diff --git a/paddle/operators/multiplex_op.cc b/paddle/operators/multiplex_op.cc
index 342b3911ae6d5e0a0a3da14fd93b2f78e63cbe2c..7b50444d16dc57fd14b918d1159e3e21ecd1f1c4 100644
--- a/paddle/operators/multiplex_op.cc
+++ b/paddle/operators/multiplex_op.cc
@@ -94,7 +94,7 @@ class MultiplexGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(!ctx.MultiOutputVar(framework::GradVarName("X")).empty(),
                    "Output(X@Grad) should not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
-                            "Input(Out@GRAD) shouldn't be null.");
+                            "Input(Out@GRAD) should not be null.");
     auto d_ins = ctx.MultiOutput<Tensor>(framework::GradVarName("X"));
     auto ins = ctx.MultiInput<Tensor>("X");
     // No need to compute gradient for Input(Ids)