Latent Probabilistic Topic Discovery for Text Documents
Incorporating Segment Structure and Word Order.
A Game, a Cliff, and a Question About Reading.
Long before "technical difficulty" was a term in a peer-reviewed paper, it was a plumber jumping between platforms. Growing up playing Super Mario, I noticed something every player learns instinctively: not all jumps are equal. A gap between two nearby, low ledges is trivial. But make the next platform farther away, or put it higher up, and suddenly the jump demands real precision — more speed, a longer run-up, better timing. Distance and height, combined, are what make a jump hard.
Years later, as a PhD student thinking about how people read, that same picture came back to me. Reading a document isn't so different from crossing that terrain. Some sentences are short hops between familiar, everyday words — easy footing. Others force the reader to leap across a wide gap between unfamiliar concepts, or up onto a ledge of dense, technical vocabulary that demands far more cognitive effort to land on safely. The farther apart two ideas are conceptually, and the "higher" the vocabulary sits above everyday language, the harder the reader has to work to get from one to the next — exactly like a harder jump in the game.
That single idea — that a document could be modelled as a terrain, where technical difficulty has both a "height" and a "distance," and that ranking text by difficulty means measuring how hard the jumps across it are — became my first CIKM paper, and the seed that grew into my entire PhD research on latent, structure-aware topic models. The interactive terrain below is my attempt to make that mental picture playable.
Farther cliffs, higher cliffs, harder reading.
Each ledge below is a sentence. Click a ledge — or use the controls — to make the explorer jump to it. The height and distance of the jump map directly onto how much cognitive load that sentence demands from a reader.
Pioneering Word Order Modeling.
My thesis challenged the traditional "Bag-of-Words" assumption that dominated NLP for decades. By introducing non-parametric segment structure and word order dependency, I developed models that could capture the actual narrative flow of human language, leading to a Best Research Awards nomination across all of Hong Kong.
Foundation for Deep Learning.
The concepts explored in my doctoral work—specifically how segments and sequences define meaning—foreshadowed the architectural shift toward modern Transformers and Large Language Models (LLMs). My algorithms proved that structure is as vital as statistics in high-stakes information retrieval.
An Unsupervised Ranking Method Based on a Technical Difficulty Terrain
Shoaib Jameel, Wai Lam, Ching-man Au Yeung, and Sheaujiun Chyan. CIKM 2011, pp. 1989–1992, Glasgow, Scotland. Winner: Best Paper Award, Beijing–Hong Kong International Doctoral Forum 2011.
DOI: 10.1145/2063576.2063872- ▹ Non-parametric N-Gram Topic Models: Developed a framework that automatically discovers the appropriate length and complexity of phrases without pre-defined constraints.
- ▹ Segment Structure Integration: Pioneered the use of document internal boundaries to improve the interpretability and coherence of latent topics.
- ▹ Sequential Discourse Cohesion: Created ranking models that understand how conceptual difficulty transitions across a text, a vital component for domain-specific readability.