What Does Hong Kong's TSA Chinese Exam Actually Test? Why AI Tutoring Tools Keep Getting It Wrong

When building an AI Chinese tutoring product for Hong Kong students aged 4 to 16, the team at LUMI Chinese Institute fell into a classic edtech trap: they bulk-generated a large volume of reading passages and accompanying questions, assuming that quantity equaled product. It wasn't until they spent an entire week reviewing every single passage in the question bank against Hong Kong's real assessment frameworks — TSA (Territory-wide System Assessment), HKAT (Hong Kong Academic Test), and the DSE Chinese curriculum — that they realized almost nothing they had assumed held true.

This article documents what they found. For any team building AI products for the Hong Kong education market, these lessons are worth taking seriously.

Hong Kong Exams Test Things You Probably Don't Expect

TSA reading papers follow a strict word-count ladder by grade level:

Level	Passage Length	Benchmark
P1–P2	30–100 words	Pre-TSA stage; picture-prompted sentence writing
P3–P4	120–200 words	TSA P3 → P6 transition
P5–P6	230–320 words	TSA P6 → HKAT Secondary placement

The passages the team initially generated were consistently 25% above or below these target ranges. Sounds trivial? It isn't. A P4 student practicing on a 160-word passage is essentially training on P3 material — and will be underprepared when they sit a real exam.

The key takeaway is simple: word count isn't a formatting detail — it is the difficulty specification itself. This is the foundational logic that most AI question-generation tools miss. They treat length as a freely adjustable variable. In Hong Kong's assessment system, length is the difficulty.

TSA (Territory-wide System Assessment) is a government-administered assessment for P3, P6, and S3 students designed to monitor overall learning progress. Results are not recorded in individual student profiles, but schools use the data to adjust teaching strategies. HKAT (Hong Kong Academic Test) is a standardized test used for Secondary One placement and directly affects which secondary school a student attends. DSE (Hong Kong Diploma of Secondary Education) is the public examination for senior secondary students — equivalent to university entrance exams in other systems — and its Chinese curriculum features a reading assessment framework that breaks reading ability into quantifiable, tiered indicators. These three systems are interconnected, forming a continuous assessment pathway from primary through secondary school. This means practice materials designed for any given grade must align precisely with that grade's position along the full pathway.

Every Answer Must Be Directly Locatable in the Text

One of the most important rules in Hong Kong's top-school exam design is this: every question must have an answer that can be directly located in the passage. There are no inference questions where the answer lives in the teacher's head.

During the review, the team found that many questions flagged as "difficult" were actually just vague — unanswerable based on the passage alone. These questions don't penalize weak students; they penalize students who read carefully, because those students get stuck when they can't find textual support. Every such question was removed.

What this rule does is redefine "difficulty" as the challenge of locating and distinguishing information — not the challenge of guessing the question-setter's intent. For AI generation, this means every answer needs a traceable source within the text.

Distractors Come from Within the Passage, Not from a Vocabulary List

This one is genuinely counterintuitive. In Hong Kong-style cloze questions, the wrong answer choices should come from words that appear elsewhere in the same passage — testing whether students can distinguish between similar words in context, not whether they've memorized vocabulary lists.

Generalist AI question-generation tools almost always fail here. They tend to pull distractors from frequency-ranked vocabulary dictionaries, producing questions where the wrong choices are often more plausible than the correct answer — completely distorting what's being assessed.

The team wrote a review script that automatically flags any distractor not found in the source passage, routing it for human review. This is a telling example: AI's "default behavior" is systematically misaligned with real assessment standards and must be explicitly corrected with rules.

Cloze tests are a classic language assessment format that asks students to fill in missing words within a passage. Their measurement validity depends heavily on how distractors are designed. When distractors come from an external vocabulary list, the question effectively becomes a vocabulary breadth test. When they come from within the text, the question tests the student's ability to distinguish word meaning and logical relationships in a specific context — which is the core skill of reading comprehension. Hong Kong's assessment framework deliberately chooses the latter, turning the cloze into a measure of context sensitivity rather than vocabulary size. Large language models default to optimizing for plausibility when generating answer choices, not discriminability, because they're trained to predict reasonable words — not to deliberately generate options that are easy to rule out in a specific context. This is the fundamental misalignment between general-purpose generation tools and purpose-built assessment design.

Question Order Follows Text Order

The answer to Question 1 should appear in the early part of the passage; the answer to Question 5 should appear near the end. This "sequential principle" reduces cognitive load and mirrors the structure of real exam papers.

Most quiz apps default to randomized question order — which is precisely training students to work against the exam format. It may seem to increase "randomness" or reduce cheating, but it breaks the rhythm of working through a paper the way students actually will in the exam room. For a test-prep tool, structural alignment with the real exam matters more than any flashy feature.

The Ability Ladder: Recall → Explain → Reorganize → Extend → Evaluate

The DSE defines reading ability across five tiers. A well-designed question set should climb this ladder within a single passage: a vocabulary question, a retrieval question, an "explain why the author…" question, and then an open-ended response question.

Eighty percent of the team's original question bank consisted of pure recall-and-retrieval questions. The rebuilt bank enforces mandatory question-type mixing, covering the full gradient from lower-order to higher-order thinking. This means AI can't just "generate questions" — it needs to understand where each question sits within the ability framework. This is the dividing line between content generation and assessment design.

This tiered ability framework maps closely onto Bloom's Taxonomy from educational theory: from the lowest tier of memorization and recognition, stepping up through comprehension, application, and analysis, to the highest tiers of evaluation and creation. Hong Kong's DSE translates this framework into operational specifications for question design, allowing exam writers to deliberately cover different cognitive levels within a single question set and avoid question homogeneity. For AI question-generation systems, producing recall-level retrieval questions ("What does the passage say about…?") is technically trivial. But generating higher-order questions that require students to synthesize information from multiple parts of the text, infer the author's intent, or make evaluative judgments requires the model to understand a question's tier within the ability framework and control information density and open-endedness accordingly. This is the critical leap from AI content generation "replicating exam formats" to "understanding assessment logic."

What They Actually Built

Based on this review, the team did three things — all of which, they say, can be directly adapted by others:

1. A written specification document. Per-grade word count standards, paragraph structure rules (P4 and above require at least 3 meaningful paragraphs), and a question-type coverage matrix (each grade requires coverage of six core question types: punctuation, sentence reordering, sentence expansion, picture-prompted writing, multiple choice, and short answer).
2. A regression review script. Run on every content change: Is the word count compliant? Is the cloze answer present in the passage? Are the multiple-choice indices valid? Are quotations consistent? Does the paragraph count meet the minimum? Any single failure blocks deployment. Think of it as CI for content.
3. The will to delete. The team removed or rewrote 56 passages that failed to meet the ability-tier requirements. To deliver the right two-thirds, they were willing to cut one-third of the question bank — and they say it was the best decision they made.

What This Means for Edtech Founders

One line from the team is worth remembering: Hong Kong parents are extraordinarily sophisticated consumers of assessment design — and they will actually read your sample papers.

This points to a widely underestimated truth: what looks like a "content" product surface is actually a measure of your compliance with assessment specifications. In an era when AI can easily generate text at scale, what's scarce is no longer content volume — it's deep understanding of real exam standards and the discipline to enforce them.

For any team using AI to enter a vertical education market, this case offers a clear lesson: master the assessment framework before you talk about scaling generation. Use automated scripts to encode your specifications as enforceable quality thresholds — don't rely on a model's default output. Quality isn't generated. It's reviewed and deleted into existence.