Proposal: context-aware learning from accepted and reverted suggestions #415

Problem

LeanType already has several personalization mechanisms:

• UserHistoryDictionary
• personal dictionary suggestions
• session word boost
• bigram/trigram history
• configurable personal-suggestion boost

However, these mechanisms mainly learn that a word was used. They do not reliably learn that a particular candidate was the correct choice for a specific mistyped input and context.

For users who mostly enter text by selecting suggestions, this can produce a recurring pattern:

1. The intended word is present in the suggestion strip, but ranked second or third.
2. The user repeatedly selects it manually.
3. The same less suitable candidate remains ranked first on future occurrences.
4. A suggestion reverted immediately with Backspace may still remain highly ranked.

This appears related to the personalization balance discussed in #382, but the proposed change would not increase the global weight of personal words.

Proposed direction

Keep the existing dictionary candidate generation unchanged and add a small, bounded feedback layer for suggestion ranking.

Possible feedback signals:

• Selecting suggestion Add This Feature please #2 or Both offline onnx models crash #3: weak positive feedback for that candidate for the same typed form.
• Repeatedly selecting the same candidate: gradually increase its contextual preference.
• Reverting an autocorrection immediately with Backspace: strong negative feedback for that correction.
• Replacing a selected suggestion immediately: negative feedback for the removed candidate and positive feedback for the replacement.
• Matching previous-word context: allow a larger adjustment than a typed-form-only match.

The feedback should only reorder candidates already produced by the existing dictionaries. It should not generate new words or replace the current decoder.

A simplified model could be:

finalScore = originalScore + contextualFeedbackBoost - rejectionPenalty

The adjustment should be capped so that learned feedback cannot override a clearly better dictionary/spatial match.

Shadow-mode evaluation

Before changing visible suggestions, the candidate ranker could first run in an optional local shadow mode:

baseline order: A, B, C
candidate order: B, A, C
user selected: B
outcome: helped

This would make the feature measurable rather than opaque:

• helped: the candidate ranker would move the selected word higher;
• hurt: the baseline ranked the selected word higher;
• neutral: no meaningful rank change.

Useful local metrics would include Top-1/Top-3 accuracy, mean reciprocal rank, immediate revert rate, helped vs hurt, and suggestion latency. Visible reranking could remain disabled until enough local evidence shows a meaningful positive margin.

Integration constraints

I would like to keep this compatible with the existing architecture:

• candidate generation remains in DictionaryFacilitatorImpl;
• final suggestion construction remains in Suggest;
• reuse existing user-history mechanisms where practical;
• no network access or telemetry;
• no learning in password fields or incognito mode;
• bounded local storage with decay or eviction;
• no complete-message storage;
• no noticeable latency in the suggestion hot path;
• optional/experimental behavior if it may not be wanted by everyone.

Non-goals

This proposal does not include:

• replacing the AOSP/native dictionary decoder;
• adding an LLM or large neural model;
• introducing a new unified dictionary format;
• changing gesture typing;
• globally increasing the priority of all personal words.

Suggested implementation sequence

PR 1: rejection feedback

Teach LeanType not to repeatedly prefer a suggestion that the user immediately reverted.

This would be intentionally small and independently useful:

• detect immediate revert of an autocorrection or selected suggestion;
• record a bounded negative association;
• apply a capped penalty when the same correction is proposed again;
• add focused regression tests;
• do not learn in sensitive fields or incognito mode.

Candidate generation would remain unchanged; this PR would only apply a bounded penalty to an already generated candidate.

PR 2: accepted-suggestion feedback

After the first behavior is validated:

• learn when the user repeatedly selects a lower-ranked candidate;
• include previous-word context;
• cap and decay old feedback;
• add ranking tests with synthetic candidates and contexts.

Touch-position adaptation could be considered separately only if evidence shows that the intended word is frequently absent from the candidate list rather than merely ranked poorly.

Questions before implementation

1. Would a small feedback layer on top of the existing suggestion scores fit the direction of LeanType?
2. Should rejection feedback be stored through UserHistoryDictionary, or would a separate bounded local store be preferable?
3. Is immediate Backspace after autocorrection a sufficiently reliable rejection signal for the first PR?
4. Would you prefer this behind the existing personalized-suggestions setting or a separate experimental toggle?

I can prepare the first narrowly scoped PR after agreeing on the storage and integration point.