diff --git a/PaddleNLP/lexical_analysis/README.md b/PaddleNLP/lexical_analysis/README.md
index e35c9d58c68922a329c6f6675f0d39e47259ffea..3a5034a50568a4ddc91d5fb39cb3a9c21d582c1a 100644
--- a/PaddleNLP/lexical_analysis/README.md
+++ b/PaddleNLP/lexical_analysis/README.md
@@ -18,6 +18,8 @@ Lexical Analysis of Chinese，简称 LAC，是一个联合的词法分析模型
 
 本项目依赖 PaddlePaddle 1.3.2 及以上版本，安装请参考官网 [快速安装](http://www.paddlepaddle.org/paddle#quick-start)。
 
+> Warning: GPU 和 CPU 版本的 PaddlePaddle 分别是 paddlepaddle-gpu 和 paddlepaddle，请安装时注意区别。
+
 #### 2. 克隆代码
 克隆工具集代码库到本地
 ```bash
diff --git a/PaddleNLP/lexical_analysis/reader.py b/PaddleNLP/lexical_analysis/reader.py
index e7099eed440c054811d92a6a8dd97cbe9a906412..4655c5ebd282e3c7be43f702d7917cb23be1b931 100644
--- a/PaddleNLP/lexical_analysis/reader.py
+++ b/PaddleNLP/lexical_analysis/reader.py
@@ -93,7 +93,7 @@ class Dataset(object):
                 for line in fread:
                     words = line.strip("\n").split("\002")
                     word_ids = self.word_to_ids(words)
-                    yield word_ids[0:max_seq_len]
+                    yield word_ids[0:max_seq_len], [0 for _ in word_ids][0: max_seq_len]
             else:
                 assert len(headline) == 2 and headline[0] == "text_a" and headline[1] == "label"
                 for line in fread:
diff --git a/PaddleNLP/lexical_analysis/run_ernie_sequence_labeling.py b/PaddleNLP/lexical_analysis/run_ernie_sequence_labeling.py
index 135c6b27fc058f20584572b3ae1c2276dc31e176..96092dc29e9dc81f1b853d735ca18fe80b52b2fb 100644
--- a/PaddleNLP/lexical_analysis/run_ernie_sequence_labeling.py
+++ b/PaddleNLP/lexical_analysis/run_ernie_sequence_labeling.py
@@ -75,6 +75,10 @@ run_type_g.add_arg("do_infer", bool, True, "Whether to perform inference.")
 args = parser.parse_args()
 # yapf: enable.
 
+sys.path.append('../models/')
+from model_check import check_cuda
+check_cuda(args.use_cuda)
+
 def ernie_pyreader(args, pyreader_name):
     """define standard ernie pyreader"""
     pyreader = fluid.layers.py_reader(
diff --git a/PaddleNLP/lexical_analysis/run_sequence_labeling.py b/PaddleNLP/lexical_analysis/run_sequence_labeling.py
index 55265b99259a4bb387979a890f4c80e9d4b273d0..6f20cfe34a4044ffa376f07d63efcebdc5a6ffb7 100644
--- a/PaddleNLP/lexical_analysis/run_sequence_labeling.py
+++ b/PaddleNLP/lexical_analysis/run_sequence_labeling.py
@@ -25,7 +25,6 @@ import utils
 sys.path.append("../")
 from models.sequence_labeling import nets
 
-
 # yapf: disable
 parser = argparse.ArgumentParser(__doc__)
 
@@ -71,6 +70,10 @@ parser.add_argument('--enable_ce', action='store_true', help='If set, run the ta
 args = parser.parse_args()
 # yapf: enable.
 
+sys.path.append('../models/')
+from model_check import check_cuda
+check_cuda(args.use_cuda)
+
 print(args)
 
 
diff --git a/PaddleNLP/lexical_analysis/utils.py b/PaddleNLP/lexical_analysis/utils.py
index 1ef20b22eb17e3081f9d39f7b30ee751d3c504a0..e3085fe0e788f8979d6e0dddc8d56c85446b8c3f 100644
--- a/PaddleNLP/lexical_analysis/utils.py
+++ b/PaddleNLP/lexical_analysis/utils.py
@@ -81,13 +81,27 @@ def parse_result(words, crf_decode, dataset):
     for sent_index in range(batch_size):
         sent_out_str = ""
         sent_len = offset_list[sent_index + 1] - offset_list[sent_index]
+        last_word = ""
+        last_tag = ""
         for tag_index in range(sent_len): # iterate every word in sent
             index = tag_index + offset_list[sent_index]
             cur_word_id = str(words[index][0])
             cur_tag_id = str(crf_decode[index][0])
             cur_word = dataset.id2word_dict[cur_word_id]
             cur_tag = dataset.id2label_dict[cur_tag_id]
-            sent_out_str += cur_word + u"/" + cur_tag + u" "
+            if last_word == "":
+                last_word = cur_word
+                last_tag = cur_tag[:-2]
+            elif cur_tag.endswith("-B") or cur_tag == "O":
+                sent_out_str += last_word + u"/" + last_tag + u" "
+                last_word = cur_word
+                last_tag = cur_tag[:-2]
+            elif cur_tag.endswith("-I"):
+                last_word += cur_word
+            else:
+                raise ValueError("invalid tag: %s" % (cur_tag))
+        if cur_word != "":
+            sent_out_str += last_word + u"/" + last_tag + u" "
         sent_out_str = to_str(sent_out_str.strip())
         batch_out_str.append(sent_out_str)
     return batch_out_str