diff --git a/models/contentunderstanding/classification/config.yaml b/models/contentunderstanding/classification/config.yaml
index d8ec77807ac4cd4f1b1584160fce38f3596d36e7..9e0bdd1e851ada704eb2377efe0a82154fd2b371 100644
--- a/models/contentunderstanding/classification/config.yaml
+++ b/models/contentunderstanding/classification/config.yaml
@@ -25,6 +25,7 @@ hyper_parameters:
   optimizer:
     class: Adagrad
     learning_rate: 0.001
+  is_sparse: False
 
 mode: runner1
 
diff --git a/models/contentunderstanding/classification/model.py b/models/contentunderstanding/classification/model.py
index f7e6b6c344c913bf082be501ae16470488758160..45c50b9bd746200ee51391a32d6fb59cd9876828 100644
--- a/models/contentunderstanding/classification/model.py
+++ b/models/contentunderstanding/classification/model.py
@@ -27,6 +27,7 @@ class Model(ModelBase):
         self.emb_dim = 8
         self.hid_dim = 128
         self.class_dim = 2
+        self.is_sparse = envs.get_global_env("hyper_parameters.is_sparse", False)
 
     def input_data(self, is_infer=False, **kwargs):
         data = fluid.data(
@@ -42,7 +43,7 @@ class Model(ModelBase):
         seq_len = input[2]
 
         # embedding layer
-        emb = fluid.embedding(input=data, size=[self.dict_dim, self.emb_dim])
+        emb = fluid.embedding(input=data, size=[self.dict_dim, self.emb_dim], is_sparse=self.is_sparse)
         emb = fluid.layers.sequence_unpad(emb, length=seq_len)
         # convolution layer
         conv = fluid.nets.sequence_conv_pool(