1. 词向量上的操作(Operations on word vectors)

因为词嵌入的训练是非常耗资源的，所以ML从业者通常 都是 选择加载训练好 的 词嵌入(Embedding)数据集。(不用自己训练啦~~~)

任务：

• 导入 预训练词向量，使用余弦相似性(cosine similarity)计算相似度

• 使用词嵌入来解决 “Man is to Woman as King is to __.” 之类的 词语类比问题

• 修改词嵌入 来减少它们的性别歧视

import numpy as np
from w2v_utils import *

导入词向量，这个任务中，使用 50维的GloVe向量 来表示单词，导入 load the word_to_vec_map.

words, word_to_vec_map = read_glove_vecs('data/glove.6B.50d.txt')  # Embedding vector已知

print(list(words)[:10])
print(word_to_vec_map['mauzac'])

['1945gmt', 'mauzac', 'kambojas', '4-b', 'wakan', 'lorikeet', 'paratroops', 'wittkower', 'messageries', 'oliver']
[ 0.049225  -0.36274   -0.31555   -0.2424    -0.58761    0.277330.059622  -0.37908   -0.59505    0.78046    0.3348    -0.904010.7552    -0.30247    0.21053    0.03027    0.22069    0.406350.11387   -0.79478   -0.57738    0.14817    0.054704   0.973-0.22502    1.3677     0.14288    0.83708   -0.31258    0.25514-1.2681    -0.41173    0.0058966 -0.64135    0.32456   -0.84562-0.68853   -0.39517   -0.17035   -0.54659    0.014695   0.0736970.1433    -0.38125    0.22585   -0.70205    0.9841     0.19452-0.21459    0.65096  ]

导入的数据：

• words: 词汇表中单词集.

• word_to_vec_map: dictionary 映射单词到它们的 GloVe vector 表示.

Embedding vectors vs one-hot vectors

• one-hot向量不能很好捕捉单词之间的相似度水平(每一个one-hot向量与任何其他one-hot向量有相同的欧几里得距离(Euclidean distance))

• Embedding vector，如Glove vector提供了许多关于 单个单词含义 的有用信息

• 下面介绍如何使用 GloVe向量 来度量两个单词之间的 相似性

1.1 余弦相似度(Cosine similarity)

为了测量两个单词之间的相似性, 我们需要一个方法来测量两个单词的两个embedding vectors的相似性程度。 给定两个向量 \(u\) 和 \(v\), cosine similarity 定义如下：

\[\text{CosineSimilarity(u, v)} = \frac {u \cdot v} {||u||_2 ||v||_2} = cos(\theta) \tag{1} \]

• \(u \cdot v\) 是两个向量的点积(内积)

• \(||u||_2\) 向量 \(u\) 的范数(长度)

• \(\theta\) 是 \(u\) 与 \(v\) 之间的夹角角度

• 余弦相似性 依赖于 \(u\) and \(v\) 的角度.

  • 如果 \(u\) 和 \(v\) 很相似， 那么 \(cos(\theta)\) 越接近1.
  • 如果 \(u\) 和 \(v\) 不相似， 那么 \(cos(\theta)\) 得到一个很小的值.

**Figure 1**: The cosine of the angle between two vectors is a measure their similarity

Exercise： 实现函数 cosine_similarity() 来计算两个词向量之间的 相似性.

Reminder： \(u\) 的范式定义为 \(||u||_2 = \sqrt{\sum_{i=1}^{n} u_i^2}\)

提示： 使用 np.dot, np.sum, or np.sqrt 很有用.

# GRADED FUNCTION: cosine_similaritydef cosine_similarity(u, v):"""Cosine similarity reflects the degree of similarity between u and vArguments:u -- a word vector of shape (n,)          v -- a word vector of shape (n,)Returns:cosine_similarity -- the cosine similarity between u and v defined by the formula above."""distance = 0.0### START CODE HERE #### Compute the dot product between u and v (≈1 line)dot = np.sum(u * v)# Compute the L2 norm of u (≈1 line)norm_u = np.sqrt(np.sum(np.square(u)))# Compute the L2 norm of v (≈1 line)norm_v = np.sqrt(np.sum(np.square(v)))# Compute the cosine similarity defined by formula (1) (≈1 line)cosine_similarity = dot / (norm_u * norm_v)### END CODE HERE ###return cosine_similarity

测试：

father = word_to_vec_map["father"]
mother = word_to_vec_map["mother"]
ball = word_to_vec_map["ball"]
crocodile = word_to_vec_map["crocodile"]
france = word_to_vec_map["france"]
italy = word_to_vec_map["italy"]
paris = word_to_vec_map["paris"]
rome = word_to_vec_map["rome"]print("cosine_similarity(father, mother) = ", cosine_similarity(father, mother))
print("cosine_similarity(ball, crocodile) = ",cosine_similarity(ball, crocodile))
print("cosine_similarity(france - paris, rome - italy) = ",cosine_similarity(france - paris, rome - italy)) # (国家-首都, 首都-国家)-->接近-1
print("cosine_similarity(france - paris, italy - rome) = ",cosine_similarity(france - paris, italy - rome))

cosine_similarity(father, mother) = 0.890903844289
cosine_similarity(ball, crocodile) = 0.274392462614cosine_similarity(france - paris, rome - italy) = -0.675147930817cosine_similarity(france - paris, italy - rome) = 0.675147930817

随意的修改单词，查看他们相似性。

1.2 词类比工作(Word analogy task)

• 在词类比工作(word analogy task)中，我们完成句子："a is to b as c is to ____".

• 举例： 'man is to woman as king is to queen' .

• 我们尝试找到一个单词 d，使得相关的单词向量 \(e_a, e_b, e_c, e_d\) 以下列方式关联:\(e_b - e_a \approx e_d - e_c\)

• 我们将使用cosine similarity测量 \(e_b - e_a\) 和 \(e_d - e_c\) 的相似性.

Exercise：完成函数complete_analogy 实现 word analogies.

# GRADED FUNCTION: complete_analogydef complete_analogy(word_a, word_b, word_c, word_to_vec_map):"""Performs the word analogy task as explained above: a is to b as c is to ____. Arguments:word_a -- a word, stringword_b -- a word, stringword_c -- a word, stringword_to_vec_map -- dictionary that maps words to their corresponding vectors. Returns:best_word --  the word such that v_b - v_a is close to v_best_word - v_c, as measured by cosine similarity"""# convert words to lower caseword_a, word_b, word_c = word_a.lower(), word_b.lower(), word_c.lower()### START CODE HERE #### Get the word embeddings v_a, v_b and v_c (≈1-3 lines)e_a, e_b, e_c = word_to_vec_map[word_a],word_to_vec_map[word_b],word_to_vec_map[word_c]### END CODE HERE ###words = word_to_vec_map.keys()max_cosine_sim = -100              # Initialize max_cosine_sim to a large negative numberbest_word = None                   # Initialize best_word with None, it will help keep track of the word to output# loop over the whole word vector setfor w in words:        # to avoid best_word being one of the input words, pass on them.if w in [word_a, word_b, word_c] :continue### START CODE HERE #### Compute cosine similarity between the vector (e_b - e_a) and the vector ((w's vector representation) - e_c)  (≈1 line)cosine_sim = cosine_similarity(e_b - e_a,word_to_vec_map[w] - e_c)# If the cosine_sim is more than the max_cosine_sim seen so far,# then: set the new max_cosine_sim to the current cosine_sim and the best_word to the current word (≈3 lines)if cosine_sim > max_cosine_sim:max_cosine_sim = cosine_simbest_word = w### END CODE HERE ###return best_word

测试：

triads_to_try = [('italy', 'italian', 'spain'), ('india', 'delhi', 'japan'), ('man', 'woman', 'boy'), ('small', 'smaller', 'large')]
for triad in triads_to_try:print ('{} -> {} :: {} -> {}'.format( *triad, complete_analogy(*triad,word_to_vec_map)))

italy -> italian :: spain -> spanish
india -> delhi :: japan -> tokyo
man -> woman :: boy -> girl
small -> smaller :: large -> larger

也存在一些单词，算法不能给出正确答案：

triad = ['small', 'smaller', 'big']
print ('{} -> {} :: {} -> {}'.format( *triad, complete_analogy(*triad, word_to_vec_map)))

small -> smaller :: big -> competitors

1.3 总结

• Cosine similarity 求两个词向量的相似度不错

• 对于NLP应用，通常使用预训练好的词向量数据集

2. 除偏词向量(Debiasing word vectors)

在这一部分，我们将研究反映在词嵌入中的性别偏差，并试着去去除这一些偏差.

首先看一下 GloVe词嵌入如何关联性别的，你将计算一个向量 \(g = e_{woman}-e_{man}\)，\(e_{woman}\) 代表 woman 的词向量，\(e_{man}\)代表man的词向量，得到的结果 \(g\) 粗略的包含性别这一概念，计算 \(g_1 = e_{mother}-e_{father}\)，\(g_2 = e_{girl}-e_{boy}\) 的平均值可能会更准确点，现在使用 \(e_{woman}-e_{man}\) 足够了。

g = word_to_vec_map['woman'] - word_to_vec_map['man']    # 计算w与g的余弦相似度时，为正更接近女人，为负更接近男人
print(g)

[-0.087144    0.2182     -0.40986    -0.03922    -0.1032      0.94165-0.06042     0.32988     0.46144    -0.35962     0.31102    -0.868240.96006     0.01073     0.24337     0.08193    -1.02722    -0.211220.695044   -0.00222     0.29106     0.5053     -0.099454    0.404450.30181     0.1355     -0.0606     -0.07131    -0.19245    -0.06115-0.3204      0.07165    -0.13337    -0.25068714 -0.14293    -0.224957-0.149       0.048882    0.12191    -0.27362    -0.165476   -0.204260.54376    -0.271425   -0.10245    -0.32108     0.2516     -0.33455-0.04371     0.01258   ]

考虑不同单词与 \(g\) 的余弦相似度， 考虑一下正相似值 与 负余弦相似值的关系。

print ('List of names and their similarities with constructed vector:')# girls and boys name
name_list = ['john', 'marie', 'sophie', 'ronaldo', 'priya', 'rahul', 'danielle', 'reza', 'katy', 'yasmin']for w in name_list:print (w, cosine_similarity(word_to_vec_map[w], g))

List of names and their similarities with constructed vector:
john -0.23163356146
marie 0.315597935396
sophie 0.318687898594
ronaldo -0.312447968503
priya 0.17632041839
rahul -0.169154710392
danielle 0.243932992163
reza -0.079304296722
katy 0.283106865957
yasmin 0.233138577679

可以看到，女性的名字与 \(g\) 的余弦相似度为正，男性的名字与 \(g\) 的余弦相似度为负。

尝试其他

print('Other words and their similarities:')
word_list = ['lipstick', 'guns', 'science', 'arts', 'literature', 'warrior','doctor', 'tree', 'receptionist', 'technology',  'fashion', 'teacher', 'engineer', 'pilot', 'computer', 'singer']
for w in word_list:print (w, cosine_similarity(word_to_vec_map[w], g))

Other words and their similarities:
lipstick 0.276919162564
guns -0.18884855679
science -0.0608290654093
arts 0.00818931238588
literature 0.0647250443346
warrior -0.209201646411
doctor 0.118952894109
tree -0.0708939917548
receptionist 0.330779417506
technology -0.131937324476
fashion 0.0356389462577
teacher 0.179209234318
engineer -0.0803928049452
pilot 0.00107644989919
computer -0.103303588739
singer 0.185005181365

可以发现，比如“computer”就接近于“man”，“literature ”接近于“woman”，但是这些都是不对的一些观念，那么我们该如何减少这些偏差呢？

对于一些特殊的词汇而言，比如“男演员(actor)”与“女演员(actress)”或者“祖母(grandmother)”与“祖父(grandfather)”之间应该是具有性别差异的，而其他词汇比如“接待员(receptionist)”与“技术(technology )”是不应该有性别差异的，当我们处理这些词汇的时候应该区别对待。

2.1 中和非性别特定词汇的偏见(Neutralize bias for non-gender specific words)

下面的一张图表示了消除偏差之后的效果。如果我们使用的是50维的词嵌入，那么50维的空间可以分为两个部分： 偏置方向 \(g\)，和 剩下的49维 \(g_{\perp}\)。 在线性代数中，将 49维的 \(g_{\perp}\) 与 \(g\) 垂直(perpendicular)或正交("orthogonal")，即 \(g_{\perp}\) 与 \(g\) 成90°角 。在中和步骤中，取一个向量，如 \(e_{receptionist}\)向量， 将 \(g\) 方向的组成 归零，得到 \(e_{receptionist}^{debiased}\) 向量。

即使 \(g_{\perp}\) 是 49维， 鉴于只能在2D屏幕上绘制图像的局限性，我们使用下面的一维坐标轴来说明它。

**Figure 2**: The word vector for "receptionist" represented before and after applying the neutralize operation.

Exercise: 实现 neutralize() 来消除词汇 的偏见，如 "receptionist" 和 "scientist"。 给定一个词嵌入输入\(e\)，可以使用下面的公式计算 \(e^{debiased}\):

\[e^{bias\_component} = \frac{e \cdot g}{||g||_2^2} * g\tag{2} \]

\[e^{debiased} = e - e^{bias\_component}\tag{3} \]

\(e^{bias\_component}\) 是 \(e\) 在 \(g\) 方向的投影。

def neutralize(word, g, word_to_vec_map):"""Removes the bias of "word" by projecting it on the space orthogonal to the bias axis. This function ensures that gender neutral words are zero in the gender subspace.Arguments:word -- string indicating the word to debiasg -- numpy-array of shape (50,), corresponding to the bias axis (such as gender)word_to_vec_map -- dictionary mapping words to their corresponding vectors.Returns:e_debiased -- neutralized word vector representation of the input "word""""### START CODE HERE #### Select word vector representation of "word". Use word_to_vec_map. (≈ 1 line)e = word_to_vec_map[word]# Compute e_biascomponent using the formula give above. (≈ 1 line)e_biascomponent = np.divide(np.sum(e * g), np.square(np.linalg.norm(g))) * g# Neutralize e by substracting e_biascomponent from it # e_debiased should be equal to its orthogonal projection. (≈ 1 line)e_debiased = e - e_biascomponent### END CODE HERE ###return e_debiased

测试：

e = "receptionist"
print("cosine similarity between " + e + " and g, before neutralizing: ", cosine_similarity(word_to_vec_map["receptionist"], g))e_debiased = neutralize("receptionist", g, word_to_vec_map)
print("cosine similarity between " + e + " and g, after neutralizing: ", cosine_similarity(e_debiased, g))

cosine similarity between receptionist and g, before neutralizing: 0.330779417506
cosine similarity between receptionist and g, after neutralizing: -5.60374039375e-17

2.2 性别词的均衡算法(Equalization algorithm for gender-specific words)

接下来我们来看看在关于有特定性别词组中，如何将它们进行均衡，比如“男演员”与“女演员”中，与“保姆”一词更接近的是“女演员”，我们可以消去“保姆”的性别偏差，但是这并不能保证“保姆”一词与“男演员”与“女演员”之间的距离相等，我们要学的均衡算法将解决这个问题。

均衡(equalization)关键思想：确保 一对特定的单词 与 49维 \(g_\perp\) 距离相等。均衡步骤还确保 两个被均衡的step 现在与 \(e_{receptionist}^{debiased}\) 或 任何已被中和的其他工作的距离相同。下面展示equalization如何工作：

线性代数的推导过程很复杂(See Bolukbasi et al., 2016 for details.) 关键的公式：

\[\mu = \frac{e_{w1} + e_{w2}}{2}\tag{4} \]

\[ \mu_{B} = \frac {\mu \cdot \text{bias_axis}}{||\text{bias_axis}||_2^2} *\text{bias_axis} \tag{5}\]

\[\mu_{\perp} = \mu - \mu_{B} \tag{6} \]

\[ e_{w1B} = \frac {e_{w1} \cdot \text{bias_axis}}{||\text{bias_axis}||_2^2} *\text{bias_axis} \tag{7}\]

\[ e_{w2B} = \frac {e_{w2} \cdot \text{bias_axis}}{||\text{bias_axis}||_2^2} *\text{bias_axis} \tag{8}\]

\[e_{w1B}^{corrected} = \sqrt{ |{1 - ||\mu_{\perp} ||^2_2} |} * \frac{e_{\text{w1B}} - \mu_B} {|(e_{w1} - \mu_{\perp}) - \mu_B|} \tag{9} \]

\[e_{w2B}^{corrected} = \sqrt{ |{1 - ||\mu_{\perp} ||^2_2} |} * \frac{e_{\text{w2B}} - \mu_B} {|(e_{w2} - \mu_{\perp}) - \mu_B|} \tag{10} \]

\[e_1 = e_{w1B}^{corrected} + \mu_{\perp} \tag{11} \]

\[e_2 = e_{w2B}^{corrected} + \mu_{\perp} \tag{12} \]

Exercise: 使用这些上述公式 得到这对单词的最终均衡版本.

def equalize(pair, bias_axis, word_to_vec_map):"""Debias gender specific words by following the equalize method described in the figure above.Arguments:pair -- pair of strings of gender specific words to debias, e.g. ("actress", "actor") bias_axis -- numpy-array of shape (50,), vector corresponding to the bias axis, e.g. genderword_to_vec_map -- dictionary mapping words to their corresponding vectorsReturnse_1 -- word vector corresponding to the first worde_2 -- word vector corresponding to the second word"""### START CODE HERE #### Step 1: Select word vector representation of "word". Use word_to_vec_map. (≈ 2 lines)w1, w2 = paire_w1, e_w2 = word_to_vec_map[w1],word_to_vec_map[w2]# Step 2: Compute the mean of e_w1 and e_w2 (≈ 1 line)mu = (e_w1 + e_w2) / 2# Step 3: Compute the projections of mu over the bias axis and the orthogonal axis (≈ 2 lines)mu_B = np.divide(np.dot(mu, bias_axis) * bias_axis, np.sum(bias_axis**2))mu_orth = mu - mu_B# Step 4: Use equations (7) and (8) to compute e_w1B and e_w2B (≈2 lines)e_w1B = (np.dot(e_w1,bias_axis) * bias_axis) / (np.sum(bias_axis**2))   e_w2B = (np.dot(e_w2,bias_axis) * bias_axis) / (np.sum(bias_axis**2))# Step 5: Adjust the Bias part of e_w1B and e_w2B using the formulas (9) and (10) given above (≈2 lines)corrected_e_w1B = (np.sqrt(np.abs(1 - (np.sum(mu_orth**2))))) * np.divide(e_w1B - mu_B, np.abs((e_w1 - mu_orth) - mu_B))corrected_e_w2B = (np.sqrt(np.abs(1 - (np.sum(mu_orth**2))))) * np.divide(e_w2B - mu_B, np.abs((e_w2 - mu_orth) - mu_B))# Step 6: Debias by equalizing e1 and e2 to the sum of their corrected projections (≈2 lines)e1 = corrected_e_w1B + mu_orthe2 = corrected_e_w2B + mu_orth### END CODE HERE ###return e1, e2

测试：

print("cosine similarities before equalizing:")
print("cosine_similarity(word_to_vec_map[\"man\"], gender) = ", cosine_similarity(word_to_vec_map["man"], g))
print("cosine_similarity(word_to_vec_map[\"woman\"], gender) = ", cosine_similarity(word_to_vec_map["woman"], g))
print()
e1, e2 = equalize(("man", "woman"), g, word_to_vec_map)
print("cosine similarities after equalizing:")
print("cosine_similarity(e1, gender) = ", cosine_similarity(e1, g))
print("cosine_similarity(e2, gender) = ", cosine_similarity(e2, g))

cosine similarities before equalizing:
cosine_similarity(word_to_vec_map["man"], gender) =  -0.117110957653
cosine_similarity(word_to_vec_map["woman"], gender) =  0.356666188463cosine similarities after equalizing:
cosine_similarity(e1, gender) =  -0.716572752584
cosine_similarity(e2, gender) =  0.739659647493

Debiasing algorithms 减少偏见很有用，但并不完美，不能消除所有的偏置痕迹

• 如：这种实现的一个缺点： \(g\) 的偏差方向 只使用 单词对 woman 和 man 来定义。如果通过计算 \(g_1 = e_{woman} - e_{man}\)，\(g_2 = e_{mother} - e_{father}\)，\(g_3 = e_{girl} - e_{boy}\)，然后使用 \(g = avg(g_1, g_2, g_3)\) 来计算他它们平均，你将更好的估计50维词嵌入空间中的 "gender" 维度。