Indonesian healthcare RAG systems face unique challenges: language mismatch, scarcity of digitized clinical knowledge in Bahasa, and lack of evaluation benchmarks.
Gap: No study evaluates RAG for Indonesian clinical decision support.
Finding 6: Indonesian Speaker Diarization Has Room for Improvement
Domain adaptation of pyannote (arXiv:2601.03684) achieves 29.24% DER for conversational Indonesian, which is still far from the <10% DER standard for English. Synthetic data augmentation reduces the gap but cross-domain generalization remains poor.
Gap: Indonesian audio processing still requires significant domain adaptation research.
Recommended Thesis Directions
- IndoEval: An Indonesian NLU Benchmark Suite
- Indonesian Clinical NER with PEFT
- LoRA Fine-Tuning of Indonesian LLMs: A Comparative Study
- RAG-Enabled Indonesian Healthcare Chatbot
- Fine-Grained Emotion Recognition in Indonesian Social Media
- Cross-Lingual Transfer for Indonesian NLP: When Does It Work?
- Federated Fine-Tuning of Indonesian Language Models on Heterogeneous Edge Devices
- Indonesian Medical QA Dataset and Retrieval-Augmented Generation System
Key Papers Referenced
- “Sarang at DEFACTIFY 4.0: Detecting AI-Generated Text Using Noised Data and an Ensemble of DeBERTa Models” –
2502.16857v1 - “Faith in AI can narrow the futures individuals consider” –
2603.28944v2 - “mdok of KInIT: Robustly Fine-tuned LLM for Binary and Multiclass AI-Generated Text Detection” –
2506.01702v2 - “Foundations of GenIR” –
2501.02842v1 - “Human vs. AI: A Novel Benchmark and a Comparative Study on the Detection of Generated Images and the Impact of Prompts” –
2412.09715v1 - “Competing Visions of Ethical AI: A Case Study of OpenAI” –
2601.16513v1 - “NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild” –
2604.11487v1 - “Multi-Hierarchical Feature Detection for Large Language Model Generated Text” –
2509.18862v1 - “HEDGE: Heterogeneous Ensemble for Detection of AI-GEnerated Images in the Wild” –
2604.03555v1 - “SilverSpeak: Evading AI-Generated Text Detectors using Homoglyphs” –
2406.11239v3 - “Detecting AI-Generated Sentences in Human-AI Collaborative Hybrid Texts: Challenges, Strategies, and Insights” –
2403.03506v4 - “TextCohesion: Detecting Text for Arbitrary Shapes” –
1904.12640v2 - “FeatDistill: A Feature Distillation Enhanced Multi-Expert Ensemble Framework for Robust AI-generated Image Detection” –
2603.21939v1 - “TextSleuth: Towards Explainable Tampered Text Detection” –
2412.14816v3 - “Towards The Ultimate Brain: Exploring Scientific Discovery with ChatGPT AI” –
2308.12400v1
Note: S2 API was rate-limited (HTTP 429). Full S2 cross-referencing unavailable. See README.md for details.