Add a comment on a caveat for future reference.
Change-Id: I328a0cd4346275aac960e1369bf370688a004e11
diff --git a/native/jni/src/bigram_dictionary.cpp b/native/jni/src/bigram_dictionary.cpp
index 8d0c859..eb4bf8d1 100644
--- a/native/jni/src/bigram_dictionary.cpp
+++ b/native/jni/src/bigram_dictionary.cpp
@@ -126,6 +126,11 @@
// codesSize == 0 means we are trying to find bigram predictions.
if (codesSize < 1 || checkFirstCharacter(bigramBuffer)) {
const int bigramFreq = UnigramDictionary::MASK_ATTRIBUTE_FREQUENCY & bigramFlags;
+ // Due to space constraints, the frequency for bigrams is approximate - the lower the
+ // unigram frequency, the worse the precision. The theoritical maximum error in
+ // resulting frequency is 8 - although in the practice it's never bigger than 3 or 4
+ // in very bad cases. This means that sometimes, we'll see some bigrams interverted
+ // here, but it can't get too bad.
const int frequency =
BinaryFormat::computeFrequencyForBigram(unigramFreq, bigramFreq);
if (addWordBigram(bigramBuffer, length, frequency)) {