About
Shu-Kai Hsieh is a professor at the Graduate Institute of Linguistics at National Taiwan University, joint-appointed at the Brain and Mind Research Institute, and Vice Dean of the College of Liberal Arts. He received his PhD in Computational Linguistics from the University of Tübingen, Germany. He leads the LOPE Lab, which focuses on computational linguistics, lexical semantics, and Chinese language resources. His research explores the intersection of large language models and linguistic theory, including the concept of 'language compression' addressing LLM-driven linguistic inequality. He also leads Taiwan's team at the International Linguistics Olympiad.
Key Contributions
- Builds Chinese lexical and semantic resources, including Chinese Wordnet, for computational linguistics and language technology
- Developed the fluid lexicon idea, challenging fixed word-boundary assumptions in semantic modeling
- Leads LOPE Lab projects that connect ontologies, multimodal corpora, language processing, and e-humanities
- Frames LLMs as a language-compression problem, drawing attention to linguistic inequality under dominant model pipelines
- Co-authored 讓 AI 好好說話, translating LLM construction and language theory for Traditional Chinese readers
- Connects linguistics, humanities computing, and AI, though his critique asks harder questions than current evaluation benchmarks can easily answer
Videos & Interviews
Papers & Publications
Connections
Emily Bender
In contrastComputational Linguist
Both are computational linguists who were trained to build language technology by hand, and they read the LLM very differently. Bender holds that a system trained on form alone cannot reach meaning. Hsieh calls it genuinely stunning that a model with no rules or features supplied learned language better than the old pipelines did, and says linguistics now has to rethink what language is. They agree about who pays: Bender insists the field name the languages it studies, and Hsieh shows how tokenization charges speakers of non-dominant languages more for every sentence.
humanityisland.nccu.edu.tw · case.ntu.edu.tw · The #BenderRule: On Naming the Languages We Study and Why It Matters (The Gradient, 2019)
Yun-Nung Vivian Chen
KindredProfessor, National Taiwan University
These are two NTU responses to the same imbalance. Hsieh, a linguist, diagnoses it: dominant languages need fewer tokens and have more training data, so a Taiwanese user pays more and gets less. Chen, an engineer, builds against it. Taiwan-LLM, from her lab, is a Traditional Mandarin model aligned to Taiwanese culture, released openly 'to bridge the linguistic divide.' One names the cost and the other pays some of it down.
case.ntu.edu.tw · arxiv.org
