diff options
Diffstat (limited to 'java/src/com/android/inputmethod/latin/Suggest.java')
| -rw-r--r-- | java/src/com/android/inputmethod/latin/Suggest.java | 287 |
1 files changed, 229 insertions, 58 deletions
diff --git a/java/src/com/android/inputmethod/latin/Suggest.java b/java/src/com/android/inputmethod/latin/Suggest.java index b03818c1d..0bf0f687a 100644 --- a/java/src/com/android/inputmethod/latin/Suggest.java +++ b/java/src/com/android/inputmethod/latin/Suggest.java @@ -20,16 +20,21 @@ import android.text.TextUtils; import com.android.inputmethod.keyboard.ProximityInfo; import com.android.inputmethod.latin.SuggestedWords.SuggestedWordInfo; +import com.android.inputmethod.latin.common.Constants; +import com.android.inputmethod.latin.common.StringUtils; import com.android.inputmethod.latin.define.DebugFlags; import com.android.inputmethod.latin.settings.SettingsValuesForSuggestion; import com.android.inputmethod.latin.utils.AutoCorrectionUtils; import com.android.inputmethod.latin.utils.BinaryDictionaryUtils; -import com.android.inputmethod.latin.utils.StringUtils; import com.android.inputmethod.latin.utils.SuggestionResults; import java.util.ArrayList; +import java.util.HashMap; import java.util.Locale; +import javax.annotation.Nonnull; +import javax.annotation.Nullable; + /** * This class loads a dictionary and provides a list of suggestions for a given sequence of * characters. This includes corrections and completions. @@ -49,34 +54,55 @@ public final class Suggest { private static final boolean DBG = DebugFlags.DEBUG_ENABLED; private final DictionaryFacilitator mDictionaryFacilitator; + private static final int MAXIMUM_AUTO_CORRECT_LENGTH_FOR_GERMAN = 12; + private static final HashMap<String, Integer> sLanguageToMaximumAutoCorrectionWithSpaceLength = + new HashMap<>(); + static { + // TODO: should we add Finnish here? + // TODO: This should not be hardcoded here but be written in the dictionary header + sLanguageToMaximumAutoCorrectionWithSpaceLength.put(Locale.GERMAN.getLanguage(), + MAXIMUM_AUTO_CORRECT_LENGTH_FOR_GERMAN); + } + private float mAutoCorrectionThreshold; + private float mPlausibilityThreshold; public Suggest(final DictionaryFacilitator dictionaryFacilitator) { mDictionaryFacilitator = dictionaryFacilitator; } - public Locale getLocale() { - return mDictionaryFacilitator.getLocale(); - } - + /** + * Set the normalized-score threshold for a suggestion to be considered strong enough that we + * will auto-correct to this. + * @param threshold the threshold + */ public void setAutoCorrectionThreshold(final float threshold) { mAutoCorrectionThreshold = threshold; } + /** + * Set the normalized-score threshold for what we consider a "plausible" suggestion, in + * the same dimension as the auto-correction threshold. + * @param threshold the threshold + */ + public void setPlausibilityThreshold(final float threshold) { + mPlausibilityThreshold = threshold; + } + public interface OnGetSuggestedWordsCallback { public void onGetSuggestedWords(final SuggestedWords suggestedWords); } public void getSuggestedWords(final WordComposer wordComposer, - final PrevWordsInfo prevWordsInfo, final ProximityInfo proximityInfo, + final NgramContext ngramContext, final ProximityInfo proximityInfo, final SettingsValuesForSuggestion settingsValuesForSuggestion, final boolean isCorrectionEnabled, final int inputStyle, final int sequenceNumber, final OnGetSuggestedWordsCallback callback) { if (wordComposer.isBatchMode()) { - getSuggestedWordsForBatchInput(wordComposer, prevWordsInfo, proximityInfo, + getSuggestedWordsForBatchInput(wordComposer, ngramContext, proximityInfo, settingsValuesForSuggestion, inputStyle, sequenceNumber, callback); } else { - getSuggestedWordsForNonBatchInput(wordComposer, prevWordsInfo, proximityInfo, + getSuggestedWordsForNonBatchInput(wordComposer, ngramContext, proximityInfo, settingsValuesForSuggestion, inputStyle, isCorrectionEnabled, sequenceNumber, callback); } @@ -84,7 +110,7 @@ public final class Suggest { private static ArrayList<SuggestedWordInfo> getTransformedSuggestedWordInfoList( final WordComposer wordComposer, final SuggestionResults results, - final int trailingSingleQuotesCount) { + final int trailingSingleQuotesCount, final Locale defaultLocale) { final boolean shouldMakeSuggestionsAllUpperCase = wordComposer.isAllUpperCase() && !wordComposer.isResumed(); final boolean isOnlyFirstCharCapitalized = @@ -96,16 +122,19 @@ public final class Suggest { || 0 != trailingSingleQuotesCount) { for (int i = 0; i < suggestionsCount; ++i) { final SuggestedWordInfo wordInfo = suggestionsContainer.get(i); + final Locale wordLocale = wordInfo.mSourceDict.mLocale; final SuggestedWordInfo transformedWordInfo = getTransformedSuggestedWordInfo( - wordInfo, results.mLocale, shouldMakeSuggestionsAllUpperCase, - isOnlyFirstCharCapitalized, trailingSingleQuotesCount); + wordInfo, null == wordLocale ? defaultLocale : wordLocale, + shouldMakeSuggestionsAllUpperCase, isOnlyFirstCharCapitalized, + trailingSingleQuotesCount); suggestionsContainer.set(i, transformedWordInfo); } } return suggestionsContainer; } - private static String getWhitelistedWordOrNull(final ArrayList<SuggestedWordInfo> suggestions) { + private static SuggestedWordInfo getWhitelistedWordInfoOrNull( + @Nonnull final ArrayList<SuggestedWordInfo> suggestions) { if (suggestions.isEmpty()) { return null; } @@ -113,75 +142,161 @@ public final class Suggest { if (!firstSuggestedWordInfo.isKindOf(SuggestedWordInfo.KIND_WHITELIST)) { return null; } - return firstSuggestedWordInfo.mWord; + return firstSuggestedWordInfo; } + // Quality constants for dictionary match + // In increasing order of quality + // This source dictionary does not match the typed word. + private static final int QUALITY_NO_MATCH = 0; + // This source dictionary has a null locale, and the preferred locale is also null. + private static final int QUALITY_MATCH_NULL = 1; + // This source dictionary has a non-null locale different from the preferred locale. The + // preferred locale may be null : this is still better than MATCH_NULL. + private static final int QUALITY_MATCH_OTHER_LOCALE = 2; + // This source dictionary matches the preferred locale. + private static final int QUALITY_MATCH_PREFERRED_LOCALE = 3; + // Retrieves suggestions for non-batch input (typing, recorrection, predictions...) // and calls the callback function with the suggestions. private void getSuggestedWordsForNonBatchInput(final WordComposer wordComposer, - final PrevWordsInfo prevWordsInfo, final ProximityInfo proximityInfo, + final NgramContext ngramContext, final ProximityInfo proximityInfo, final SettingsValuesForSuggestion settingsValuesForSuggestion, final int inputStyleIfNotPrediction, final boolean isCorrectionEnabled, final int sequenceNumber, final OnGetSuggestedWordsCallback callback) { - final String typedWord = wordComposer.getTypedWord(); - final int trailingSingleQuotesCount = StringUtils.getTrailingSingleQuotesCount(typedWord); + final String typedWordString = wordComposer.getTypedWord(); + final int trailingSingleQuotesCount = + StringUtils.getTrailingSingleQuotesCount(typedWordString); final String consideredWord = trailingSingleQuotesCount > 0 - ? typedWord.substring(0, typedWord.length() - trailingSingleQuotesCount) - : typedWord; + ? typedWordString.substring(0, typedWordString.length() - trailingSingleQuotesCount) + : typedWordString; final SuggestionResults suggestionResults = mDictionaryFacilitator.getSuggestionResults( - wordComposer, prevWordsInfo, proximityInfo, settingsValuesForSuggestion, - SESSION_ID_TYPING); + wordComposer, ngramContext, proximityInfo.getNativeProximityInfo(), + settingsValuesForSuggestion, SESSION_ID_TYPING); + final Locale mostProbableLocale = mDictionaryFacilitator.getMostProbableLocale(); final ArrayList<SuggestedWordInfo> suggestionsContainer = getTransformedSuggestedWordInfoList(wordComposer, suggestionResults, - trailingSingleQuotesCount); - final boolean didRemoveTypedWord = - SuggestedWordInfo.removeDups(wordComposer.getTypedWord(), suggestionsContainer); + trailingSingleQuotesCount, + // For transforming suggestions that don't come for any dictionary, we + // use the currently most probable locale as it's our best bet. + mostProbableLocale); + + boolean typedWordExistsInAnotherLanguage = false; + int qualityOfFoundSourceDictionary = QUALITY_NO_MATCH; + @Nullable Dictionary sourceDictionaryOfRemovedWord = null; + for (final SuggestedWordInfo info : suggestionsContainer) { + // Search for the best dictionary, defined as the first one with the highest match + // quality we can find. + if (typedWordString.equals(info.mWord)) { + if (mostProbableLocale.equals(info.mSourceDict.mLocale)) { + if (qualityOfFoundSourceDictionary < QUALITY_MATCH_PREFERRED_LOCALE) { + // Use this source if the old match had lower quality than this match + sourceDictionaryOfRemovedWord = info.mSourceDict; + qualityOfFoundSourceDictionary = QUALITY_MATCH_PREFERRED_LOCALE; + } + } else { + final int matchQuality = (null == info.mSourceDict.mLocale) + ? QUALITY_MATCH_NULL : QUALITY_MATCH_OTHER_LOCALE; + if (qualityOfFoundSourceDictionary < matchQuality) { + // Use this source if the old match had lower quality than this match + sourceDictionaryOfRemovedWord = info.mSourceDict; + qualityOfFoundSourceDictionary = matchQuality; + } + typedWordExistsInAnotherLanguage = true; + } + } + } + + SuggestedWordInfo.removeDups(typedWordString, suggestionsContainer); - final String whitelistedWord = getWhitelistedWordOrNull(suggestionsContainer); + final SuggestedWordInfo whitelistedWordInfo = + getWhitelistedWordInfoOrNull(suggestionsContainer); + final String whitelistedWord; + if (null != whitelistedWordInfo && + (mDictionaryFacilitator.isConfidentAboutCurrentLanguageBeing( + whitelistedWordInfo.mSourceDict.mLocale) + || (!typedWordExistsInAnotherLanguage + && !hasPlausibleCandidateInAnyOtherLanguage(suggestionsContainer, + consideredWord, whitelistedWordInfo)))) { + // We'll use the whitelist candidate if we are confident the user is typing in the + // language of the dictionary it's coming from, or if there is no plausible candidate + // coming from another language. + whitelistedWord = whitelistedWordInfo.mWord; + } else { + // If on the contrary we are not confident in the current language and we have + // at least a plausible candidate in any other language, then we don't use this + // whitelist candidate. + whitelistedWord = null; + } final boolean resultsArePredictions = !wordComposer.isComposingWord(); // We allow auto-correction if we have a whitelisted word, or if the word had more than // one char and was not suggested. final boolean allowsToBeAutoCorrected = (null != whitelistedWord) - || (consideredWord.length() > 1 && !didRemoveTypedWord); + || (consideredWord.length() > 1 && (null == sourceDictionaryOfRemovedWord)); final boolean hasAutoCorrection; - // TODO: using isCorrectionEnabled here is not very good. It's probably useless, because - // any attempt to do auto-correction is already shielded with a test for this flag; at the - // same time, it feels wrong that the SuggestedWord object includes information about - // the current settings. It may also be useful to know, when the setting is off, whether - // the word *would* have been auto-corrected. - if (!isCorrectionEnabled || !allowsToBeAutoCorrected || resultsArePredictions - || suggestionResults.isEmpty() || wordComposer.hasDigits() - || wordComposer.isMostlyCaps() || wordComposer.isResumed() - || !mDictionaryFacilitator.hasInitializedMainDictionary() + // If correction is not enabled, we never auto-correct. This is for example for when + // the setting "Auto-correction" is "off": we still suggest, but we don't auto-correct. + if (!isCorrectionEnabled + // If the word does not allow to be auto-corrected, then we don't auto-correct. + || !allowsToBeAutoCorrected + // If we are doing prediction, then we never auto-correct of course + || resultsArePredictions + // If we don't have suggestion results, we can't evaluate the first suggestion + // for auto-correction + || suggestionResults.isEmpty() + // If the word has digits, we never auto-correct because it's likely the word + // was type with a lot of care + || wordComposer.hasDigits() + // If the word is mostly caps, we never auto-correct because this is almost + // certainly intentional (and careful input) + || wordComposer.isMostlyCaps() + // We never auto-correct when suggestions are resumed because it would be unexpected + || wordComposer.isResumed() + // If we don't have a main dictionary, we never want to auto-correct. The reason + // for this is, the user may have a contact whose name happens to match a valid + // word in their language, and it will unexpectedly auto-correct. For example, if + // the user types in English with no dictionary and has a "Will" in their contact + // list, "will" would always auto-correct to "Will" which is unwanted. Hence, no + // main dict => no auto-correct. Also, it would probably get obnoxious quickly. + // TODO: now that we have personalization, we may want to re-evaluate this decision + || !mDictionaryFacilitator.hasAtLeastOneInitializedMainDictionary() + // If the first suggestion is a shortcut we never auto-correct to it, regardless + // of how strong it is (whitelist entries are not KIND_SHORTCUT but KIND_WHITELIST). + // TODO: we may want to have shortcut-only entries auto-correct in the future. || suggestionResults.first().isKindOf(SuggestedWordInfo.KIND_SHORTCUT)) { - // If we don't have a main dictionary, we never want to auto-correct. The reason for - // this is, the user may have a contact whose name happens to match a valid word in - // their language, and it will unexpectedly auto-correct. For example, if the user - // types in English with no dictionary and has a "Will" in their contact list, "will" - // would always auto-correct to "Will" which is unwanted. Hence, no main dict => no - // auto-correct. - // Also, shortcuts should never auto-correct unless they are whitelist entries. - // TODO: we may want to have shortcut-only entries auto-correct in the future. hasAutoCorrection = false; } else { - hasAutoCorrection = AutoCorrectionUtils.suggestionExceedsAutoCorrectionThreshold( - suggestionResults.first(), consideredWord, mAutoCorrectionThreshold); + final SuggestedWordInfo firstSuggestion = suggestionResults.first(); + if (!AutoCorrectionUtils.suggestionExceedsThreshold( + firstSuggestion, consideredWord, mAutoCorrectionThreshold)) { + // Score is too low for autocorrect + hasAutoCorrection = false; + } else { + // We have a high score, so we need to check if this suggestion is in the correct + // form to allow auto-correcting to it in this language. For details of how this + // is determined, see #isAllowedByAutoCorrectionWithSpaceFilter. + // TODO: this should not have its own logic here but be handled by the dictionary. + hasAutoCorrection = isAllowedByAutoCorrectionWithSpaceFilter(firstSuggestion); + } } - if (!TextUtils.isEmpty(typedWord)) { - suggestionsContainer.add(0, new SuggestedWordInfo(typedWord, - SuggestedWordInfo.MAX_SCORE, SuggestedWordInfo.KIND_TYPED, - Dictionary.DICTIONARY_USER_TYPED, - SuggestedWordInfo.NOT_AN_INDEX /* indexOfTouchPointOfSecondWord */, - SuggestedWordInfo.NOT_A_CONFIDENCE /* autoCommitFirstWordConfidence */)); + final SuggestedWordInfo typedWordInfo = new SuggestedWordInfo(typedWordString, + SuggestedWordInfo.MAX_SCORE, SuggestedWordInfo.KIND_TYPED, + null == sourceDictionaryOfRemovedWord ? Dictionary.DICTIONARY_USER_TYPED + : sourceDictionaryOfRemovedWord, + SuggestedWordInfo.NOT_AN_INDEX /* indexOfTouchPointOfSecondWord */, + SuggestedWordInfo.NOT_A_CONFIDENCE /* autoCommitFirstWordConfidence */); + if (!TextUtils.isEmpty(typedWordString)) { + suggestionsContainer.add(0, typedWordInfo); } final ArrayList<SuggestedWordInfo> suggestionsList; if (DBG && !suggestionsContainer.isEmpty()) { - suggestionsList = getSuggestionsInfoListWithDebugInfo(typedWord, suggestionsContainer); + suggestionsList = getSuggestionsInfoListWithDebugInfo(typedWordString, + suggestionsContainer); } else { suggestionsList = suggestionsContainer; } @@ -195,7 +310,7 @@ public final class Suggest { inputStyle = inputStyleIfNotPrediction; } callback.onGetSuggestedWords(new SuggestedWords(suggestionsList, - suggestionResults.mRawSuggestions, + suggestionResults.mRawSuggestions, typedWordInfo, // TODO: this first argument is lying. If this is a whitelisted word which is an // actual word, it says typedWordValid = false, which looks wrong. We should either // rename the attribute or change the value. @@ -204,16 +319,32 @@ public final class Suggest { false /* isObsoleteSuggestions */, inputStyle, sequenceNumber)); } + private boolean hasPlausibleCandidateInAnyOtherLanguage( + final ArrayList<SuggestedWordInfo> suggestionsContainer, final String consideredWord, + final SuggestedWordInfo whitelistedWordInfo) { + for (final SuggestedWordInfo info : suggestionsContainer) { + if (whitelistedWordInfo.mSourceDict.mLocale.equals(info.mSourceDict.mLocale)) { + continue; + } + return AutoCorrectionUtils.suggestionExceedsThreshold(info, consideredWord, + mPlausibilityThreshold); + } + // No candidate in another language + return false; + } + // Retrieves suggestions for the batch input // and calls the callback function with the suggestions. private void getSuggestedWordsForBatchInput(final WordComposer wordComposer, - final PrevWordsInfo prevWordsInfo, final ProximityInfo proximityInfo, + final NgramContext ngramContext, final ProximityInfo proximityInfo, final SettingsValuesForSuggestion settingsValuesForSuggestion, final int inputStyle, final int sequenceNumber, final OnGetSuggestedWordsCallback callback) { final SuggestionResults suggestionResults = mDictionaryFacilitator.getSuggestionResults( - wordComposer, prevWordsInfo, proximityInfo, settingsValuesForSuggestion, - SESSION_ID_GESTURE); + wordComposer, ngramContext, proximityInfo.getNativeProximityInfo(), + settingsValuesForSuggestion, SESSION_ID_GESTURE); + // For transforming words that don't come from a dictionary, because it's our best bet + final Locale defaultLocale = mDictionaryFacilitator.getMostProbableLocale(); final ArrayList<SuggestedWordInfo> suggestionsContainer = new ArrayList<>(suggestionResults); final int suggestionsCount = suggestionsContainer.size(); @@ -222,9 +353,10 @@ public final class Suggest { if (isFirstCharCapitalized || isAllUpperCase) { for (int i = 0; i < suggestionsCount; ++i) { final SuggestedWordInfo wordInfo = suggestionsContainer.get(i); + final Locale wordlocale = wordInfo.mSourceDict.mLocale; final SuggestedWordInfo transformedWordInfo = getTransformedSuggestedWordInfo( - wordInfo, suggestionResults.mLocale, isAllUpperCase, isFirstCharCapitalized, - 0 /* trailingSingleQuotesCount */); + wordInfo, null == wordlocale ? defaultLocale : wordlocale, isAllUpperCase, + isFirstCharCapitalized, 0 /* trailingSingleQuotesCount */); suggestionsContainer.set(i, transformedWordInfo); } } @@ -237,7 +369,7 @@ public final class Suggest { SuggestedWordInfo.removeDups(null /* typedWord */, suggestionsContainer); // For some reason some suggestions with MIN_VALUE are making their way here. - // TODO: Find a more robust way to detect distractors. + // TODO: Find a more robust way to detect distracters. for (int i = suggestionsContainer.size() - 1; i >= 0; --i) { if (suggestionsContainer.get(i).mScore < SUPPRESS_SUGGEST_THRESHOLD) { suggestionsContainer.remove(i); @@ -248,8 +380,12 @@ public final class Suggest { // (typedWordValid=true), not as an "auto correct word" (willAutoCorrect=false). // Note that because this method is never used to get predictions, there is no need to // modify inputType such in getSuggestedWordsForNonBatchInput. + final SuggestedWordInfo pseudoTypedWordInfo = suggestionsContainer.isEmpty() ? null + : suggestionsContainer.get(0); + callback.onGetSuggestedWords(new SuggestedWords(suggestionsContainer, suggestionResults.mRawSuggestions, + pseudoTypedWordInfo, true /* typedWordValid */, false /* willAutoCorrect */, false /* isObsoleteSuggestions */, @@ -283,6 +419,41 @@ public final class Suggest { return suggestionsList; } + /** + * Computes whether this suggestion should be blocked or not in this language + * + * This function implements a filter that avoids auto-correcting to suggestions that contain + * spaces that are above a certain language-dependent character limit. In languages like German + * where it's possible to concatenate many words, it often happens our dictionary does not + * have the longer words. In this case, we offer a lot of unhelpful suggestions that contain + * one or several spaces. Ideally we should understand what the user wants and display useful + * suggestions by improving the dictionary and possibly having some specific logic. Until + * that's possible we should avoid displaying unhelpful suggestions. But it's hard to tell + * whether a suggestion is useful or not. So at least for the time being we block + * auto-correction when the suggestion is long and contains a space, which should avoid the + * worst damage. + * This function is implementing that filter. If the language enforces no such limit, then it + * always returns true. If the suggestion contains no space, it also returns true. Otherwise, + * it checks the length against the language-specific limit. + * + * @param info the suggestion info + * @return whether it's fine to auto-correct to this. + */ + private static boolean isAllowedByAutoCorrectionWithSpaceFilter(final SuggestedWordInfo info) { + final Locale locale = info.mSourceDict.mLocale; + if (null == locale) { + return true; + } + final Integer maximumLengthForThisLanguage = + sLanguageToMaximumAutoCorrectionWithSpaceLength.get(locale.getLanguage()); + if (null == maximumLengthForThisLanguage) { + // This language does not enforce a maximum length to auto-correction + return true; + } + return info.mWord.length() <= maximumLengthForThisLanguage + || -1 == info.mWord.indexOf(Constants.CODE_SPACE); + } + /* package for test */ static SuggestedWordInfo getTransformedSuggestedWordInfo( final SuggestedWordInfo wordInfo, final Locale locale, final boolean isAllUpperCase, final boolean isOnlyFirstCharCapitalized, final int trailingSingleQuotesCount) { |