Enable count based dynamic ngram language model for v403.

Bug: 14425059

Change-Id: Icc15e14cfd77d37cd75f75318fd0fa36f9ca7a5b

6 files changed, 128 insertions, 89 deletions

diff --git a/native/jni/src/suggest/core/dictionary/dictionary.cpp b/native/jni/src/suggest/core/dictionary/dictionary.cpp
index bfe17cc4c..6a5df9d95 100644
--- a/native/jni/src/suggest/core/dictionary/dictionary.cpp
+++ b/native/jni/src/suggest/core/dictionary/dictionary.cpp
@@ -81,6 +81,9 @@ void Dictionary::NgramListenerForPrediction::onVisitEntry(const int ngramProbabi
     }
     const WordAttributes wordAttributes = mDictStructurePolicy->getWordAttributesInContext(
             mPrevWordIds, targetWordId, nullptr /* multiBigramMap */);
+    if (wordAttributes.getProbability() == NOT_A_PROBABILITY) {
+        return;
+    }
     mSuggestionResults->addPrediction(targetWordCodePoints, codePointCount,
             wordAttributes.getProbability());
 }
diff --git a/native/jni/src/suggest/core/dictionary/ngram_listener.h b/native/jni/src/suggest/core/dictionary/ngram_listener.h
index e9b3c1aaf..2eb5e9fd1 100644
--- a/native/jni/src/suggest/core/dictionary/ngram_listener.h
+++ b/native/jni/src/suggest/core/dictionary/ngram_listener.h
@@ -26,6 +26,8 @@ namespace latinime {
  */
 class NgramListener {
  public:
+    // ngramProbability is always 0 for v403 decaying dictionary.
+    // TODO: Remove ngramProbability.
     virtual void onVisitEntry(const int ngramProbability, const int targetWordId) = 0;
     virtual ~NgramListener() {};
 
diff --git a/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content.cpp b/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content.cpp
index 05a3a6356..31b1ea696 100644
--- a/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content.cpp
+++ b/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content.cpp
@@ -19,11 +19,11 @@
 #include <algorithm>
 #include <cstring>
 
-#include "suggest/policyimpl/dictionary/utils/forgetting_curve_utils.h"
+#include "suggest/policyimpl/dictionary/structure/v4/content/dynamic_language_model_probability_utils.h"
+#include "suggest/policyimpl/dictionary/utils/probability_utils.h"
 
 namespace latinime {
 
-const int LanguageModelDictContent::DUMMY_PROBABILITY_FOR_VALID_WORDS = 1;
 const int LanguageModelDictContent::TRIE_MAP_BUFFER_INDEX = 0;
 const int LanguageModelDictContent::GLOBAL_COUNTERS_BUFFER_INDEX = 1;
 
@@ -39,7 +39,8 @@ bool LanguageModelDictContent::runGC(
 }
 
 const WordAttributes LanguageModelDictContent::getWordAttributes(const WordIdArrayView prevWordIds,
-        const int wordId, const HeaderPolicy *const headerPolicy) const {
+        const int wordId, const bool mustMatchAllPrevWords,
+        const HeaderPolicy *const headerPolicy) const {
     int bitmapEntryIndices[MAX_PREV_WORD_COUNT_FOR_N_GRAM + 1];
     bitmapEntryIndices[0] = mTrieMap.getRootBitmapEntryIndex();
     int maxPrevWordCount = 0;
@@ -53,7 +54,15 @@ const WordAttributes LanguageModelDictContent::getWordAttributes(const WordIdArr
         bitmapEntryIndices[i + 1] = nextBitmapEntryIndex;
     }
 
+    const ProbabilityEntry unigramProbabilityEntry = getProbabilityEntry(wordId);
+    if (mHasHistoricalInfo && unigramProbabilityEntry.getHistoricalInfo()->getCount() == 0) {
+        // The word should be treated as a invalid word.
+        return WordAttributes();
+    }
     for (int i = maxPrevWordCount; i >= 0; --i) {
+        if (mustMatchAllPrevWords && prevWordIds.size() > static_cast<size_t>(i)) {
+            break;
+        }
         const TrieMap::Result result = mTrieMap.get(wordId, bitmapEntryIndices[i]);
         if (!result.mIsValid) {
             continue;
@@ -62,36 +71,39 @@ const WordAttributes LanguageModelDictContent::getWordAttributes(const WordIdArr
                 ProbabilityEntry::decode(result.mValue, mHasHistoricalInfo);
         int probability = NOT_A_PROBABILITY;
         if (mHasHistoricalInfo) {
-            const int rawProbability = ForgettingCurveUtils::decodeProbability(
-                    probabilityEntry.getHistoricalInfo(), headerPolicy);
-            if (rawProbability == NOT_A_PROBABILITY) {
-                // The entry should not be treated as a valid entry.
-                continue;
-            }
+            const HistoricalInfo *const historicalInfo = probabilityEntry.getHistoricalInfo();
+            int contextCount = 0;
             if (i == 0) {
                 // unigram
-                probability = rawProbability;
+                contextCount = mGlobalCounters.getTotalCount();
             } else {
                 const ProbabilityEntry prevWordProbabilityEntry = getNgramProbabilityEntry(
                         prevWordIds.skip(1 /* n */).limit(i - 1), prevWordIds[0]);
                 if (!prevWordProbabilityEntry.isValid()) {
                     continue;
                 }
-                if (prevWordProbabilityEntry.representsBeginningOfSentence()) {
-                    probability = rawProbability;
-                } else {
-                    const int prevWordRawProbability = ForgettingCurveUtils::decodeProbability(
-                            prevWordProbabilityEntry.getHistoricalInfo(), headerPolicy);
-                    probability = std::min(MAX_PROBABILITY - prevWordRawProbability
-                            + rawProbability, MAX_PROBABILITY);
+                if (prevWordProbabilityEntry.representsBeginningOfSentence()
+                        && historicalInfo->getCount() == 1) {
+                    // BoS ngram requires multiple contextCount.
+                    continue;
                 }
+                contextCount = prevWordProbabilityEntry.getHistoricalInfo()->getCount();
             }
+            const float rawProbability =
+                    DynamicLanguageModelProbabilityUtils::computeRawProbabilityFromCounts(
+                            historicalInfo->getCount(), contextCount, i + 1);
+            const int encodedRawProbability =
+                    ProbabilityUtils::encodeRawProbability(rawProbability);
+            const int decayedProbability =
+                    DynamicLanguageModelProbabilityUtils::getDecayedProbability(
+                            encodedRawProbability, *historicalInfo);
+            probability = DynamicLanguageModelProbabilityUtils::backoff(
+                    decayedProbability, i + 1 /* n */);
         } else {
             probability = probabilityEntry.getProbability();
         }
         // TODO: Some flags in unigramProbabilityEntry should be overwritten by flags in
         // probabilityEntry.
-        const ProbabilityEntry unigramProbabilityEntry = getProbabilityEntry(wordId);
         return WordAttributes(probability, unigramProbabilityEntry.isBlacklisted(),
                 unigramProbabilityEntry.isNotAWord(),
                 unigramProbabilityEntry.isPossiblyOffensive());
@@ -167,7 +179,8 @@ void LanguageModelDictContent::exportAllNgramEntriesRelatedToWordInner(
                 ProbabilityEntry::decode(entry.value(), mHasHistoricalInfo);
         if (probabilityEntry.isValid()) {
             const WordAttributes wordAttributes = getWordAttributes(
-                    WordIdArrayView(*prevWordIds), wordId, headerPolicy);
+                    WordIdArrayView(*prevWordIds), wordId, true /* mustMatchAllPrevWords */,
+                    headerPolicy);
             outBummpedFullEntryInfo->emplace_back(*prevWordIds, wordId,
                     wordAttributes, probabilityEntry);
         }
@@ -231,7 +244,7 @@ bool LanguageModelDictContent::updateAllEntriesOnInputWord(const WordIdArrayView
             return false;
         }
         mGlobalCounters.updateMaxValueOfCounters(
-                updatedUnigramProbabilityEntry.getHistoricalInfo()->getCount());
+                updatedNgramProbabilityEntry.getHistoricalInfo()->getCount());
         if (!originalNgramProbabilityEntry.isValid()) {
             entryCountersToUpdate->incrementNgramCount(i + 2);
         }
@@ -242,10 +255,9 @@ bool LanguageModelDictContent::updateAllEntriesOnInputWord(const WordIdArrayView
 const ProbabilityEntry LanguageModelDictContent::createUpdatedEntryFrom(
         const ProbabilityEntry &originalProbabilityEntry, const bool isValid,
         const HistoricalInfo historicalInfo, const HeaderPolicy *const headerPolicy) const {
-    const HistoricalInfo updatedHistoricalInfo = ForgettingCurveUtils::createUpdatedHistoricalInfo(
-            originalProbabilityEntry.getHistoricalInfo(), isValid ?
-                    DUMMY_PROBABILITY_FOR_VALID_WORDS : NOT_A_PROBABILITY,
-            &historicalInfo, headerPolicy);
+    const HistoricalInfo updatedHistoricalInfo = HistoricalInfo(historicalInfo.getTimestamp(),
+            0 /* level */, originalProbabilityEntry.getHistoricalInfo()->getCount()
+                    + historicalInfo.getCount());
     if (originalProbabilityEntry.isValid()) {
         return ProbabilityEntry(originalProbabilityEntry.getFlags(), &updatedHistoricalInfo);
     } else {
@@ -311,7 +323,7 @@ int LanguageModelDictContent::getBitmapEntryIndex(const WordIdArrayView prevWord
 
 bool LanguageModelDictContent::updateAllProbabilityEntriesForGCInner(const int bitmapEntryIndex,
         const int prevWordCount, const HeaderPolicy *const headerPolicy,
-        MutableEntryCounters *const outEntryCounters) {
+        const bool needsToHalveCounters, MutableEntryCounters *const outEntryCounters) {
     for (const auto &entry : mTrieMap.getEntriesInSpecifiedLevel(bitmapEntryIndex)) {
         if (prevWordCount > MAX_PREV_WORD_COUNT_FOR_N_GRAM) {
             AKLOGE("Invalid prevWordCount. prevWordCount: %d, MAX_PREV_WORD_COUNT_FOR_N_GRAM: %d.",
@@ -328,33 +340,41 @@ bool LanguageModelDictContent::updateAllProbabilityEntriesForGCInner(const int b
             }
             continue;
         }
-        if (mHasHistoricalInfo && !probabilityEntry.representsBeginningOfSentence()
-                && probabilityEntry.isValid()) {
-            const HistoricalInfo historicalInfo = ForgettingCurveUtils::createHistoricalInfoToSave(
-                    probabilityEntry.getHistoricalInfo(), headerPolicy);
-            if (ForgettingCurveUtils::needsToKeep(&historicalInfo, headerPolicy)) {
-                // Update the entry.
-                const ProbabilityEntry updatedEntry(probabilityEntry.getFlags(), &historicalInfo);
-                if (!mTrieMap.put(entry.key(), updatedEntry.encode(mHasHistoricalInfo),
-                        bitmapEntryIndex)) {
-                    return false;
-                }
-            } else {
+        if (mHasHistoricalInfo && probabilityEntry.isValid()) {
+            const HistoricalInfo *originalHistoricalInfo = probabilityEntry.getHistoricalInfo();
+            if (DynamicLanguageModelProbabilityUtils::shouldRemoveEntryDuringGC(
+                    *originalHistoricalInfo)) {
                 // Remove the entry.
                 if (!mTrieMap.remove(entry.key(), bitmapEntryIndex)) {
                     return false;
                 }
                 continue;
             }
+            if (needsToHalveCounters) {
+                const int updatedCount = originalHistoricalInfo->getCount() / 2;
+                if (updatedCount == 0) {
+                    // Remove the entry.
+                    if (!mTrieMap.remove(entry.key(), bitmapEntryIndex)) {
+                        return false;
+                    }
+                    continue;
+                }
+                const HistoricalInfo historicalInfoToSave(originalHistoricalInfo->getTimestamp(),
+                        originalHistoricalInfo->getLevel(), updatedCount);
+                const ProbabilityEntry updatedEntry(probabilityEntry.getFlags(),
+                        &historicalInfoToSave);
+                if (!mTrieMap.put(entry.key(), updatedEntry.encode(mHasHistoricalInfo),
+                        bitmapEntryIndex)) {
+                    return false;
+                }
+            }
         }
-        if (!probabilityEntry.representsBeginningOfSentence()) {
-            outEntryCounters->incrementNgramCount(prevWordCount + 1);
-        }
+        outEntryCounters->incrementNgramCount(prevWordCount + 1);
         if (!entry.hasNextLevelMap()) {
             continue;
         }
         if (!updateAllProbabilityEntriesForGCInner(entry.getNextLevelBitmapEntryIndex(),
-                prevWordCount + 1, headerPolicy, outEntryCounters)) {
+                prevWordCount + 1, headerPolicy, needsToHalveCounters, outEntryCounters)) {
             return false;
         }
     }
@@ -408,11 +428,11 @@ bool LanguageModelDictContent::getEntryInfo(const HeaderPolicy *const headerPoli
         }
         const ProbabilityEntry probabilityEntry =
                 ProbabilityEntry::decode(entry.value(), mHasHistoricalInfo);
-        const int probability = (mHasHistoricalInfo) ?
-                ForgettingCurveUtils::decodeProbability(probabilityEntry.getHistoricalInfo(),
-                        headerPolicy) : probabilityEntry.getProbability();
-        outEntryInfo->emplace_back(probability,
-                probabilityEntry.getHistoricalInfo()->getTimestamp(),
+        const int priority = mHasHistoricalInfo
+                ? DynamicLanguageModelProbabilityUtils::getPriorityToPreventFromEviction(
+                        *probabilityEntry.getHistoricalInfo())
+                : probabilityEntry.getProbability();
+        outEntryInfo->emplace_back(priority, probabilityEntry.getHistoricalInfo()->getCount(),
                 entry.key(), targetLevel, prevWordIds->data());
     }
     return true;
@@ -420,11 +440,11 @@ bool LanguageModelDictContent::getEntryInfo(const HeaderPolicy *const headerPoli
 
 bool LanguageModelDictContent::EntryInfoToTurncate::Comparator::operator()(
         const EntryInfoToTurncate &left, const EntryInfoToTurncate &right) const {
-    if (left.mProbability != right.mProbability) {
-        return left.mProbability < right.mProbability;
+    if (left.mPriority != right.mPriority) {
+        return left.mPriority < right.mPriority;
     }
-    if (left.mTimestamp != right.mTimestamp) {
-        return left.mTimestamp > right.mTimestamp;
+    if (left.mCount != right.mCount) {
+        return left.mCount < right.mCount;
     }
     if (left.mKey != right.mKey) {
         return left.mKey < right.mKey;
@@ -441,10 +461,9 @@ bool LanguageModelDictContent::EntryInfoToTurncate::Comparator::operator()(
     return false;
 }
 
-LanguageModelDictContent::EntryInfoToTurncate::EntryInfoToTurncate(const int probability,
-        const int timestamp, const int key, const int prevWordCount, const int *const prevWordIds)
-        : mProbability(probability), mTimestamp(timestamp), mKey(key),
-          mPrevWordCount(prevWordCount) {
+LanguageModelDictContent::EntryInfoToTurncate::EntryInfoToTurncate(const int priority,
+        const int count, const int key, const int prevWordCount, const int *const prevWordIds)
+        : mPriority(priority), mCount(count), mKey(key), mPrevWordCount(prevWordCount) {
     memmove(mPrevWordIds, prevWordIds, mPrevWordCount * sizeof(mPrevWordIds[0]));
 }
 
diff --git a/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content.h b/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content.h
index 5b92b96e3..9678c35f9 100644
--- a/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content.h
+++ b/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content.h
@@ -151,13 +151,14 @@ class LanguageModelDictContent {
             const LanguageModelDictContent *const originalContent);
 
     const WordAttributes getWordAttributes(const WordIdArrayView prevWordIds, const int wordId,
-            const HeaderPolicy *const headerPolicy) const;
+            const bool mustMatchAllPrevWords, const HeaderPolicy *const headerPolicy) const;
 
     ProbabilityEntry getProbabilityEntry(const int wordId) const {
         return getNgramProbabilityEntry(WordIdArrayView(), wordId);
     }
 
     bool setProbabilityEntry(const int wordId, const ProbabilityEntry *const probabilityEntry) {
+        mGlobalCounters.addToTotalCount(probabilityEntry->getHistoricalInfo()->getCount());
         return setNgramProbabilityEntry(WordIdArrayView(), wordId, probabilityEntry);
     }
 
@@ -180,8 +181,15 @@ class LanguageModelDictContent {
 
     bool updateAllProbabilityEntriesForGC(const HeaderPolicy *const headerPolicy,
             MutableEntryCounters *const outEntryCounters) {
-        return updateAllProbabilityEntriesForGCInner(mTrieMap.getRootBitmapEntryIndex(),
-                0 /* prevWordCount */, headerPolicy, outEntryCounters);
+        if (!updateAllProbabilityEntriesForGCInner(mTrieMap.getRootBitmapEntryIndex(),
+                0 /* prevWordCount */, headerPolicy, mGlobalCounters.needsToHalveCounters(),
+                outEntryCounters)) {
+            return false;
+        }
+        if (mGlobalCounters.needsToHalveCounters()) {
+            mGlobalCounters.halveCounters();
+        }
+        return true;
     }
 
     // entryCounts should be created by updateAllProbabilityEntries.
@@ -206,11 +214,12 @@ class LanguageModelDictContent {
             DISALLOW_ASSIGNMENT_OPERATOR(Comparator);
         };
 
-        EntryInfoToTurncate(const int probability, const int timestamp, const int key,
+        EntryInfoToTurncate(const int priority, const int count, const int key,
                 const int prevWordCount, const int *const prevWordIds);
 
-        int mProbability;
-        int mTimestamp;
+        int mPriority;
+        // TODO: Remove.
+        int mCount;
         int mKey;
         int mPrevWordCount;
         int mPrevWordIds[MAX_PREV_WORD_COUNT_FOR_N_GRAM + 1];
@@ -219,8 +228,6 @@ class LanguageModelDictContent {
         DISALLOW_DEFAULT_CONSTRUCTOR(EntryInfoToTurncate);
     };
 
-    // TODO: Remove
-    static const int DUMMY_PROBABILITY_FOR_VALID_WORDS;
     static const int TRIE_MAP_BUFFER_INDEX;
     static const int GLOBAL_COUNTERS_BUFFER_INDEX;
 
@@ -233,7 +240,8 @@ class LanguageModelDictContent {
     int createAndGetBitmapEntryIndex(const WordIdArrayView prevWordIds);
     int getBitmapEntryIndex(const WordIdArrayView prevWordIds) const;
     bool updateAllProbabilityEntriesForGCInner(const int bitmapEntryIndex, const int prevWordCount,
-            const HeaderPolicy *const headerPolicy, MutableEntryCounters *const outEntryCounters);
+            const HeaderPolicy *const headerPolicy, const bool needsToHalveCounters,
+            MutableEntryCounters *const outEntryCounters);
     bool turncateEntriesInSpecifiedLevel(const HeaderPolicy *const headerPolicy,
             const int maxEntryCount, const int targetLevel, int *const outEntryCount);
     bool getEntryInfo(const HeaderPolicy *const headerPolicy, const int targetLevel,
diff --git a/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content_global_counters.h b/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content_global_counters.h
index 9953aa425..283c2691a 100644
--- a/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content_global_counters.h
+++ b/native/jni/src/suggest/policyimpl/dictionary/structure/v4/content/language_model_dict_content_global_counters.h
@@ -63,6 +63,10 @@ class LanguageModelDictContentGlobalCounters {
         mTotalCount += 1;
     }
 
+    void addToTotalCount(const int count) {
+        mTotalCount += count;
+    }
+
     void updateMaxValueOfCounters(const int count) {
         mMaxValueOfCounters = std::max(count, mMaxValueOfCounters);
     }
diff --git a/native/jni/src/suggest/policyimpl/dictionary/structure/v4/ver4_patricia_trie_policy.cpp b/native/jni/src/suggest/policyimpl/dictionary/structure/v4/ver4_patricia_trie_policy.cpp
index d3de322f9..96d789f58 100644
--- a/native/jni/src/suggest/policyimpl/dictionary/structure/v4/ver4_patricia_trie_policy.cpp
+++ b/native/jni/src/suggest/policyimpl/dictionary/structure/v4/ver4_patricia_trie_policy.cpp
@@ -110,7 +110,7 @@ const WordAttributes Ver4PatriciaTriePolicy::getWordAttributesInContext(
         return WordAttributes();
     }
     return mBuffers->getLanguageModelDictContent()->getWordAttributes(prevWordIds, wordId,
-            mHeaderPolicy);
+            false /* mustMatchAllPrevWords */, mHeaderPolicy);
 }
 
 int Ver4PatriciaTriePolicy::getProbabilityOfWord(const WordIdArrayView prevWordIds,
@@ -118,18 +118,13 @@ int Ver4PatriciaTriePolicy::getProbabilityOfWord(const WordIdArrayView prevWordI
     if (wordId == NOT_A_WORD_ID || prevWordIds.contains(NOT_A_WORD_ID)) {
         return NOT_A_PROBABILITY;
     }
-    const ProbabilityEntry probabilityEntry =
-            mBuffers->getLanguageModelDictContent()->getNgramProbabilityEntry(prevWordIds, wordId);
-    if (!probabilityEntry.isValid() || probabilityEntry.isBlacklisted()
-            || probabilityEntry.isNotAWord()) {
+    const WordAttributes wordAttributes =
+            mBuffers->getLanguageModelDictContent()->getWordAttributes(prevWordIds, wordId,
+                    true /* mustMatchAllPrevWords */, mHeaderPolicy);
+    if (wordAttributes.isBlacklisted() || wordAttributes.isNotAWord()) {
         return NOT_A_PROBABILITY;
     }
-    if (mHeaderPolicy->hasHistoricalInfoOfWords()) {
-        return ForgettingCurveUtils::decodeProbability(probabilityEntry.getHistoricalInfo(),
-                mHeaderPolicy);
-    } else {
-        return probabilityEntry.getProbability();
-    }
+    return wordAttributes.getProbability();
 }
 
 BinaryDictionaryShortcutIterator Ver4PatriciaTriePolicy::getShortcutIterator(
@@ -152,9 +147,7 @@ void Ver4PatriciaTriePolicy::iterateNgramEntries(const WordIdArrayView prevWordI
                 continue;
             }
             const int probability = probabilityEntry.hasHistoricalInfo() ?
-                    ForgettingCurveUtils::decodeProbability(
-                            probabilityEntry.getHistoricalInfo(), mHeaderPolicy) :
-                    probabilityEntry.getProbability();
+                    0 : probabilityEntry.getProbability();
             listener->onVisitEntry(probability, entry.getWordId());
         }
     }
@@ -386,25 +379,35 @@ bool Ver4PatriciaTriePolicy::updateEntriesForWordWithNgramContext(
             AKLOGE("Cannot add unigarm entry in updateEntriesForWordWithNgramContext().");
             return false;
         }
+        if (!isValidWord) {
+            return true;
+        }
         wordId = getWordId(wordCodePoints, false /* tryLowerCaseSearch */);
     }
 
     WordIdArray<MAX_PREV_WORD_COUNT_FOR_N_GRAM> prevWordIdArray;
     const WordIdArrayView prevWordIds = ngramContext->getPrevWordIds(this, &prevWordIdArray,
             false /* tryLowerCaseSearch */);
-    if (prevWordIds.firstOrDefault(NOT_A_WORD_ID) == NOT_A_WORD_ID
-            && ngramContext->isNthPrevWordBeginningOfSentence(1 /* n */)) {
-        const UnigramProperty beginningOfSentenceUnigramProperty(
-                true /* representsBeginningOfSentence */,
-                true /* isNotAWord */, false /* isPossiblyOffensive */, NOT_A_PROBABILITY,
-                HistoricalInfo(historicalInfo.getTimestamp(), 0 /* level */, 0 /* count */));
-        if (!addUnigramEntry(ngramContext->getNthPrevWordCodePoints(1 /* n */),
-                &beginningOfSentenceUnigramProperty)) {
-            AKLOGE("Cannot add BoS entry in updateEntriesForWordWithNgramContext().");
+    if (ngramContext->isNthPrevWordBeginningOfSentence(1 /* n */)) {
+        if (prevWordIds.firstOrDefault(NOT_A_WORD_ID) == NOT_A_WORD_ID) {
+            const UnigramProperty beginningOfSentenceUnigramProperty(
+                    true /* representsBeginningOfSentence */,
+                    true /* isNotAWord */, false /* isPossiblyOffensive */, NOT_A_PROBABILITY,
+                    HistoricalInfo(historicalInfo.getTimestamp(), 0 /* level */, 0 /* count */));
+            if (!addUnigramEntry(ngramContext->getNthPrevWordCodePoints(1 /* n */),
+                    &beginningOfSentenceUnigramProperty)) {
+                AKLOGE("Cannot add BoS entry in updateEntriesForWordWithNgramContext().");
+                return false;
+            }
+            // Refresh word ids.
+            ngramContext->getPrevWordIds(this, &prevWordIdArray, false /* tryLowerCaseSearch */);
+        }
+        // Update entries for beginning of sentence.
+        if (!mBuffers->getMutableLanguageModelDictContent()->updateAllEntriesOnInputWord(
+                prevWordIds.skip(1 /* n */), prevWordIds[0], true /* isVaild */, historicalInfo,
+                mHeaderPolicy, &mEntryCounters)) {
             return false;
         }
-        // Refresh word ids.
-        ngramContext->getPrevWordIds(this, &prevWordIdArray, false /* tryLowerCaseSearch */);
     }
     if (!mBuffers->getMutableLanguageModelDictContent()->updateAllEntriesOnInputWord(prevWordIds,
             wordId, updateAsAValidWord, historicalInfo, mHeaderPolicy, &mEntryCounters)) {
@@ -542,7 +545,7 @@ const WordProperty Ver4PatriciaTriePolicy::getWordProperty(
         }
     }
     const WordAttributes wordAttributes = languageModelDictContent->getWordAttributes(
-            WordIdArrayView(), wordId, mHeaderPolicy);
+            WordIdArrayView(), wordId, true /* mustMatchAllPrevWords */, mHeaderPolicy);
     const ProbabilityEntry probabilityEntry = languageModelDictContent->getProbabilityEntry(wordId);
     const HistoricalInfo *const historicalInfo = probabilityEntry.getHistoricalInfo();
     const UnigramProperty unigramProperty(probabilityEntry.representsBeginningOfSentence(),