1. Question 1

When stacking LSTMs, how do you feed the next LSTM in the sequence?

• ❌ Ensure same number of units

• ❌ Do nothing

• ❌ return_sequences = True on all

• ✅ return_sequences = True only on units that feed another LSTM

Explanation:
Only LSTM layers feeding into another LSTM must output a sequence → set return_sequences=True.

---

2. Question 2

How does an LSTM capture long-range meaning?

• ❌ Load all words into cell state

• ❌ They don’t

• ✅ Earlier word values can be carried forward via the cell state

• ❌ Shuffle words randomly

Explanation:
LSTMs maintain memory through a cell state that flows across time steps.

---

3. Question 3

Best way to avoid overfitting in NLP datasets?

• ❌ LSTMs

• ❌ GRUs

• ❌ Conv1D

• ✅ None of the above

Explanation:
Overfitting isn’t solved by architecture choice → use regularization, dropout, augmentation, more data.

---

4. Question 4

Which Keras layer allows LSTMs to read forward and backward?

• ❌ Bilateral

• ❌ Unilateral

• ❌ Bothdirection

• ✅ Bidirectional

Explanation:
Bidirectional() wraps an LSTM so it processes the sequence forward and backward.

---

5. Question 5

Why does sequence matter for semantics?

• ❌ It doesn’t

• ❌ Because order dictates their impact on meaning

• ❌ Order doesn’t matter

• ✅ Because the order in which words appear dictates their meaning

Explanation:
Word order changes meaning entirely:
“dog bites man” ≠ “man bites dog”.

---

6. Question 6

How do RNNs help understand sequence meaning?

• ❌ They don’t

• ❌ They shuffle the words

• ❌ They look at whole sentence at once

• ✅ They carry meaning from one cell to the next

Explanation:
RNNs pass hidden states between time steps → sequence-aware understanding.

---

7. Question 7

Output shape of a bidirectional LSTM with 64 units?

• ❌ (128,1)

• ❌ (128,None)

• ❌ (None,64)

• ✅ (None,128)

Explanation:
Bidirectional doubles the units:
64 forward + 64 backward = 128.

---

8. Question 8

Sentence = 120 tokens → Conv1D with 128 filters, kernel size = 5.
What is output shape?

• ❌ (None, 120, 124)

• ✅ (None, 116, 128)

• ❌ (None, 116, 124)

• ❌ (None, 120, 128)

Explanation:
Conv1D output length (no padding):
120 − 5 + 1 = 116
Filters = 128 channels.

---

🧾 Summary Table

---