From cb99376307f0d57e2935449f93fc162253dcdd01 Mon Sep 17 00:00:00 2001 From: Jean Chalard Date: Tue, 29 May 2012 16:04:07 +0900 Subject: [PATCH] Fix a bug where the bigram freq would be underevaluated The difference in score is not large, but it's still a bug Change-Id: Ie22c2b6e1206e829c1c8af096469df05af14d47b --- native/jni/src/binary_format.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/native/jni/src/binary_format.h b/native/jni/src/binary_format.h index e42589b81..85fdd9418 100644 --- a/native/jni/src/binary_format.h +++ b/native/jni/src/binary_format.h @@ -537,7 +537,7 @@ inline int BinaryFormat::computeFrequencyForBigram(const int unigramFreq, const // while a value of 15 represents the middle of the top step. // See makedict.BinaryDictInputOutput for details. const float stepSize = ((float)MAX_FREQ - unigramFreq) / (1.5f + MAX_BIGRAM_FREQ); - return (int)(unigramFreq + bigramFreq * stepSize); + return (int)(unigramFreq + (bigramFreq + 1) * stepSize); } // This returns a probability in log space.