Lecture 8 Deep Learning for NLP: Recurrent Networks

Recurrent Networks 循环神经网络

Problem of N-gram Language Model N-gram 语言模型的问题

• Cen be implemented using counts with smoothing 可以用平滑计数实现

• Can be implemented using feed-forward neural networks 可以用前馈神经网络实现

• Problem: limited context 问题：上下文限制

• E.g. Generate sentences using trigram model: 例如：使用 trigram 模型生成句子：

Recurrent Neural Network(RNN) 循环神经网络

• Allow representation of arbitrarily sized inputs 允许表示任意大小的输入

• Core idea: processes the input sequence one at a time, by applying a recurrence formula 核心思想：一次处理一个输入序列，通过应用递归公式

• Uses a state vector to represent contexts that have been previously processed 使用状态向量表示之前处理过的上下文

• RNN Neuron: RNN 神经元

  

• RNN States: RNN 状态

  
  
  
  
  
  
  Activation 激活函数:
  
  
  

• RNN Unrolled: 展开的 RNN

  

  • Same parameters are used across all time steps 同一参数 在所有时间步长中都被使用

• Training RNN: 训练 RNN

  • An unrolled RNN is a very deep neural network. But parameters are shared across all time steps 展开的 RNN 是一个非常深的神经网络。但是参数在所有时间步中都是共享的
  • To train RNN, just need to create the unrolled computation graph given an input sequence and use backpropagation algorithm to compute gradients as usual. 要训练 RNN，只需根据输入序列创建展开的计算图，并使用反向传播算法计算梯度
  • This procedure is called backpropagation through time. 这个过程叫做时间反向传播

    E.g of unrolled equation: 展开方程的例子
    
    
    
    

RNN Language Model: RNN 语言模型

• is current word (e.g. eats) mapped to an embedding 是当前词（例如 eats）映射到一个嵌入

• contains information of the previous words (e.g. a and cow) 包含前面词的信息（例如 a 和 cow）

• is the next word (e.g. grass) 是下一个词（例如 grass）

• Training:

  • Vocabulary 词汇: [a, cow, eats, grass]

  • Training example 训练样本: a cow eats grass

  • Training process 训练过程:

    
    
    
    
    
    

  • Losses:

    • 
    • 
    • 

    • Total loss:

• Generation:

  

• Problems of RNN: RNN 的问题

  • Error Propagation: Unable to recover from errors in intermediate steps 错误传播：无法从中间步骤的错误中恢复
  • Low diversity in generated language 生成的语言多样性低
  • Tend to generate bland or generic language 倾向于生成乏味或通用的语言

Long Short-Term Memory Networks

Long Short-Term Memory Model (LSTM) 长短期记忆模型（LSTM）

• RNN has the capability to model infinite context. But it cannot capture long-range dependencies in practice due to the vanishing gradients RNN 具有建模无限上下文的能力。但由于梯度消失，实际上无法捕捉长距离依赖性

• Vanishing Gradient: Gradients in later steps diminish quickly during backpropagation. Earlier inputs do not get much update. 梯度消失：在反向传播过程中，后续步骤的梯度快速减小。较早的输入没有得到太多更新。

• LSTM is introduced to solve vanishing gradients LSTM 用来解决梯度消失问题

• Core idea: have memory cells that preserve gradients across time. Access to the memory cells is controlled by gates. 核心思想：拥有跨时间保存梯度的记忆单元。通过门控制对记忆单元的访问。

• Gates: For each input, a gate decides: 门：对于每个输入，门决定

  • How much the new input should be written to the memory cell 应该将多少新输入写入记忆单元
  • How much content of the current memory cell should be forgotten 应该忘记当前记忆单元的多少内容

• Comparison between simple RNN and LSTM: 简单 RNN 和 LSTM 的比较

  

Gating Vector 门向量

• A gate is a vector. Each element of the gate has values between 0 and 1. Use sigmoid function to produce . 门 是一个向量。门的每个元素的值在 0 到 1 之间。使用 sigmoid 函数来产生 。

• is multiplied component-wise with vector to determine how much information to keep for 和向量 乘以 component-wise 来确定对 保留多少信息

  

Forget Gate 忘记门

• Controls how much information to forget in the memory cell 控制在记忆单元 中忘记多少信息

• E.g. Given Tha cas that the boy predict the next word likes 例如，给定 Tha cas that the boy 预测下一个词 likes

  • Memory cell was storing noun information cats 记忆单元正在存储名词信息 cats
  • The cell should now forget cats and store boy to correctly predict the singular verb likes 该单元现在应该忘记 cats 并存储 boy 以正确预测单数动词 likes

Input Gate 输入门

• Input gate controls how much new information to put to memory cell 输入门控制将多少新信息放入记忆单元

• is new distilled information to be added 是要添加的新提炼信息

Update Memory Cell 更新记忆单元

• Use the forget and input gates to update memory cell 使用忘记门和输入门来更新记忆单元

Output Gate 输出门

• Output gate controls how much to distill the content of the memory cell to create the next state 输出门控制如何提炼记忆单元的内容以创建下一个状态

Disadvantages of LSTM LSTM 的缺点

• Introduces some but not many parameters 引入了一些但并不多的参数
• Still unable to capture very long range dependencies 仍无法捕获非常长的依赖性
• Slower but not much slower than simple RNN 比简单的 RNN 慢，但并不比 RNN 慢太多

Applications of RNN RNN 的应用

Example Applications 示例应用

• Shakespeare Generator 莎士比亚生成器:

  • Training data: all works fo Shakespeare 训练数据：莎士比亚的所有作品
  • Model: Character RNN, hidden dimension = 512 模型：Character RNN，隐藏维度 = 512

• Wikipedia Generator: 维基百科生成器

  • Training data: 100MB of Wikipedia raw data 训练数据：100MB的维基百科原始数据

• Code Generator 代码生成器

• Text Classification 文本分类

  • RNNs can be used in variety NLP tasks. Particularly suited for tasks where order of words matter. E.g. sentiment analysis RNNs可以用于各种NLP任务。特别适合于单词顺序很重要的任务。例如，情感分析

  

• Sequence Labeling: E.g. POS tagging 序列标记：例如，词性标注

  

Variants of LSTM LSTM的变种

• Peephole connections: allow gates to look at cell state 窥视孔连接：允许门看到单元状态

  

• Gated recurrent unit (GRU): Simplified variant with only 2 gates and no memory cell 门控循环单元（GRU）：简化的变体，只有2个门，没有记忆单元

  

• Multi-layer LSTM 多层LSTM

  

• Bidirectional LSTM 双向LSTM