epic: autocorrect 🚂

android/KMAPro/kMAPro/src/main/java/com/tavultesoft/kmapro/LanguageSettingsActivity.java

@@ -108,11 +108,11 @@ public void onCreate(Bundle savedInstanceState) {
     RadioGroup radioGroup = (RadioGroup) findViewById(R.id.suggestion_radio_group);
     radioGroup.clearCheck();
 
-    // Auto-correct disabled for Keyman 18.0 #12767
     int[] RadioButtonArray = {
       R.id.suggestion_radio_0,
       R.id.suggestion_radio_1,
-      R.id.suggestion_radio_2};
+      R.id.suggestion_radio_2,
+      R.id.suggestion_radio_3};
     RadioButton radioButton = (RadioButton)radioGroup.findViewById(RadioButtonArray[maySuggest]);
     radioButton.setChecked(true);
 

android/KMEA/app/src/main/res/layout/language_settings_list_layout.xml

@@ -90,13 +90,12 @@
                 android:layout_gravity="center_vertical"
                 android:text="@string/suggestions_radio_2" />
 
-            <!-- Auto-correct disabled for Keyman 18.0 #12767 -->
-            <!--com.google.android.material.radiobutton.MaterialRadioButton
+            <com.google.android.material.radiobutton.MaterialRadioButton
                 android:id="@+id/suggestion_radio_3"
                 android:layout_width="wrap_content"
                 android:layout_height="wrap_content"
                 android:layout_gravity="center_vertical"
-                android:text="@string/suggestions_radio_3" /-->
+                android:text="@string/suggestions_radio_3" />
 
         </RadioGroup>
 

web/README.md

@@ -1,5 +1,5 @@
 # Keyman Engine for Web
-The Original Code is (C) SIL International
+The Original Code is (C) SIL Global
 
 ## Prerequisites
 See [build configuration](../docs/build/index.md) for details on how to

web/src/engine/main/src/headless/languageProcessor.ts

@@ -21,7 +21,7 @@ export class LanguageProcessor extends EventEmitter<LanguageProcessorEventMap> {
 
   private _mayPredict: boolean = true;
   private _mayCorrect: boolean = true;
-  private _mayAutoCorrect: boolean = false; // initialized to false - #12767
+  private _mayAutoCorrect: boolean = true;
 
   private _state: StateChangeEnum = 'inactive';
 

web/src/engine/predictive-text/wordbreakers/src/main/default/index.ts

@@ -38,7 +38,7 @@ export interface DefaultWordBreakerOptions {
  * @see http://unicode.org/reports/tr29/#Word_Boundaries
  * @see https://github.com/eddieantonio/unicode-default-word-boundary/tree/v12.0.0
  */
-export default function default_(text: string, options?: DefaultWordBreakerOptions): LexicalModelTypes.Span[] {
+ function default_(text: string, options?: DefaultWordBreakerOptions): LexicalModelTypes.Span[] {
   let boundaries = findBoundaries(text, options);
   if (boundaries.length == 0) {
     return [];
@@ -64,6 +64,16 @@ export default function default_(text: string, options?: DefaultWordBreakerOptio
   return spans;
 }
 
+// Exposes `searchForProperty` for external use while associating it with this wordbreaker.
+const def = Object.assign(default_, { 
+  /**
+   * This method returns enum values corresponding to the character type as perceived by the wordbreaking algorithm.
+   */
+  searchForProperty: searchForProperty
+});
+
+export default def;
+
 /**
  * A span that does not cut out the substring until it absolutely has to!
  */

web/src/engine/predictive-text/wordbreakers/src/main/index.ts

@@ -1,5 +1,6 @@
 import placeholder from "./placeholder.js";
 import ascii from "./ascii.js";
 import default_ from "./default/index.js";
+import { WordBreakProperty } from "./default/data.inc.js";
 
-export { placeholder, ascii, default_ as default, default_ as defaultWordbreaker };
+export { placeholder, ascii, default_ as default, default_ as defaultWordbreaker, WordBreakProperty };

web/src/engine/predictive-text/worker-thread/src/main/predict-helpers.ts

@@ -6,6 +6,9 @@ import { ContextTracker, TrackedContextState } from './correction/context-tracke
 import { ExecutionTimer } from './correction/execution-timer.js';
 import ModelCompositor from './model-compositor.js';
 import { LexicalModelTypes } from '@keymanapp/common-types';
+import { defaultWordbreaker, WordBreakProperty } from '@keymanapp/models-wordbreakers';
+const searchForProperty = defaultWordbreaker.searchForProperty;
+
 import Context = LexicalModelTypes.Context;
 import Distribution = LexicalModelTypes.Distribution;
 import Keep = LexicalModelTypes.Keep;
@@ -615,6 +618,35 @@ export function processSimilarity(
   });
 }
 
+/**
+ * This function may be used to prevent auto-selection/auto-correct from applying in
+ * unexpected ways.  For example, when typing numbers in English, we don't expect
+ * '5' to auto-correct to '5th' just because there are no pure-number entries in
+ * the lexicon rooted on '5'.
+ * @param correction 
+ * @returns 
+ */
+export function correctionValidForAutoSelect(correction: string) {
+  let chars = [...correction];
+
+  // If the _correction_ - the actual, existing text - does not include any letters,
+  // then predictions built upon it should not be considered valid for auto-correction.
+  for(let c of chars) {
+    // Found even one letter?  We'll consider it valid.
+    switch(searchForProperty(c.codePointAt(0))) {
+      case WordBreakProperty.ALetter:
+      case WordBreakProperty.Hebrew_Letter:
+      case WordBreakProperty.Katakana:
+        return true;
+      default:
+    }
+  }
+
+  // Only reached when the correction has nothing that passes as a letter in-context.
+  // (MidLet and MidNumLet only count when there are adjacent letters.)
+  return false;
+}
+
 export function predictionAutoSelect(suggestionDistribution: CorrectionPredictionTuple[]) {
   if(suggestionDistribution.length == 0) {
     return;
@@ -633,6 +665,11 @@ export function predictionAutoSelect(suggestionDistribution: CorrectionPredictio
   suggestionDistribution = suggestionDistribution.slice(1);
 
   if(suggestionDistribution.length == 1) {
+    // Prevent auto-acceptance when the root doesn't meet validation criteria.
+    if(!correctionValidForAutoSelect(suggestionDistribution[0].correction.sample)) {
+      return;
+    }
+
     // Mark for auto-acceptance; there are no alternatives.
     suggestionDistribution[0].prediction.sample.autoAccept = true;
     return;
@@ -676,6 +713,10 @@ export function predictionAutoSelect(suggestionDistribution: CorrectionPredictio
     return;
   }
 
+  if(!correctionValidForAutoSelect(bestSuggestion.correction.sample)) {
+    return;
+  }
+
   // compare correction-cost aspects?  We disable if the base correction is lower than best,
   // but should we do other comparisons too?