Learning from choices raises a word later, but can also cement typo candidates

What it is

An IME treats every commit as supervision: pick this word, and the next time the same pinyin arrives it moves forward. Proper names and catchphrases reach slot one quickly. The same mechanism learns mis-picks: Space committing a homophone typo, or a finger slipping onto the second candidate, writes that wrong word into the user lexicon, where it becomes ever easier to pick again. Personalization treats history as truth; error mixed into history cements.

Why it happens

User lexicons usually update (pinyin, word, count). A mis-pick and an intentional pick both add one; there is no “was this a slip” label. A self-excitation loop follows: the wrong word ranks higher → Space is more likely to commit it again → the count rises. Cloud sync ships one device's mis-pick to another. Some systems decay or down-weight when the user immediately backspaces and picks another word, but backspace may also be read as editing earlier text, not as a rejection of that commit. Unlike static frequency, this pollution is personal; someone else's IME will not show that wrong word on top.

Studying it

On a resettable user lexicon, inject N wrong commits, then measure how many it takes for that wrong word to enter top-1, and how many correct picks it takes to push it back. Compare with and without “undo commit down-weights.” Diaries can show whether a wrong word still owns slot one days later. Do not only report that personalization raised hit rate—report cemented error items separately.

Where it stops holding

With user learning off or in a private session, cementing does not happen, and cold-start names also slow down. Enterprise lexicons pushed by an admin keep personal mis-picks out, but the admin's error cements for everyone. If speech-recognition errors write into the same user lexicon, the keyboard is not the only pollution source. On a child's or a shared device, one person's mis-pick becomes the next person's first candidate.

Applying it

• Offer a way to inspect and delete user entries; a wrong word must be searchable and removable in one step.
• A backspace-and-re-pick within a short window after commit should lower that word's weight, not increment it again.
• Shared devices should disable or isolate the user lexicon by default.
• Verify by committing a homophone typo with Space five times, restarting the IME, and checking whether it is now first; after deleting the user entry it should leave. If deletion is impossible, the learning is not recoverable.

Related

• Same group: C6.26.1 Candidate ranking fuses word frequency, a contextual language model, and personal history · C6.26.3 For polyphones and homophones, candidate order dominates first-hit rate · 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 · C6.11 Autocorrect
• Search: user dictionary · personalization drift · IME learning