1. Question 1

What function creates one-hot encoded arrays of labels?

• ✅ tf.keras.utils.to_categorical

• ❌ tf.keras.utils.SequenceEnqueuer

• ❌ tf.keras.utils.img_to_array

• ❌ tf.keras.preprocessing.text.one_hot

Explanation:
to_categorical() converts integer labels into one-hot encoded vectors, which is required for multi-class classification.

---

2. Question 2

Major drawback of word-based training vs character-based?

• ❌ Character generation is more accurate

• ✅ Because there are far more words in a typical corpus than characters, it is much more memory intensive

• ❌ No drawback

• ❌ Word-based is more accurate

Explanation:
Word vocabularies can reach tens of thousands, making:
✔ memory heavier
✔ embeddings larger
✔ output layers huge
✔ training slower

Characters usually: 40–100 symbols only.

---

3. Question 3

Critical steps in preparing input sequences?

• ✅ Generating subphrases using n_gram_sequences

• ❌ Converting the seed text with texts_to_sequences

• ❌ Splitting into training/testing (not part of core sequence prep)

• ✅ Pre-padding the subphrase sequences

Correct Steps:

1. Create n-gram subphrases

2. Pre-pad sequences so they are equal length

Explanation:
These are essential for sequence prediction model training.

---

4. Question 4

Why does predicting more words cause gibberish?

• ❌ Probability compounds

• ❌ Matching fewer known phrases

• ❌ Likelihood doesn’t change

• ✅ Because you are more likely to hit words not in the training set

Explanation:
The model becomes less confident as predictions move forward → probability of selecting uncommon or unseen words increases, and small prediction errors accumulate → gibberish grows.

---

5. Question 5

Do we use a sigmoid output layer with one neuron per word?

• ❌ True

• ✅ False

Explanation:
For next-word prediction:
✔ Use Softmax
✔ One neuron per word
✔ Softmax gives a proper probability distribution across all words.

Sigmoid is for binary classification, not multi-class.

---

🧾 Summary Table

---