diff --git a/icefall/shared/make_kn_lm.py b/icefall/shared/make_kn_lm.py
index a52e8da71..8d0170f8b 100755
--- a/icefall/shared/make_kn_lm.py
+++ b/icefall/shared/make_kn_lm.py
@@ -165,9 +165,9 @@ class NgramCounts:
                 n1 += stat[1]
                 n2 += stat[2]
             assert n1 + 2 * n2 > 0
-            self.d.append(max(0.001, n1 * 1.0) / (n1 + 2 * n2))   # We are doing this max(0.001, xxx) to avoid zero discounting constant D, 
-                                                                  # which could happen if the number of symbols is small and all w in the vocab 
-                                                                  # has been seen after certain h. This can cause division by zero in computing BOW.
+            self.d.append(max(0.001, n1 * 1.0) / (n1 + 2 * n2))   # We are doing this max(0.001, xxx) to avoid zero discounting constant D due to n1=0, 
+                                                                  # which could happen if the number of symbols is small.
+                                                                  # Otherwise, zero discounting constant can cause division by zero in computing BOW.
 
     def cal_f(self):
         # f(a_z) is a probability distribution of word sequence a_z.