Research

Sungkyunkwan professor's new book shows LLMs flatten word meaning

Sungkyunkwan University professor Lee Han-jung's book, due next month from De Gruyter Brill, finds LLMs like ChatGPT reduce word meanings to frequent patterns and fail to capture full context-dependent sense.

Sungkyunkwan University professor explores limits of AI language models in new book
Sungkyunkwan University professor explores limits of AI language models in new bookArturo_Anez / pixabay
By Rebecca Stone4 min read

Why it matters

  • Lee Han-jung's book 'Polysemy in Semantics with Contextualized Language Models' is published next month by De Gruyter Brill.
  • The book joins the 'Language and Computers' series, a peer-reviewed digital linguistics series launched in 1988.
  • Lee presented the findings in a keynote at the 2026 Seoul International Conference on Linguistics in August.
  • Follow-up research is scheduled for the 2027 LSA Annual Meeting in San Francisco in January.
  • The research found state-of-the-art LLMs reduce word meanings to generalized representations based on frequent patterns.

Sungkyunkwan University professor Lee Han-jung will publish an English-language book next month arguing that large language models, including ChatGPT, reduce word meanings to generalized representations built from frequently occurring patterns — and fail to capture the full range of human meaning.

The book, titled "Polysemy in Semantics with Contextualized Language Models," will be published by De Gruyter Brill, the university said. It presents an AI-based methodology for studying linguistic meaning and offers a scientific analysis of the fundamental limitations of large language models.

The stakes are straightforward. Lexical semantics has traditionally sat in the humanities; computational models now shape how millions of people read, write and search. A rigorous account of where those models fall short on context-dependent meaning matters both to linguists testing their theories against machines and to developers trying to understand what LLMs actually encode when they process language.

What does the book actually contain?

Lee, a professor in the Department of English Language and Literature at Sungkyunkwan University, integrates linguistic frameworks with artificial neural networks — the systems that allow computers to process language by representing words and expressions as numerical data. The combination produces a new approach to understanding how word meanings shift depending on context.

The book joins "Language and Computers," an academic series launched in 1988 that publishes outstanding works in digital linguistics selected through a rigorous peer-review process. The university described the placement as significant: the book bridges lexical semantics, traditionally rooted in the humanities, with computational models powered by advanced AI technologies.

Alongside the methodology, the book explores ways to advance the future development of LLMs. It treats the models' weaknesses not as curiosities but as measurable properties that future systems could address.

"Through this book, I sought to explore both the potential and limitations of AI language models and shed light on how AI affects human language and thought," Lee said. "I will continue to pursue interdisciplinary research that bridges linguistics, AI and cognitive science, expanding the horizons of English studies and the humanities."

Why is context-dependent meaning hard for machines?

Humans resolve ambiguity in words effortlessly. Korean speakers, the university noted, can easily determine from context whether the word "nun" refers to an eye or snow — the two are homophones in Korean. Speakers do this without conscious effort.

For AI language models, the same task is not settled. Accurately identifying context-dependent meanings and resolving lexical ambiguity remain major challenges, according to the research. The gap between human and machine performance on such cases forms the empirical core of the book.

Lee's research team approached the problem from three directions at once:

  • Large-scale textual data analysis, examining how words behave across corpora.
  • Psychological experiments on how humans comprehend words.
  • In-depth investigations into the internal semantic representations of AI systems.

The combined study produced what the university described as scientific evidence that even widely used, state-of-the-art AI language models fail to fully capture the richness and complexity of human language. Instead, the models tend to reduce word meanings to generalized representations based on frequently occurring patterns — a compression that limits their ability to capture the full range of meanings a word can carry.

That finding lands amid a broader debate over what LLMs understand versus what they reproduce. By grounding the claim in psychological experiments and inspection of the models' internal representations, rather than in surface-level output alone, the work gives the debate empirical footing.

Where has the research been presented?

Lee presented the findings in a keynote lecture to an international audience of linguists at the 2026 Seoul International Conference on Linguistics, held in Seoul in August. A keynote slot at an international conference signals that the work has already cleared peer scrutiny beyond the book series itself.

More results are coming. The university said follow-up research will be presented at the 2027 Annual Meeting of the Linguistic Society of America — one of the world's leading linguistics conferences — in San Francisco in January. That timeline gives researchers and AI developers a fixed date for the next installment of the findings.

Why does this matter beyond linguistics?

The book's central claim — that pattern frequency substitutes for meaning in current models — speaks directly to how LLMs are built and evaluated. If models generalize from frequent patterns, they will perform unevenly on rare senses, ambiguous words and context-shifted usage, exactly the cases where human readers excel. Lee's methodology offers a way to measure that unevenness rather than assume it.

The publication also marks a shift in who studies these systems. A professor of English language and literature, working through a digital-linguistics series founded in 1988, is now publishing peer-reviewed evidence on the internal semantics of AI models with an academic house like De Gruyter Brill. The humanities are not just commenting on LLMs; they are instrumenting them.

Lee framed her agenda in interdisciplinary terms. Her stated plan is to continue research bridging linguistics, AI and cognitive science — work that treats language models as objects of scientific study, with testable limits, rather than as finished products.

The January 2027 session in San Francisco will show whether the follow-up research narrows the gap it documents: whether models can be pushed past frequent-pattern generalization toward the context-sensitive meaning that human readers deploy without thinking.

Source: feed.koreatimes.co.kr

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

802 articles

Related articles

  1. Study finds six habits that make AI writing exhausting to read
  2. AI English Tutors Could Close Korea's Speaking Gap, Veteran Teacher Says
  3. Google's Language Explorer Maps 7,000 Languages, With Caveats
  4. A Startup Wants to Train AI World Models on Your Gameplay
  5. Graphite finds 13,000 phrases that flag AI writing, including 'this matters'

« Previous articleNext article »