【NLP實戰】手把手帶你RNN文本分類
- 2020 年 3 月 13 日
- 筆記
寫在前面
這是NLP保姆級教程的第二篇----基於RNN的文本分類實現(Text RNN)
參考的的論文是來自2016年復旦大學IJCAI上的發表的關於循環神經網路在多任務文本分類上的應用:Recurrent Neural Network for Text Classification with Multi-Task Learning[1]
論文概覽
在先前的許多工作中,模型的學習都是基於單任務,對於複雜的問題,也可以分解為簡單且相互獨立的子問題來單獨解決,然後再合併結果,得到最初複雜問題的結果。這樣做看似合理,其實是不正確的,因為現實世界中很多問題不能分解為一個一個獨立的子問題,即使可以分解,各個子問題之間也是相互關聯的,通過一些共享因素或「共享表示(share representation)」 聯繫在一起。把現實問題當做一個個獨立的單任務處理,往往會忽略了問題之間所富含的豐富的關聯資訊。
上面的問題引出了本文的重點——「多任務學習(Multi-task learning)」,把多個相關(related)的任務(task)放在一起學習。多個任務之間共享一些因素,它們可以在學習過程中,共享它們所學到的資訊,這是單任務學習沒有具備的。相關聯的多任務學習比單任務學習能去的更好的泛化(generalization)效果。本文基於 RNN 循環神經網路,提出三種不同的資訊共享機制,整體網路是基於所有的任務共同學習得到。
下圖展示的是單任務學習和多任務學習的流程圖,可以對比一下區別。

下面具體介紹一下文章中的三個模型。
Model I: Uniform-Layer Architecture

在他們提出的第一個模型中,不同的任務共享一個LSTM網路層和一個embedding layer,此外每個任務還有其自己的embedding layer。所以對於上圖中的任務m,輸入x包括了兩個部分:
其中等號右側第一項和第二項分別表示該任務「特有」的word embedding和該模型中「共享」的word embedding,兩者做一個concatenation。
LSTM網路層是所有任務所共享的,對於任務m的最後sequence representation為LSTM的輸出:
Model II: Coupled-Layer Architecture

在第二個模型中,為每個任務都指定了「特定」的LSTM layer,但是不同任務間的LSTM layer可以共享資訊。
為了更好地控制在不同LSTM layer之間的資訊流動,作者提出了一個global gating unit,使得模型具有決定資訊流動程度的能力。
為此,他們改寫了LSTM中的表達式:
其中,
Model III: Shared-Layer Architecture

與模型二相似,作者也為每個單獨的任務指派了特定的LSTM層,但是對於整體的模型使用了雙向的LSTM,這樣可以使得資訊共享更為準確。
模型表現
論文作者在4個數據集上對上述模型做了評價,並和其他state-of-the-art的網路模型進行了對比,均顯示最好的效果。
程式碼實現
RNN的程式碼框架和上一篇介紹的CNN類似,首先定義一個RNN類來實現論文中的模型
class RNN(BaseModel): """ A RNN class for sentence classification With an embedding layer + Bi-LSTM layer + FC layer + softmax """ def __init__(self, sequence_length, num_classes, vocab_size, embed_size, learning_rate, decay_steps, decay_rate, hidden_size, is_training, l2_lambda, grad_clip, initializer=tf.random_normal_initializer(stddev=0.1)):
這裡的模型包括了一層embedding,一層雙向LSTM,一層全連接層最後接上一個softmax分類函數。
然後依次定義模型,訓練,損失等函數在後續調用。
def inference(self): """ 1. embedding layer 2. Bi-LSTM layer 3. concat Bi-LSTM output 4. FC(full connected) layer 5. softmax layer """ # embedding layer with tf.name_scope('embedding'): self.embedded_words = tf.nn.embedding_lookup(self.Embedding, self.input_x) # Bi-LSTM layer with tf.name_scope('Bi-LSTM'): lstm_fw_cell = rnn.BasicLSTMCell(self.hidden_size) lstm_bw_cell = rnn.BasicLSTMCell(self.hidden_size) if self.dropout_keep_prob is not None: lstm_fw_cell = rnn.DropoutWrapper(lstm_fw_cell, output_keep_prob=self.dropout_keep_prob) lstm_bw_cell = rnn.DropoutWrapper(lstm_bw_cell, output_keep_prob=self.dropout_keep_prob) outputs, output_states = tf.nn.bidirectional_dynamic_rnn(lstm_fw_cell, lstm_bw_cell, self.embedded_words, dtype=tf.float32) output = tf.concat(outputs, axis=2) output_last = tf.reduce_mean(output, axis=1) # FC layer with tf.name_scope('output'): self.score = tf.matmul(output_last, self.W_projection) + self.b_projection return self.score def loss(self): # loss with tf.name_scope('loss'): losses = tf.nn.softmax_cross_entropy_with_logits(labels=self.input_y, logits=self.score) data_loss = tf.reduce_mean(losses) l2_loss = tf.add_n([tf.nn.l2_loss(cand_v) for cand_v in tf.trainable_variables() if 'bias' not in cand_v.name]) * self.l2_lambda data_loss += l2_loss return data_loss def train(self): learning_rate = tf.train.exponential_decay(self.learning_rate, self.global_step, self.decay_steps, self.decay_rate, staircase=True) optimizer = tf.train.AdamOptimizer(learning_rate) grads_and_vars = optimizer.compute_gradients(self.loss_val) grads_and_vars = [(tf.clip_by_norm(grad, self.grad_clip), val) for grad, val in grads_and_vars] train_op = optimizer.apply_gradients(grads_and_vars, global_step=self.global_step) return train_op訓練部分的數據集這裡就直接採用CNN那篇文章相同的數據集(懶...),預處理的方式與函數等都是一樣的,,,
def train(x_train, y_train, vocab_processor, x_dev, y_dev): with tf.Graph().as_default(): session_conf = tf.ConfigProto( # allows TensorFlow to fall back on a device with a certain operation implemented allow_soft_placement= FLAGS.allow_soft_placement, # allows TensorFlow log on which devices (CPU or GPU) it places operations log_device_placement=FLAGS.log_device_placement ) sess = tf.Session(config=session_conf) with sess.as_default(): # initialize cnn rnn = RNN(sequence_length=x_train.shape[1], num_classes=y_train.shape[1], vocab_size=len(vocab_processor.vocabulary_), embed_size=FLAGS.embed_size, l2_lambda=FLAGS.l2_reg_lambda, is_training=True, grad_clip=FLAGS.grad_clip, learning_rate=FLAGS.learning_rate, decay_steps=FLAGS.decay_steps, decay_rate=FLAGS.decay_rate, hidden_size=FLAGS.hidden_size ) # output dir for models and summaries timestamp = str(time.time()) out_dir = os.path.abspath(os.path.join(os.path.curdir, 'run', timestamp)) if not os.path.exists(out_dir): os.makedirs(out_dir) print('Writing to {} n'.format(out_dir)) # checkpoint dir. checkpointing – saving the parameters of your model to restore them later on. checkpoint_dir = os.path.abspath(os.path.join(out_dir, FLAGS.ckpt_dir)) checkpoint_prefix = os.path.join(checkpoint_dir, 'model') if not os.path.exists(checkpoint_dir): os.makedirs(checkpoint_dir) saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints) # Write vocabulary vocab_processor.save(os.path.join(out_dir, 'vocab')) # Initialize all sess.run(tf.global_variables_initializer()) def train_step(x_batch, y_batch): """ A single training step :param x_batch: :param y_batch: :return: """ feed_dict = { rnn.input_x: x_batch, rnn.input_y: y_batch, rnn.dropout_keep_prob: FLAGS.dropout_keep_prob } _, step, loss, accuracy = sess.run( [rnn.train_op, rnn.global_step, rnn.loss_val, rnn.accuracy], feed_dict=feed_dict ) time_str = datetime.datetime.now().isoformat() print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy)) def dev_step(x_batch, y_batch): """ Evaluate model on a dev set Disable dropout :param x_batch: :param y_batch: :param writer: :return: """ feed_dict = { rnn.input_x: x_batch, rnn.input_y: y_batch, rnn.dropout_keep_prob: 1.0 } step, loss, accuracy = sess.run( [rnn.global_step, rnn.loss_val, rnn.accuracy], feed_dict=feed_dict ) time_str = datetime.datetime.now().isoformat() print("dev results:{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy)) # generate batches batches = data_process.batch_iter(list(zip(x_train, y_train)), FLAGS.batch_size, FLAGS.num_epochs) # training loop for batch in batches: x_batch, y_batch = zip(*batch) train_step(x_batch, y_batch) current_step = tf.train.global_step(sess, rnn.global_step) if current_step % FLAGS.validate_every == 0: print('n Evaluation:') dev_step(x_dev, y_dev) print('') path = saver.save(sess, checkpoint_prefix, global_step=current_step) print('Save model checkpoint to {} n'.format(path)) def main(argv=None): x_train, y_train, vocab_processor, x_dev, y_dev = prepocess() train(x_train, y_train, vocab_processor, x_dev, y_dev) if __name__ == '__main__': tf.app.run()本文參考資料
[1]
Recurrent Neural Network for Text Classification with Multi-Task Learning: https://arxiv.org/abs/1605.05101
- END -

The End

