diff --git a/HAN-document_classification/README.md b/HAN-document_classification/README.md
new file mode 100644
index 00000000..11a93297
--- /dev/null
+++ b/HAN-document_classification/README.md
@@ -0,0 +1,36 @@
+## Introduction
+This is the implementation of [Hierarchical Attention Networks for Document Classification](https://www.cs.cmu.edu/~diyiy/docs/naacl16.pdf) paper in PyTorch.
+* Dataset is 600k documents extracted from [Yelp 2018](https://www.yelp.com/dataset) customer reviews
+* Use [NLTK](http://www.nltk.org/) and [Stanford CoreNLP](https://stanfordnlp.github.io/CoreNLP/) to tokenize documents and sentences
+* Both CPU & GPU support
+* The best accuracy is 71%, reaching the same performance in the paper
+
+## Requirement
+* python 3.6
+* pytorch = 0.3.0
+* numpy
+* gensim
+* nltk
+* coreNLP
+
+## Parameters
+According to the paper and experiment, I set model parameters:
+|word embedding dimension|GRU hidden size|GRU layer|word/sentence context vector dimension|
+|---|---|---|---|
+|200|50|1|100|
+
+And the training parameters:
+|Epoch|learning rate|momentum|batch size|
+|---|---|---|---|
+|3|0.01|0.9|64|
+
+## Run
+1. Prepare dataset. Download the [data set](https://www.yelp.com/dataset), and unzip the custom reviews as a file. Use preprocess.py to transform file into data set foe model input.
+2. Train the model. Word enbedding of train data in 'yelp.word2vec'. The model will trained and autosaved in 'model.dict'
+```
+python train
+```
+3. Test the model.
+```
+python evaluate
+```
\ No newline at end of file
diff --git a/HAN-document_classification/data/test_samples.pkl b/HAN-document_classification/data/test_samples.pkl
new file mode 100644
index 00000000..7a904f09
Binary files /dev/null and b/HAN-document_classification/data/test_samples.pkl differ
diff --git a/HAN-document_classification/data/train_samples.pkl b/HAN-document_classification/data/train_samples.pkl
new file mode 100644
index 00000000..a4ad849b
Binary files /dev/null and b/HAN-document_classification/data/train_samples.pkl differ
diff --git a/HAN-document_classification/data/yelp.word2vec b/HAN-document_classification/data/yelp.word2vec
new file mode 100644
index 00000000..38eaf2b4
Binary files /dev/null and b/HAN-document_classification/data/yelp.word2vec differ
diff --git a/HAN-document_classification/evaluate.py b/HAN-document_classification/evaluate.py
new file mode 100644
index 00000000..1f10a9a2
--- /dev/null
+++ b/HAN-document_classification/evaluate.py
@@ -0,0 +1,44 @@
+from model import *
+from train import *
+
+def evaluate(net, dataset, bactch_size=64, use_cuda=False):
+    dataloader = DataLoader(dataset, batch_size=bactch_size, collate_fn=collate, num_workers=0)
+    count = 0
+    if use_cuda:
+        net.cuda()
+    for i, batch_samples in enumerate(dataloader):
+        x, y = batch_samples
+        doc_list = []
+        for sample in x:
+            doc = []
+            for sent_vec in sample:
+                if use_cuda:
+                    sent_vec = sent_vec.cuda()
+                doc.append(Variable(sent_vec, volatile=True))
+            doc_list.append(pack_sequence(doc))
+        if use_cuda:
+            y = y.cuda()
+        predicts = net(doc_list)
+        p, idx = torch.max(predicts, dim=1)
+        idx = idx.data
+        count += torch.sum(torch.eq(idx, y))
+    return count
+
+if __name__ == '__main__':
+    '''
+    Evaluate the performance of model
+    '''
+    from gensim.models import Word2Vec
+    import gensim
+    from gensim import models
+    embed_model = Word2Vec.load('yelp.word2vec')
+    embedding = Embedding_layer(embed_model.wv, embed_model.wv.vector_size)
+    del embed_model
+
+    net = HAN(input_size=200, output_size=5, 
+            word_hidden_size=50, word_num_layers=1, word_context_size=100,
+            sent_hidden_size=50, sent_num_layers=1, sent_context_size=100)
+    net.load_state_dict(torch.load('model.dict'))
+    test_dataset = YelpDocSet('reviews', 199, 4, embedding)
+    correct = evaluate(net, test_dataset, True)
+    print('accuracy {}'.format(correct/len(test_dataset)))
diff --git a/HAN-document_classification/model.py b/HAN-document_classification/model.py
new file mode 100644
index 00000000..ce047f8e
--- /dev/null
+++ b/HAN-document_classification/model.py
@@ -0,0 +1,110 @@
+import torch
+import torch.nn as nn
+from torch.autograd import Variable
+import torch.nn.functional as F
+
+def pack_sequence(tensor_seq, padding_value=0.0):
+    if len(tensor_seq) <= 0:
+        return
+    length = [v.size(0) for v in tensor_seq]
+    max_len = max(length)
+    size = [len(tensor_seq), max_len]
+    size.extend(list(tensor_seq[0].size()[1:]))
+    ans = torch.Tensor(*size).fill_(padding_value)
+    if tensor_seq[0].data.is_cuda:
+        ans = ans.cuda()
+    ans = Variable(ans)
+    for i, v in enumerate(tensor_seq):
+        ans[i, :length[i], :] = v
+    return ans
+
+class HAN(nn.Module):
+    def __init__(self, input_size, output_size, 
+                word_hidden_size, word_num_layers, word_context_size, 
+                sent_hidden_size, sent_num_layers, sent_context_size):
+        super(HAN, self).__init__()
+
+        self.word_layer = AttentionNet(input_size, 
+                                    word_hidden_size, 
+                                    word_num_layers, 
+                                    word_context_size)
+        self.sent_layer = AttentionNet(2* word_hidden_size, 
+                                    sent_hidden_size, 
+                                    sent_num_layers, 
+                                    sent_context_size)
+        self.output_layer = nn.Linear(2* sent_hidden_size, output_size)
+        self.softmax = nn.LogSoftmax(dim=1)
+
+    def forward(self, batch_doc):
+        # input is a sequence of matrix
+        doc_vec_list = []
+        for doc in batch_doc: 
+            sent_mat = self.word_layer(doc) # doc's dim (num_sent, seq_len, word_dim)
+            doc_vec_list.append(sent_mat) # sent_mat's dim (num_sent, vec_dim)
+        doc_vec = self.sent_layer(pack_sequence(doc_vec_list))
+        output = self.softmax(self.output_layer(doc_vec))
+        return output
+
+class AttentionNet(nn.Module):
+    def __init__(self, input_size, gru_hidden_size, gru_num_layers, context_vec_size):
+        super(AttentionNet, self).__init__()
+        
+        self.input_size = input_size
+        self.gru_hidden_size = gru_hidden_size
+        self.gru_num_layers = gru_num_layers
+        self.context_vec_size = context_vec_size
+
+        # Encoder
+        self.gru = nn.GRU(input_size=input_size, 
+                        hidden_size=gru_hidden_size, 
+                        num_layers=gru_num_layers, 
+                        batch_first=True, 
+                        bidirectional=True)
+        # Attention
+        self.fc = nn.Linear(2* gru_hidden_size, context_vec_size)
+        self.tanh = nn.Tanh()
+        self.softmax = nn.Softmax(dim=1)
+        # context vector
+        self.context_vec = nn.Parameter(torch.Tensor(context_vec_size, 1))
+        self.context_vec.data.uniform_(-0.1, 0.1)
+
+    def forward(self, inputs):
+        # GRU part
+        h_t, hidden = self.gru(inputs) # inputs's dim (batch_size, seq_len,  word_dim)
+        u = self.tanh(self.fc(h_t))
+        # Attention part
+        alpha = self.softmax(torch.matmul(u, self.context_vec)) # u's dim (batch_size, seq_len, context_vec_size)
+        output = torch.bmm(torch.transpose(h_t, 1, 2), alpha) # alpha's dim (batch_size, seq_len, 1)
+        return torch.squeeze(output, dim=2) # output's dim (batch_size, 2*hidden_size, 1)
+
+
+if __name__ == '__main__':
+    '''
+    Test the model correctness
+    '''
+    import numpy as np
+    use_cuda = True
+    net = HAN(input_size=200, output_size=5, 
+                word_hidden_size=50, word_num_layers=1, word_context_size=100,
+                sent_hidden_size=50, sent_num_layers=1, sent_context_size=100)
+    optimizer = torch.optim.SGD(net.parameters(), lr=0.01, momentum=0.9)
+    criterion = nn.NLLLoss()
+    test_time = 10
+    batch_size = 64
+    if use_cuda:
+        net.cuda()
+    print('test training')
+    for step in range(test_time):
+        x_data = [torch.randn(np.random.randint(1,10), 200, 200) for i in range(batch_size)]
+        y_data = torch.LongTensor([np.random.randint(0, 5) for i in range(batch_size)])
+        if use_cuda:
+            x_data = [x_i.cuda() for x_i in x_data]
+            y_data = y_data.cuda()
+        x = [Variable(x_i) for x_i in x_data]
+        y = Variable(y_data)
+        predict = net(x)
+        loss = criterion(predict, y)
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+        print(loss.data[0])
\ No newline at end of file
diff --git a/HAN-document_classification/preprocess.py b/HAN-document_classification/preprocess.py
new file mode 100644
index 00000000..0de0d74e
--- /dev/null
+++ b/HAN-document_classification/preprocess.py
@@ -0,0 +1,51 @@
+''''
+    Tokenize yelp dataset's documents using stanford core nlp
+'''
+
+import pickle
+import json
+import nltk
+from nltk.tokenize import stanford
+import os
+
+input_filename = 'review.json'
+
+# config for stanford core nlp
+os.environ['JAVAHOME'] = 'D:\\java\\bin\\java.exe'
+path_to_jar = 'E:\\College\\fudanNLP\\stanford-corenlp-full-2018-02-27\\stanford-corenlp-3.9.1.jar'
+tokenizer = stanford.CoreNLPTokenizer()
+
+in_dirname = 'review'
+out_dirname = 'reviews'
+
+
+f = open(input_filename, encoding='utf-8')
+samples = []
+j = 0
+for i, line in enumerate(f.readlines()):
+    review = json.loads(line)
+    samples.append((review['stars'], review['text']))
+    if (i+1) % 5000 == 0:
+        print(i)
+        pickle.dump(samples, open(in_dirname + '/samples%d.pkl'%j, 'wb'))
+        j += 1
+        samples = []
+pickle.dump(samples, open(in_dirname + '/samples%d.pkl'%j, 'wb'))
+# samples = pickle.load(open(out_dirname + '/samples0.pkl', 'rb'))
+# print(samples[0])
+
+
+for fn in os.listdir(in_dirname):
+    print(fn)
+    precessed = []
+    for stars, text in pickle.load(open(os.path.join(in_dirname, fn), 'rb')):
+        tokens = []
+        sents = nltk.tokenize.sent_tokenize(text)
+        for s in sents:
+            tokens.append(tokenizer.tokenize(s))
+        precessed.append((stars, tokens))
+        # print(tokens)
+        if len(precessed) % 100 == 0:
+            print(len(precessed))
+    pickle.dump(precessed, open(os.path.join(out_dirname, fn), 'wb'))
+    
diff --git a/HAN-document_classification/train.py b/HAN-document_classification/train.py
new file mode 100644
index 00000000..72e2a696
--- /dev/null
+++ b/HAN-document_classification/train.py
@@ -0,0 +1,167 @@
+import os
+import pickle
+
+import nltk
+import numpy as np
+import torch
+
+from model import *
+
+class SentIter:
+    def __init__(self, dirname, count):
+        self.dirname = dirname
+        self.count = int(count)
+
+    def __iter__(self):
+        for f in os.listdir(self.dirname)[:self.count]:
+            with open(os.path.join(self.dirname, f), 'rb') as f:
+                for y, x in pickle.load(f):
+                    for sent in x:
+                        yield sent
+
+def train_word_vec():
+    # load data
+    dirname = 'reviews'
+    sents = SentIter(dirname, 238)
+    # define model and train
+    model = models.Word2Vec(size=200, sg=0, workers=4, min_count=5)
+    model.build_vocab(sents)
+    model.train(sents, total_examples=model.corpus_count, epochs=10)
+    model.save('yelp.word2vec')
+    print(model.wv.similarity('woman', 'man'))
+    print(model.wv.similarity('nice', 'awful'))
+
+class Embedding_layer:
+    def __init__(self, wv, vector_size):
+        self.wv = wv
+        self.vector_size = vector_size
+
+    def get_vec(self, w):
+        try:
+            v = self.wv[w]
+        except KeyError as e:
+            v = np.random.randn(self.vector_size)
+        return v
+
+
+from torch.utils.data import DataLoader, Dataset
+class YelpDocSet(Dataset):
+    def __init__(self, dirname, start_file, num_files, embedding):
+        self.dirname = dirname
+        self.num_files = num_files
+        self._files = os.listdir(dirname)[start_file:start_file + num_files]
+        self.embedding = embedding
+        self._cache = [(-1, None) for i in range(5)]
+
+    def get_doc(self, n):
+        file_id = n // 5000
+        idx = file_id % 5
+        if self._cache[idx][0] != file_id:
+            with open(os.path.join(self.dirname, self._files[file_id]), 'rb') as f:
+                self._cache[idx] = (file_id, pickle.load(f))
+        y, x = self._cache[idx][1][n % 5000]
+        sents = []
+        for s_list in x:
+            sents.append(' '.join(s_list))
+        x = '\n'.join(sents)
+        return x, y-1
+
+    def __len__(self):
+        return len(self._files)*5000
+
+    def __getitem__(self, n):
+        file_id = n // 5000
+        idx = file_id % 5
+        if self._cache[idx][0] != file_id:
+            print('load {} to {}'.format(file_id, idx))
+            with open(os.path.join(self.dirname, self._files[file_id]), 'rb') as f:
+                self._cache[idx] = (file_id, pickle.load(f))
+        y, x = self._cache[idx][1][n % 5000]
+        doc = []
+        for sent in x:
+            if len(sent) == 0:
+                continue
+            sent_vec = []
+            for word in sent:
+                vec = self.embedding.get_vec(word)
+                sent_vec.append(vec.tolist())
+            sent_vec = torch.Tensor(sent_vec)
+            doc.append(sent_vec)
+        if len(doc) == 0:
+            doc = [torch.zeros(1,200)]
+        return doc, y-1
+
+def collate(iterable):
+    y_list = []
+    x_list = []
+    for x, y in iterable:
+        y_list.append(y)
+        x_list.append(x)
+    return x_list, torch.LongTensor(y_list)
+
+def train(net, dataset, num_epoch, batch_size, print_size=10, use_cuda=False):
+    optimizer = torch.optim.SGD(net.parameters(), lr=0.01, momentum=0.9)
+    criterion = nn.NLLLoss()
+
+    dataloader = DataLoader(dataset, 
+                            batch_size=batch_size, 
+                            collate_fn=collate,
+                            num_workers=0)
+    running_loss = 0.0
+
+    if use_cuda:
+        net.cuda()
+    print('start training')
+    for epoch in range(num_epoch):
+        for i, batch_samples in enumerate(dataloader):
+            x, y = batch_samples
+            doc_list = []
+            for sample in x:
+                doc = []
+                for sent_vec in sample:
+                    if use_cuda:
+                        sent_vec = sent_vec.cuda()
+                    doc.append(Variable(sent_vec))
+                doc_list.append(pack_sequence(doc))
+            if use_cuda:
+                y = y.cuda()
+            y = Variable(y)
+            predict = net(doc_list)
+            loss = criterion(predict, y)
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+            running_loss += loss.data[0]
+            if i % print_size == print_size-1:
+                print('{}, {}'.format(i+1, running_loss/print_size))
+                running_loss = 0.0
+                torch.save(net.state_dict(), 'model.dict')
+    torch.save(net.state_dict(), 'model.dict')
+    
+
+if __name__ == '__main__':
+    '''
+    Train process
+    '''
+    from gensim.models import Word2Vec
+    import gensim
+    from gensim import models
+
+    train_word_vec()
+
+    embed_model = Word2Vec.load('yelp.word2vec')
+    embedding = Embedding_layer(embed_model.wv, embed_model.wv.vector_size)
+    del embed_model
+    start_file = 0
+    dataset = YelpDocSet('reviews', start_file, 120-start_file, embedding)
+    print('training data size {}'.format(len(dataset)))
+    net = HAN(input_size=200, output_size=5, 
+            word_hidden_size=50, word_num_layers=1, word_context_size=100,
+            sent_hidden_size=50, sent_num_layers=1, sent_context_size=100)
+    try:
+        net.load_state_dict(torch.load('model.dict'))
+        print("last time trained model has loaded")
+    except Exception:
+        print("cannot load model, train the inital model")
+    
+    train(net, dataset, num_epoch=5, batch_size=64, use_cuda=True)
\ No newline at end of file