Research Areas

IISR advances intelligent information services through three major thrusts (2021–2025), integrating LLMs, Retrieval (RAG), Data-centric AI, and Cross-domain Representation Learning.

Thrust 1 — Taiwan-Localized LLMs & Low-Resource Languages

Motivation: Mainstream LLMs often lack Traditional Chinese (Taiwan) cultural grounding and have limited coverage for local languages (e.g., Taiwanese Hokkien).

🔑 Key Contributions

• Chat Vector (ACL 2024) [📄 Paper] [💻 Code]
  A lightweight, economical alignment technique that transfers instruction-following ability using weight-space additivity—reducing dependence on expensive RLHF.
• TWBias (EMNLP 2024) [📄 Paper] [💻 Code]
  The first benchmark for assessing social bias in Traditional Chinese LLMs through a Taiwan cultural lens, supporting trustworthy evaluation for Taiwan-localized models.
• Hokkien NLP & Multimodal Learning
  • Standardizing and bridging multiple writing systems (COLING 2024). [📄 Paper]
  • ATAIGI (NAACL 2025): A multimodal learning app leveraging generative models for low-resource Hokkien education. [📄 Paper] [💻 Code]

Thrust 2 — Biomedical NLP (BioNLP): Data-Centric & Privacy-Aware

Motivation: Biomedical data are often high-noise / low-resource, and clinical text requires strict privacy protection.

🔑 Key Contributions

• BioASQ Biomedical QA (2020–2025) [🌐 Project]
  Achieved six consecutive years as the top team in tasks involving evidence retrieval, semantic search, and evidence-grounded answering.
• Relation Extraction: Survey & New Resource (PEDD) [📄 Paper] [💾 Data]
  Proposed PEDD, a high-quality document-level dataset originally developed for the AI CUP 2019 competition, addressing critical inconsistencies in legacy benchmarks (Briefings in Bioinformatics, 2024).
• Clinical De-identification & Temporal Normalization with LLMs [📄 Paper]
  • Discovery of inverse scaling beyond ~6B parameters without targeted adaptation (npj Digital Medicine, 2025).
  • Efficient solution via PEFT/LoRA, improving privacy protection while preserving clinical utility.
• Data-centric Ensemble Learning [📄 Paper]
  Enhancing biomedical relation extraction through data-centric and preprocessing-robust ensemble learning approach (Database, 2025).

Thrust 3 — Digital Humanities & Historical GIS

Motivation: Humanities data are frequently unstructured, archaic in language, and lack annotations, making standard supervised learning ineffective.

🔑 Key Contributions

• Historical Maps Without Labels (IJGIS 2025) [📄 Paper] [💻 Code]
  Unsupervised domain adaptation (UDA) bridging modern labeled maps and historical unlabeled maps for land-use understanding.
• MingOfficial (EMNLP 2023) [📄 Paper] [💾 Data (Depository)]
  A historical context-aware representation learning framework embedding time, space, and events into career trajectory representations.
• Computational Analysis of Ming Military Power [📄 Paper] [📰 Media (中文)]
  A digital humanities approach to analyzing the roles of Supreme Commanders and Grand Coordinators in Ming Shilu (Digital Scholarship in the Humanities, 2025).
• LLM-based Long-document Analysis
  Designing grounded RAG workflows with context fields (metadata + text) and human-in-the-loop benchmarks to improve traceability in historical sources.

🚀 Selected Deployment Context