erabo/spiegel-keyman
1
0
Fork 0
mirror of https://github.com/keymanapp/keyman.git synced 2026-08-11 03:15:33 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 20

Merge pull request #4206 from keymanapp/feat/common/models/basic-backspace-modeling

Browse source 
feat(common/models): naive backspace workaround, naive multi-char Transform mitigation
This commit is contained in:
Joshua Horton 2020-12-23 11:44:52 +07:00 • committed by GitHub
commit 7e2f5bf3a5
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 34 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
common/predictive-text/worker/correction/classical-calculation.ts
View file

@ -52,7 +52,8 @@ namespace correction {
* Otherwise, this object represents a heuristic that _may_ overestimate the true edit distance. Note that it will
* never underestimate.
*/
diagonalWidth: number = 1;
diagonalWidth: number = 2; // TODO: Ideally, should start at 1... but we'll start at 2 for now
// as a naive workaround for multi-char transform limitations.
// The sequence of characters input so far.
inputSequence: TInput[] = [];

33
common/predictive-text/worker/correction/context-tracker.ts
View file

@ -358,7 +358,38 @@ namespace correction {
token.transformDistributions = [transformDistribution];
state.pushTail(token);
} else {
state.updateTail(transformDistribution, tokenizedContext[0]);
// Consider backspace entry for this case?
let hasDistribution = transformDistribution && Array.isArray(transformDistribution);
let primaryInput = hasDistribution ? transformDistribution[0].sample : null;
if(primaryInput && primaryInput.insert == "" && primaryInput.deleteLeft > 0) {
// It's a backspace transform; time for special handling!
//
// For now, with 14.0, we simply compress all remaining Transforms for the token into
// multiple single-char transforms. Probabalistically modeling BKSP is quite complex,
// so we simplify by assuming everything remaining after a BKSP is 'true' and 'intended' text.
//
// Note that we cannot just use a single, monolithic transform at this point b/c
// of our current edit-distance optimization strategy; diagonalization is currently...
// not very compatible with that.
let backspacedTokenContext = tokenizedContext[0].split('').map(function(char) {
let transform: Transform = {
insert: char,
deleteLeft: 0,
id: primaryInput.id // Not exactly optimal for every transform to have the same ID,
// but is actually accurate here.
};
return [{sample: transform, p: 1.0}];
});
state.tokens.pop();
let compactedToken = new TrackedContextToken();
compactedToken.raw = tokenizedContext[0];
compactedToken.transformDistributions = backspacedTokenContext;
state.pushTail(compactedToken);
} else {
state.updateTail(transformDistribution, tokenizedContext[0]);
}
}
}
return state;