ai-engineering-from-scratch/phases/05-nlp-foundations-to-advanced/03-word-embeddings-word2vec/quiz.json at main · rohitg00/ai-engineering-from-scratch · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
{
  "lesson": "03-word-embeddings-word2vec",
  "title": "Word Embeddings — Word2Vec from Scratch",
  "questions": [
    {
      "stage": "pre",
      "question": "What is the distributional hypothesis?",
      "options": [
        "All words can be embedded in 300 dimensions",
        "Frequent words carry the most meaning",
        "You shall know a word by the company it keeps",
        "Words are uniformly distributed across documents"
      ],
      "correct": 2,
      "explanation": "Firth's (1957) distributional hypothesis: similar contexts imply similar meanings."
    },
    {
      "stage": "pre",
      "question": "Why is plain softmax over the vocabulary impractical in Word2Vec?",
      "options": [
        "Softmax cannot model probabilities",
        "Softmax has no gradient",
        "Computing softmax over 100k+ vocabulary terms is prohibitively expensive per training step",
        "It causes vanishing gradients"
      ],
      "correct": 2,
      "explanation": "Full-vocabulary softmax is too expensive; negative sampling reformulates it as binary classification."
    },
    {
      "stage": "check",
      "question": "What is the difference between skip-gram and CBOW?",
      "options": [
        "Skip-gram predicts context words from the center; CBOW predicts the center from context",
        "Skip-gram predicts the center word from context; CBOW does the reverse",
        "Skip-gram trains a deep network; CBOW is shallow",
        "Skip-gram uses bigrams; CBOW uses unigrams"
      ],
      "correct": 0,
      "explanation": "Skip-gram: center -> context. CBOW: context -> center."
    },
    {
      "stage": "check",
      "question": "In negative sampling, what is the objective for the positive pair (center, context)?",
      "options": [
        "Minimize their dot product",
        "Force them to be orthogonal",
        "Maximize sigmoid(W[center] dot W'[context]) so it is close to 1",
        "Decorrelate the vectors"
      ],
      "correct": 2,
      "explanation": "Positive pairs train sigmoid near 1; sampled negatives train sigmoid near 0."
    },
    {
      "stage": "check",
      "question": "After training Word2Vec, which weight matrix becomes the word embeddings?",
      "options": [
        "A separate post-training projection",
        "Both, multiplied together",
        "The hidden-to-output W' matrix",
        "The input-to-hidden W matrix (center-word table)"
      ],
      "correct": 3,
      "explanation": "The center-word table W is the standard embedding output; W' is often discarded or averaged in."
    },
    {
      "stage": "post",
      "question": "Why does Word2Vec fail on polysemy (e.g. 'bank')?",
      "options": [
        "Negative sampling drops rare meanings",
        "It ignores rare words",
        "The window size is too small",
        "It assigns one static vector per word, so 'river bank' and 'financial bank' share the same vector"
      ],
      "correct": 3,
      "explanation": "Static embeddings cannot disambiguate senses; contextual embeddings (ELMo/BERT) fix this."
    },
    {
      "stage": "post",
      "question": "When would you still reach for Word2Vec over a transformer in 2026?",
      "options": [
        "When dialog quality matters",
        "When you need contextual disambiguation",
        "Lightweight, on-device, domain-specific retrieval where a single row lookup is the latency budget",
        "When inputs are very long"
      ],
      "correct": 2,
      "explanation": "Word2Vec wins on tiny latency budgets, on-device inference, or fast domain-specific training."
    },
    {
      "stage": "post",
      "question": "What enables the famous analogy 'king - man + woman ~ queen'?",
      "options": [
        "Skip-gram trains on royal vocabularies",
        "Vector arithmetic captures linear directions like 'royal' that transfer across genders",
        "The model encodes royalty as a flag bit",
        "Cosine similarity is invariant to gender"
      ],
      "correct": 1,
      "explanation": "Directions in embedding space encode relational features, so adding/subtracting moves systematically."
    }
  ]
}