[吴恩达团队自然语言处理第二课_3]词嵌入与神经网络

[吴恩达团队自然语言处理第二课_3]词嵌入与神经网络

Overview#

主要应用#

• Semantic analogies and similarity
• Sentiment analysis
• Classification of
  customer feedback

高级应用#

• Machine translation
• Information extraction
• Question answering

Learning objectives#

• Identify the key concepts of word representations
• Generate word embeddings
• Prepare text for machine learning
• Implement the continuous bag-of-words model

Basic word representation#

Integers#

给词汇表每个单词分配一个唯一整数

• 简单

• Nunber按字母顺序排列，没有什么语义，如happy的值比hand大没有原因

  

One-hot vectors#

• 简单

• No implied ordering不暗示两个词有任何关系

• Huge vectors可能超过一百万行

  

• No embedded meaning 不包含单词意思，如无法计算相似度，happy应该更接近excited

​

Word embeddings#

Meaning as vectors#

可以发现happpy与excited更接近而不是paper

获得了一些语义但是也放弃了一些精确度，精确度不如one-hot，因为在此2D图中，两个单词可能位于同一点,如spider和snake

word embeddings#

• 维数低

• Embed meaning 有语义

  • e.g. semantic distance

    forest ≈tree forest≠ticket

  • e.g. analogies类比

    Paris:France::Rome:?

Terminology#

Summary#

• Words as integers
• Words as vectors.
  • One-hot vectors
  • Word embedding vectors
• Benefits of word embeddings for NLP

How to create word embedding#

Methods#

Basic word embedding methods#

• word2vec(Google,2013)
  • Continuous bag-of-words(CBOW) 根据给定的单词预测丢失的单词
  • Continuous skip-gram/Skip-gram with negative sampling(SGNS) 根据给定的词预测词
• Global Vectors(GloVe)(Stanford,2014)
• fast Text(Facebook,2016)
  • Supports out-of-vocabulary (OOV) words

Advanced word embedding methods#

Deep learning, contextual embeddings 单词根据context有不同的embedding，增加了多义词或相似词

• BERT(Google,2018)
• ELMo (Allen Institute for Al,2018)
• GPT-2(OpenAl,2018)

以上都 Tunable pre-trained models available

Continuous bag-of-words(CBOW) model#

基于周围单词得缺失单词

如果两个词周围的单词很像，那么他们可能语义相关

Create a training example#

C是模型超参，可改，将中心词加上上下文词称为window，此处window=1+2+2=5

From corpus to training#

e.g. 刚开始window内容为I am happy because I,然后window像右边滑动得am happy I am.

如此重复获得训练示例

CBOW in a nutshell#

Cleaning and tokenization#

Example in Python: corpus#

# pip install nltk
# pip instalL emoji

import nltk
from nltk.tokenize import word_tokenize
import emoji
import re

nltk.download('punkt')  # download pre-trained Punkt tokenizer for English
corpus = 'Who ♥"word embeddings" in 2020? I do!!!'
data = re.sub(r'[,!2;-]+','.', corpus)
data = nltk.word_tokenize(data)# tokenize string to words
data = [ ch.lower() for ch in data
         if ch. isalpha()
         or ch == '.'
         or emoji.get_emoji_regexp().search(ch)
         ]
print(data)
#['who', '♥', 'word', 'embeddings', 'in', 'i', 'do', '.']

#Sliding window of words
def get_windows(words,C):
    i=C
    while i<len(words)-C:
        center_word=words[i]
        context_words=words[(i-C):i]+words[(i+1):(i+C+1)]
        yield context_words,center_word
        i += 1

for x,y in get_windows(
			['i','am','happy','because','i','am','learning'],
    		2
		):
	print(f'{x}\t{y}')
#['i', 'am', 'because', 'i']	happy
#['am', 'happy', 'i', 'am']	because
#['happy', 'because', 'am', 'learning']	i

Transforming words into vectors#

Average of individual one-hot vectors#

Final prpared training set#

vector都是行向量，这样写是为了好看

Architecture of the CBOW model#

Demension(single input)#

行列向量 转置#

Dimensions (batch input)#

m : batch size 超参

组合成一个 V 行m 列的矩阵 称为 X

将y_hat分解为m个列向量

Rectified Linear Unit(ReLU)#

Softmax#

e.g.

Training a CBOW model#

Loss#

Corss-entropy loss#

常在分类中使用

$$J=-\sum_{k=1}^V y_k\log \hat y_k$$

$$J=-\log \hat y_{actual\ word}$$

forward propagation#

Training process

• forward propagation
• Cost
• Backpropagation and gradient descent

forward propagation

Cost#

Cost: mean of losses

$$J_{batch}=-\frac{1}{m}\sum_{i=1}^m\sum_{j=1}^Vy_j^{(i) }\log \hat y_j^{(i)}\\ J_{batch}=-\frac{1}{m}\sum_{i=1}^mJ^{(i)}$$

Minimizing the cost#

• Backpropagation:calculate partial derivatives of cost with respect to
  weights and biases

  $$\frac{\partial J_{batch}}{\partial W_1}, \frac{\partial J_{batch}}{\partial W_2}, \frac{\partial J_{batch}}{\partial b_1}, \frac{\partial J_{batch}}{\partial b_2}$$
  $$J_{batch}=f(W_1,W_2,b_1,b_2)$$

• Gradient descent:update weights and biases

Backpropagation#

Gradient descent#

Extracting word embedding vectors#

Option 1#

Option 2#

Option 3#

取前两个平均值

Evaluation word embedding#

Intrinsic evaluation#

Test relationships between words

• Analogies

  • Semantic analogies
    “France"is to"Paris"as"Italy"is to<?>

  • Syntactic analogies
    “seen"is to"saw"as"been"is to<?>

  • Ambiguity
    “wolf”is to"pack"as"bee"is to <?>→swarm? colony?

  • 在化学元素对应做的不好，word2vec给出：

    

• Clustering

  

• Visualization

  

Extrinsic evaluation#

Test word embeddings on external task

e.g. named entity recognition, parts-of-speech tagging

• Evaluates actual usefulness of embeddings
• 比内在评估更耗时
• More difficult to troubleshoot

Conclusion#

Recap and assignment

• Data preparation
• Word representations
• Continuous bag-of-words model
• Evaluation

Going further

• Advanced language modelling and word embeddings
• NLP and machine learning libraries