For polyphones and homophones, candidate order dominates first-hit rate

What it is

Chinese maps many syllables onto many characters; lists for shi, yi, zhi can run long. What users usually care about is not “is it in the list” but whether slot one is the intended word—Space, the 1 key, or the leftmost soft cell is bound to the first candidate. The larger the homophone set, the larger the gap between rank 1 and rank 5: the former is zero extra work; the latter is reading, a number key, or a page turn. Order here is almost the same thing as efficiency.

Why it happens

Homophone-set size is a property of the syllable's information, not of the UI. On high-ambiguity syllables, a few function words occupy the frequency head and content words crowd behind. Context models matter most here: without context, shi almost always yields 是 first; after 教师, 室 can rise. Top-1 is therefore extremely sensitive to ranking quality and not very sensitive to candidate-window height—the window can grow, Space still takes the first. On touch the first cell is also the largest, so mistaps concentrate there: a wrong rank both requires another tap and makes the big button easy to hit by mistake.

Studying it

Stratify syllables by ambiguity (low such as nü, high such as yi) and report top-1, top-3, and page-turn rate per band. Materials must be contextual sentences, not isolated syllables. Shuffling the list except slot one separates “the word exists” from “which rank it occupies.” Eye tracking shows whether high-ambiguity syllables force reading the whole row. Do not extrapolate from English autocomplete top-1; English candidates are usually not a homophone set.

Where it stops holding

Whole-sentence decode postpones ambiguity, lowering the importance of per-syllable first hits—until the user confirms word by word, when the problem returns. Speech homophones happen in recognition, with no candidate order to tune. Digit- and English-only fields have no such set. Japanese kana conversion has a similar kanji-homophone problem; Korean hangul barely does. Chinese top-1 numbers do not transfer.

Applying it

• Write the objective as domain top-1, not “the lexicon contains this word.”
• On high-ambiguity syllables with context, be willing to take function words out of slot one.
• Do not make the touch first-candidate button much larger than the second, or a wrong rank stacks a mistap penalty.
• Verify by taking the 50 most frequent syllables from product logs and scoring Space-commit accuracy; for the worst syllables, check whether context was ignored by the fusion.

Related

• Same group: C6.26.1 Candidate ranking fuses word frequency, a contextual language model, and personal history · C6.26.2 Learning from choices raises a word later, but can also cement typo candidates · C6.26.4 Live re-ranking can move the target as the user is about to click, causing a mis-pick
• Adjacent: C6.12 IME candidates · C6.08 T9 versus full keyboards
• Search: homophone · top-1 accuracy · polyphone