{
  "lesson": "08-cnns-rnns-for-text",
  "title": "CNNs and RNNs for Text",
  "questions": [
    {
      "stage": "pre",
      "question": "What is a 1D convolutional filter of width 3 over word embeddings effectively learning?",
      "options": [
        "A learnable bigram detector",
        "A learnable trigram (n-gram) detector",
        "A position embedding",
        "A unigram count"
      ],
      "correct": 1,
      "explanation": "A width-3 filter spans three consecutive tokens, acting as a learnable trigram detector."
    },
    {
      "stage": "pre",
      "question": "Why does an LSTM use a cell state with mostly additive interactions?",
      "options": [
        "Faster softmax",
        "To avoid matrix inversion",
        "Additive flow lets gradients propagate through long sequences without vanishing or exploding",
        "It saves memory"
      ],
      "correct": 2,
      "explanation": "The cell-state highway keeps gradients stable across hundreds of steps, fixing vanishing-gradient issues."
    },
    {
      "stage": "check",
      "question": "Why does TextCNN use global max-pooling after the convolutional layer?",
      "options": [
        "To reduce vocabulary",
        "Max-pool gives a fixed-size, position-invariant representation by selecting the strongest activation per filter",
        "To enable backpropagation",
        "To remove embeddings"
      ],
      "correct": 1,
      "explanation": "Global max-pool produces a fixed-size feature vector regardless of input length, with position invariance."
    },
    {
      "stage": "check",
      "question": "What is the main advantage of a bidirectional RNN over a one-way RNN for sequence labeling?",
      "options": [
        "Cheaper memory",
        "Avoids embeddings",
        "Each token's representation sees both left and right context, which is essential for tagging",
        "Faster training"
      ],
      "correct": 2,
      "explanation": "Bidirectional networks concatenate forward + backward hidden states so labels can use full context."
    },
    {
      "stage": "check",
      "question": "Why does max-pooling over an LSTM's outputs often beat using just the last hidden state for classification?",
      "options": [
        "Max-pool removes capitalization",
        "Last-state pool ignores padding",
        "Max-pool is differentiable; last-state is not",
        "Information at the end of a long sequence tends to dominate the last state, hiding earlier evidence"
      ],
      "correct": 3,
      "explanation": "Max-pool aggregates strongest signal across positions; last-state can lose earlier evidence."
    },
    {
      "stage": "post",
      "question": "Which limitation of LSTM/RNN encoder-decoders motivated attention?",
      "options": [
        "They require subword tokenizers",
        "The decoder sees only a fixed-size encoder state, losing detail on long inputs; recurrence also serializes training",
        "They cannot embed tokens",
        "They cannot use embeddings"
      ],
      "correct": 1,
      "explanation": "Fixed-size summarization plus serial training were two failures attention removed."
    },
    {
      "stage": "post",
      "question": "When does a TextCNN or BiLSTM still beat a transformer in 2026?",
      "options": [
        "Whenever the dataset is multilingual",
        "Edge / on-device, streaming token-by-token inputs, or tiny-data baselines",
        "When latency requirements are extremely loose",
        "On image inputs"
      ],
      "correct": 1,
      "explanation": "Small architectures still win on edge inference, streaming inputs, and rapid baselines."
    },
    {
      "stage": "post",
      "question": "What is the vanishing gradient problem in plain RNNs?",
      "options": [
        "Inputs become zero",
        "Repeated multiplication by recurrent weights smaller than 1 makes gradients toward early steps shrink toward zero",
        "Loss becomes negative",
        "Embedding matrix becomes singular"
      ],
      "correct": 1,
      "explanation": "Long-product gradients vanish (or explode) without gating; this is why LSTMs and GRUs exist."
    }
  ]
}