diff --git a/paddle/gserver/evaluators/ChunkEvaluator.cpp b/paddle/gserver/evaluators/ChunkEvaluator.cpp
index 3d8af5bcd419e76fb2026eddc95dc409a33c9d92..13f02e51fe9e3831103982130bfdaa3255e1d174 100644
--- a/paddle/gserver/evaluators/ChunkEvaluator.cpp
+++ b/paddle/gserver/evaluators/ChunkEvaluator.cpp
@@ -12,6 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#include <set>
 #include <vector>
 
 #include "paddle/math/Vector.h"
@@ -72,6 +73,7 @@ class ChunkEvaluator : public Evaluator {
 
   std::vector<Segment> labelSegments_;
   std::vector<Segment> outputSegments_;
+  std::set<int> excludedChunkTypes_;
 
 public:
   virtual void init(const EvaluatorConfig& config) {
@@ -105,6 +107,10 @@ public:
     }
     CHECK(config.has_num_chunk_types()) << "Missing num_chunk_types in config";
     otherChunkType_ = numChunkTypes_ = config.num_chunk_types();
+
+    // the chunks of types in excludedChunkTypes_ will not be counted
+    auto& tmp = config.excluded_chunk_types();
+    excludedChunkTypes_.insert(tmp.begin(), tmp.end());
   }
 
   virtual void start() {
@@ -156,7 +162,8 @@ public:
     getSegments(label, length, labelSegments_);
     size_t i = 0, j = 0;
     while (i < outputSegments_.size() && j < labelSegments_.size()) {
-      if (outputSegments_[i] == labelSegments_[j]) {
+      if (outputSegments_[i] == labelSegments_[j] &&
+          excludedChunkTypes_.count(outputSegments_[i].type) != 1) {
         ++numCorrect_;
       }
       if (outputSegments_[i].end < labelSegments_[j].end) {
@@ -168,8 +175,12 @@ public:
         ++j;
       }
     }
-    numLabelSegments_ += labelSegments_.size();
-    numOutputSegments_ += outputSegments_.size();
+    for (auto& segment : labelSegments_) {
+      if (excludedChunkTypes_.count(segment.type) != 1) ++numLabelSegments_;
+    }
+    for (auto& segment : outputSegments_) {
+      if (excludedChunkTypes_.count(segment.type) != 1) ++numOutputSegments_;
+    }
   }
 
   void getSegments(int* label, int length, std::vector<Segment>& segments) {
diff --git a/paddle/gserver/gradientmachines/GradientMachine.h b/paddle/gserver/gradientmachines/GradientMachine.h
index 579eca71d4cdd2545a3a8be1c7f1dacfdd5ef66b..ad82869aec8318cef42f5a51a7e29c233bd23a95 100644
--- a/paddle/gserver/gradientmachines/GradientMachine.h
+++ b/paddle/gserver/gradientmachines/GradientMachine.h
@@ -212,11 +212,7 @@ public:
    * @note    This function will only been implemented and used in a
    *          multithreaded environment.
    */
-  virtual void start(const TrainerConfig& config,
-                     DataProviderPtr dataProvider) {
-    (void)config;
-    (void)dataProvider;
-  }
+  virtual void start() {}
 
   /**
    * @brief   check  each work-thread whether is failed/error/finish,
diff --git a/paddle/gserver/gradientmachines/MultiGradientMachine.cpp b/paddle/gserver/gradientmachines/MultiGradientMachine.cpp
index 88c098b3559d8d2918309aa48329af067f79bdd5..95a4c0e16a91f572a0d78e2fee113e03a82d2056 100644
--- a/paddle/gserver/gradientmachines/MultiGradientMachine.cpp
+++ b/paddle/gserver/gradientmachines/MultiGradientMachine.cpp
@@ -441,7 +441,7 @@ TrainerThread::TrainerThread(const ModelConfig& config,
 TrainerThread::~TrainerThread() { stop(); }
 
 void TrainerThread::start() {
-  gradientMachine_->start(*(TrainerConfig*)nullptr, (DataProviderPtr) nullptr);
+  gradientMachine_->start();
 
   computeThread_.reset(new std::thread([this]() { computeThread(); }));
 
diff --git a/paddle/gserver/gradientmachines/MultiNetwork.cpp b/paddle/gserver/gradientmachines/MultiNetwork.cpp
index 6eb3d8db962161ed4123b4ef4a4bb42147bfdf19..f1308f3721f8d6cf2645d7cd54b14959311ae17f 100644
--- a/paddle/gserver/gradientmachines/MultiNetwork.cpp
+++ b/paddle/gserver/gradientmachines/MultiNetwork.cpp
@@ -109,10 +109,9 @@ void MultiNetwork::onPassEnd() {
   }
 }
 
-void MultiNetwork::start(const TrainerConfig& config,
-                         DataProviderPtr dataProvider) {
+void MultiNetwork::start() {
   for (auto& subNetwork : subNetworks_) {
-    subNetwork->start(config, dataProvider);
+    subNetwork->start();
   }
 }
 
diff --git a/paddle/gserver/gradientmachines/MultiNetwork.h b/paddle/gserver/gradientmachines/MultiNetwork.h
index 89fbf32b4f90bceab60b8335c27b369806faaee1..f04406b983746965d72071c6ea6cfd2b6afcf9dc 100644
--- a/paddle/gserver/gradientmachines/MultiNetwork.h
+++ b/paddle/gserver/gradientmachines/MultiNetwork.h
@@ -54,7 +54,7 @@ public:
     return subNetworks_;
   }
 
-  virtual void start(const TrainerConfig& config, DataProviderPtr dataProvider);
+  virtual void start();
 
   virtual void finish();
 
diff --git a/paddle/gserver/gradientmachines/ParallelNeuralNetwork.cpp b/paddle/gserver/gradientmachines/ParallelNeuralNetwork.cpp
index 980a5851a2734ce42b3417d16a37987dc5ed6b24..c6e3a3b321efb06a1678ec760c514fe1181cd3d7 100644
--- a/paddle/gserver/gradientmachines/ParallelNeuralNetwork.cpp
+++ b/paddle/gserver/gradientmachines/ParallelNeuralNetwork.cpp
@@ -131,11 +131,7 @@ void ParallelNeuralNetwork::forwardBackward(const std::vector<Argument>& inArgs,
   backward(callback);
 }
 
-void ParallelNeuralNetwork::start(const TrainerConfig& config,
-                                  DataProviderPtr dataProvider) {
-  (void)config;
-  (void)dataProvider;
-
+void ParallelNeuralNetwork::start() {
   for (auto& thread : threads_) {
     thread->start();
   }
diff --git a/paddle/gserver/gradientmachines/ParallelNeuralNetwork.h b/paddle/gserver/gradientmachines/ParallelNeuralNetwork.h
index 8f445b1ded3eb8960dc06512dd3f80b00d284acc..39f5682a58e653cabaf4f3d3382d1e0bac5dece9 100644
--- a/paddle/gserver/gradientmachines/ParallelNeuralNetwork.h
+++ b/paddle/gserver/gradientmachines/ParallelNeuralNetwork.h
@@ -56,7 +56,7 @@ public:
                        PassType passType,
                        const UpdateCallback &callback = NULL);
 
-  virtual void start(const TrainerConfig &config, DataProviderPtr dataProvider);
+  virtual void start();
 
   void addComputeThread(int deviceId);
 
diff --git a/paddle/gserver/tests/test_NetworkCompare.cpp b/paddle/gserver/tests/test_NetworkCompare.cpp
index fc60228f816e0cea30ef764c59a8c7875ed4a0e8..0d261059555c971cd509e64802d6c70abc9d2fef 100644
--- a/paddle/gserver/tests/test_NetworkCompare.cpp
+++ b/paddle/gserver/tests/test_NetworkCompare.cpp
@@ -114,7 +114,7 @@ void calcGradient(DataIn& in, DataOut& out, const std::string& configPath) {
       parameters[i]->getBuf(PARAMETER_VALUE)->copyFrom(*in.paraValues[i]);
     }
   }
-  gradientMachine->start(trainer.getConfig(), nullptr);
+  gradientMachine->start();
   gradientMachine->forward(in.inArgs, &outArgs, PASS_TRAIN);
   for (size_t i = 0; i < in.outGrads.size(); i++) {
     // If the all the layers in the config have no parameters, also
diff --git a/paddle/gserver/tests/test_RecurrentGradientMachine.cpp b/paddle/gserver/tests/test_RecurrentGradientMachine.cpp
index e19cf35cd5eb0148879fc3d0d40ea9b106947f9a..150850da4d49a2320acc70ed370cf8728d5c9def 100644
--- a/paddle/gserver/tests/test_RecurrentGradientMachine.cpp
+++ b/paddle/gserver/tests/test_RecurrentGradientMachine.cpp
@@ -28,7 +28,7 @@ class TrainerForTest : public paddle::Trainer {
 public:
   void startTrain() {
     GradientMachine& gm = *this->trainerInternal_.getGradientMachine();
-    gm.start(this->getConfig(), dataProvider_);
+    gm.start();
   }
 
   void finishTrain() {
diff --git a/paddle/trainer/Tester.cpp b/paddle/trainer/Tester.cpp
index 24fac3e5a8141cbec912d276833ec491385b97ab..13aa28ae5d9699d267858d48e46797c756487ddd 100644
--- a/paddle/trainer/Tester.cpp
+++ b/paddle/trainer/Tester.cpp
@@ -257,7 +257,7 @@ void Tester::test() {
   CHECK(testDataProvider_) << "TestData is not specified";
   testDataProvider_->setSkipShuffle();
   testDataProvider_->reset();
-  gradientMachine_->start(*config_, testDataProvider_);
+  gradientMachine_->start();
 
   // For evaluation
   std::vector<std::string> modelList;
diff --git a/paddle/trainer/Trainer.cpp b/paddle/trainer/Trainer.cpp
index 1eec2c432d235ef484b688db08aae8a39f878a85..6c57467cca30d6703a7ebe6babc624b70a427b8b 100644
--- a/paddle/trainer/Trainer.cpp
+++ b/paddle/trainer/Trainer.cpp
@@ -308,7 +308,7 @@ static double genPerturbation(real* d, real* grad, size_t dim) {
 }
 
 real Trainer::checkGradient() {
-  trainerInternal_.getGradientMachine()->start(*config_, dataProvider_);
+  trainerInternal_.getGradientMachine()->start();
   std::vector<ParameterPtr>& parameters =
       trainerInternal_.getGradientMachine()->getNonStaticParameters();
   DataBatch dataBatch;
@@ -390,7 +390,7 @@ void Trainer::startTrain() {
     dataProvider_->reset();
   }
 
-  trainerInternal_.getGradientMachine()->start(*config_, dataProvider_);
+  trainerInternal_.getGradientMachine()->start();
 }
 
 void Trainer::finishTrain() { trainerInternal_.getGradientMachine()->finish(); }
diff --git a/paddle/trainer/tests/test_Compare.cpp b/paddle/trainer/tests/test_Compare.cpp
index 72fc76bea35e433eeb08ba625b4bf6afdda491fb..e855a8fe2e09aa0f16a73f3e7bcc2f32921092f8 100644
--- a/paddle/trainer/tests/test_Compare.cpp
+++ b/paddle/trainer/tests/test_Compare.cpp
@@ -50,7 +50,7 @@ void calcGradient(bool useGpu, comData& Data) {
   trainer.getDataProvider()->getNextBatch(batchSize, &dataBatch);
   CHECK(dataBatch.getSize()) << "No data from data provider";
   vector<Argument>& inArgs = dataBatch.getStreams();
-  trainer.getGradientMachine()->start(trainer.getConfig(), nullptr);
+  trainer.getGradientMachine()->start();
   for (int i = 0; i < 2; ++i) {
     trainer.getGradientMachine()->forwardBackward(
         inArgs, &Data.outArgs, PASS_TRAIN);
diff --git a/paddle/trainer/tests/test_CompareTwoNets.cpp b/paddle/trainer/tests/test_CompareTwoNets.cpp
index 80c61e259e71dd31d7637072248b22a2910c532e..94f65e545d116c802fb4877dc14f07aaaf83a4fb 100644
--- a/paddle/trainer/tests/test_CompareTwoNets.cpp
+++ b/paddle/trainer/tests/test_CompareTwoNets.cpp
@@ -72,7 +72,7 @@ void calcGradient(ComData& data, const string configFile) {
   CHECK(dataBatch.getSize()) << "No data from data provider";
   vector<Argument>& inArgs = dataBatch.getStreams();
 
-  trainer.getGradientMachine()->start(trainer.getConfig(), nullptr);
+  trainer.getGradientMachine()->start();
   trainer.getGradientMachine()->forwardBackward(
       inArgs, &data.outArgs, PASS_TRAIN);
 
diff --git a/proto/ModelConfig.proto b/proto/ModelConfig.proto
index 552af71e76e5adf27f35bb5ad6fd8a69c71df0f1..be4d0041f91cf7d0306d14338b43bb25e052fd58 100644
--- a/proto/ModelConfig.proto
+++ b/proto/ModelConfig.proto
@@ -433,8 +433,10 @@ message EvaluatorConfig {
   repeated string input_layers = 3;
 
   // Used by ChunkEvaluator
-  optional string chunk_scheme = 4; // one of "IOB", "IOE", "IOBES"
-  optional int32 num_chunk_types = 5; // number of chunk types other than "other"
+  // one of "IOB", "IOE", "IOBES"
+  optional string chunk_scheme = 4;
+  // number of chunk types other than "other"
+  optional int32 num_chunk_types = 5;
 
   // Used by PrecisionRecallEvaluator and ClassificationErrorEvaluator
   // For multi binary labels: true if output > classification_threshold
@@ -453,6 +455,10 @@ message EvaluatorConfig {
 
   // whether to delimit the sequence in the seq_text_printer
   optional bool delimited = 11 [default = true];
+
+  // Used by ChunkEvaluator
+  // chunk of these types are not counted
+  repeated int32 excluded_chunk_types = 12;
 }
 
 message LinkConfig {
diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py
index ea3e4308fe05be464c3e8c6b84d8b7be8a30c016..39892d0533aab468d808274146ae1a0f72170495 100644
--- a/python/paddle/trainer/config_parser.py
+++ b/python/paddle/trainer/config_parser.py
@@ -1240,7 +1240,8 @@ def Evaluator(
         dict_file=None,
         result_file=None,
         num_results=None,
-        delimited=None, ):
+        delimited=None,
+        excluded_chunk_types=None, ):
     evaluator = g_config.model_config.evaluators.add()
     evaluator.type = type
     evaluator.name = MakeLayerNameInSubmodel(name)
@@ -1269,6 +1270,9 @@ def Evaluator(
     if delimited is not None:
         evaluator.delimited = delimited
 
+    if excluded_chunk_types:
+        evaluator.excluded_chunk_types.extend(excluded_chunk_types)
+
 
 class LayerBase(object):
     def __init__(
diff --git a/python/paddle/trainer_config_helpers/evaluators.py b/python/paddle/trainer_config_helpers/evaluators.py
index 3e0e88972c58e8c853e79e21f839943ae4b027d6..bd247ea9af9d8dfb2d476cdc62638bd65c11add5 100644
--- a/python/paddle/trainer_config_helpers/evaluators.py
+++ b/python/paddle/trainer_config_helpers/evaluators.py
@@ -57,19 +57,21 @@ def evaluator(*attrs):
     return impl
 
 
-def evaluator_base(input,
-                   type,
-                   label=None,
-                   weight=None,
-                   name=None,
-                   chunk_scheme=None,
-                   num_chunk_types=None,
-                   classification_threshold=None,
-                   positive_label=None,
-                   dict_file=None,
-                   result_file=None,
-                   num_results=None,
-                   delimited=None):
+def evaluator_base(
+        input,
+        type,
+        label=None,
+        weight=None,
+        name=None,
+        chunk_scheme=None,
+        num_chunk_types=None,
+        classification_threshold=None,
+        positive_label=None,
+        dict_file=None,
+        result_file=None,
+        num_results=None,
+        delimited=None,
+        excluded_chunk_types=None, ):
     """
     Evaluator will evaluate the network status while training/testing.
 
@@ -127,7 +129,8 @@ def evaluator_base(input,
         positive_label=positive_label,
         dict_file=dict_file,
         result_file=result_file,
-        delimited=delimited)
+        delimited=delimited,
+        excluded_chunk_types=excluded_chunk_types, )
 
 
 @evaluator(EvaluatorAttribute.FOR_CLASSIFICATION)
@@ -330,7 +333,8 @@ def chunk_evaluator(
         label,
         chunk_scheme,
         num_chunk_types,
-        name=None, ):
+        name=None,
+        excluded_chunk_types=None, ):
     """
     Chunk evaluator is used to evaluate segment labelling accuracy for a
     sequence. It calculates the chunk detection F1 score.
@@ -376,6 +380,8 @@ def chunk_evaluator(
     :param num_chunk_types: number of chunk types other than "other"
     :param name: The Evaluator name, it is optional.
     :type name: basename|None
+    :param excluded_chunk_types: chunks of these types are not considered
+    :type excluded_chunk_types: list of integer|None
     """
     evaluator_base(
         name=name,
@@ -383,7 +389,8 @@ def chunk_evaluator(
         input=input,
         label=label,
         chunk_scheme=chunk_scheme,
-        num_chunk_types=num_chunk_types)
+        num_chunk_types=num_chunk_types,
+        excluded_chunk_types=excluded_chunk_types, )
 
 
 @evaluator(EvaluatorAttribute.FOR_UTILS)