UNIT- 4
MCQs Suggestions Questions
1. Natural Language Processing with Sequence Models and Neural Networks for
Sentiment Analysis
1. In NLP, a sequence model aims to model dependencies between:
A. Individual letters
B. Entire documents
C. Sequential words
D. Sentiment scores
Answer: C. Sequential words
2. Which type of neural network is best suited for sentiment analysis in text?
A. Convolutional Neural Network (CNN)
B. Recurrent Neural Network (RNN)
C. Fully Connected Network
D. Decision Tree
Answer: B. Recurrent Neural Network (RNN)
3. In a binary sentiment classification task, the output layer typically has:
A. 1 unit
B. 2 units
C. 3 units
D. Variable units depending on input size
Answer: A. 1 unit
4. Which activation function is commonly used in the output layer for sentiment
classification tasks?
A. ReLU
B. Tanh
C. Sigmoid
D. Softmax
Answer: C. Sigmoid
5. Which of the following algorithms is commonly used to train neural networks in
sentiment analysis tasks?
A. Backpropagation
B. Forward propagation
C. Decision Trees
D. Linear Regression
Answer: A. Backpropagation
6. The vanishing gradient problem often occurs in RNNs when:
A. Gradients get too large
B. Gradients become too small over time
C. The model has too few layers
D. We use batch normalization
Answer: B. Gradients become too small over time
7. Sentiment analysis often uses which type of loss function?
A. Mean Squared Error
B. Cross-Entropy Loss
C. Hinge Loss
D. KL Divergence
Answer: B. Cross-Entropy Loss
8. Which of the following embedding techniques is used to convert words into vector
representations for NLP tasks?
A. One-hot encoding
B. Word2Vec
C. Binary encoding
D. Frequency encoding
Answer: B. Word2Vec
9. What is the purpose of using padding in sequence models for sentiment analysis?
A. To introduce randomness
B. To equalize sequence lengths
C. To improve model performance
D. To make predictions accurate
Answer: B. To equalize sequence lengths
10. In sentiment analysis, a positive sentiment is typically represented by:
A. The number 0
B. The number 1
C. -1
D. Any non-zero value
Answer: B. The number 1
2. Recurrent Neural Networks for Language Modeling
11. Language modeling aims to:
A. Generate new sentences
B. Predict the next word in a sequence
C. Classify text sentiment
D. Summarize text
Answer: B. Predict the next word in a sequence
12. Which of the following architectures is commonly used for language modeling?
A. CNN
B. RNN
C. SVM
D. KNN
Answer: B. RNN
13. The hidden state in RNNs is:
A. A random variable
B. A static vector
C. Updated at each time step
D. Irrelevant for predictions
Answer: C. Updated at each time step
14. GRU is an RNN variant that helps with:
A. Handling images
B. Decreasing training time
C. Addressing the vanishing gradient problem
D. Increasing the number of parameters
Answer: C. Addressing the vanishing gradient problem
15. The formula ht=σ(W⋅xt+U⋅ht−1+b)h_t = \sigma(W \cdot x_t + U \cdot h_{t-1} + b)ht
=σ(W⋅xt+U⋅ht−1+b) in RNNs computes:
A. Output vector
B. Input sequence
C. Next hidden state
D. Initial state
Answer: C. Next hidden state
16. In the context of RNNs, "sequential data" refers to:
A. Randomly arranged data points
B. Data with temporal or ordered structure
C. Independent data points
D. Unstructured text
Answer: B. Data with temporal or ordered structure
17. Which gate in an LSTM cell helps retain long-term dependencies?
A. Input gate
B. Output gate
C. Forget gate
D. Memory gate
Answer: C. Forget gate
18. GRU differs from LSTM in that it:
A. Uses fewer gates
B. Has more memory capacity
C. Is slower to train
D. Requires more data
Answer: A. Uses fewer gates
19. The purpose of a next-word generator model is to:
A. Classify documents
B. Predict the subsequent word in a text sequence
C. Detect sentiment
D. Label entities in text
Answer: B. Predict the subsequent word in a text sequence
20. In NLP, a "language model" assigns a probability to a:
A. Sentence or sequence of words
B. Single character
C. Random number
D. Topic label
Answer: A. Sentence or sequence of words
3. Named Entity Recognition (NER) Using LSTMs
21. Named Entity Recognition (NER) is the process of identifying:
A. The sentiment of a text
B. Specific entities like names, locations, and dates in text
C. The language of a document
D. Keywords in a document
Answer: B. Specific entities like names, locations, and dates in text
22. Which neural network component in LSTM prevents the vanishing gradient
problem?
A. Tanh activation
B. Cell state
C. Bias
D. Learning rate
Answer: B. Cell state
23. The purpose of an LSTM’s output gate is to:
A. Remove irrelevant information
B. Store the cell state
C. Control the output at each time step
D. Initialize hidden states
Answer: C. Control the output at each time step
24. An LSTM cell’s forget gate decides:
A. What information to discard
B. Which information to add
C. The final output
D. The learning rate
Answer: A. What information to discard
25. NER models using LSTMs are popular because they can:
A. Process data quickly
B. Retain contextual information over long sequences
C. Require no training
D. Be trained with minimal data
Answer: B. Retain contextual information over long sequences
26. In an NER system, "John" in "John bought a car" is classified as:
A. Organization
B. Location
C. Person
D. Date
Answer: C. Person
27. A Named Entity Recognition system using an LSTM generally outputs:
A. A probability distribution over possible entities
B. A single label for the whole sequence
C. A random guess
D. Only positive sentiment
Answer: A. A probability distribution over possible entities
28 Which model is often used for Named Entity Recognition?
A. Decision Trees
B. Naive Bayes
C. LSTM
D. CNN
Answer: C. LSTM
29 In a language model, "bi-LSTM" refers to an LSTM that:
A. Processes data only forwards
B. Processes data forwards and backwards
C. Ignores the input sequence
D. Requires more memory
Answer: B. Processes data forwards and backwards
30. In LSTMs, the "cell state" is responsible for:
A. Storing short-term information
B. Maintaining long-term dependencies
C. Generating random data
D. Decreasing model accuracy
Answer: B. Maintaining long-term dependencies
31. In a simple RNN, the hidden state at each time step is calculated based on:
A. Only the current input
B. The previous hidden state and the current input
C. The next hidden state
D. Only the previous hidden state
Answer: B. The previous hidden state and the current input
32. The main limitation of an RNN is the issue of:
A. High accuracy
B. Vanishing and exploding gradients
C. Low computational cost
D. High interpretability
Answer: B. Vanishing and exploding gradients
33. In an RNN, the activation function used to calculate the hidden state is typically:
A. Softmax
B. ReLU
C. Tanh or Sigmoid
D. Linear
Answer: C. Tanh or Sigmoid
34. The vanishing gradient problem in RNNs is primarily caused by:
A. Short input sequences
B. Constant gradients
C. Small gradients diminishing during backpropagation
D. A high learning rate
Answer: C. Small gradients diminishing during backpropagation
35. An RNN’s performance on long sequences can be improved using:
A. More hidden layers
B. Batch normalization
C. LSTM or GRU units
D. Higher learning rates
Answer: C. LSTM or GRU units
2. Long Short-Term Memory (LSTM) Networks
6. In an LSTM cell, the forget gate is responsible for:
A. Deciding what information to discard from the cell state
B. Adding new information to the cell state
C. Generating the final output
D. Controlling the learning rate
Answer: A. Deciding what information to discard from the cell state
7. The input gate in an LSTM:
A. Controls the current hidden state
B. Decides what new information to store in the cell state
C. Ignores the cell state
D. Controls gradient flow
Answer: B. Decides what new information to store in the cell state
8. Which activation function is typically used in the forget, input, and output gates of
an LSTM?
A. ReLU
B. Softmax
C. Sigmoid
D. Tanh
Answer: C. Sigmoid
9. The cell state in an LSTM helps to:
A. Track short-term dependencies
B. Maintain long-term dependencies
C. Initialize the model
D. Update the learning rate
Answer: B. Maintain long-term dependencies
10. The output gate in an LSTM controls:
A. The information to be output from the cell state
B. The information to be added to the cell state
C. The learning rate
D. The initialization of weights
Answer: A. The information to be output from the cell state
11. Which component in an LSTM cell is responsible for addressing the vanishing
gradient problem?
A. Forget gate
B. Output gate
C. Cell state
D. Dropout
Answer: C. Cell state
3. Gated Recurrent Unit (GRU) Networks
16. GRU is a variant of RNNs designed to address:
A. Image processing issues
B. The vanishing gradient problem
C. Increase in model complexity
D. Sequential data sorting
Answer: B. The vanishing gradient problem
17. Which gates are present in a GRU?
A. Forget and Input gates
B. Reset and Update gates
C. Output and Input gates
D. Memory and Forget gates
Answer: B. Reset and Update gates
18. The purpose of the reset gate in a GRU is to:
A. Control the amount of past information to forget
B. Update the learning rate
C. Compute the output
D. Initialize the model
Answer: A. Control the amount of past information to forget
19. The update gate in a GRU decides:
A. Whether to forget the previous hidden state
B. How much of the previous hidden state to carry forward
C. The learning rate
D. The training batch size
Answer: B. How much of the previous hidden state to carry forward
20. Compared to an LSTM, a GRU has:
A. More gates
B. Fewer gates
C. No cell state
D. Both B and C
Answer: D. Both B and C
21. Which of the following is a key advantage of GRU over LSTM?
A. Less computationally intensive
B. Higher memory requirements
C. Better accuracy on image data
D. Lower prediction accuracy
Answer: A. Less computationally intensive
22. In GRUs, the hidden state hth_tht is calculated using:
ht=(1−zt)∗ht−1+zt∗ht~
Here, tzt is the:
A. Reset gate output
B. Update gate output
C. Forget gate output
D. Learning rate
Answer: B. Update gate output
23. The candidate hidden state ht~\tilde{h_t}ht~ in GRUs is influenced by which gate?
A. Reset gate
B. Update gate
C. Output gate
D. Forget gate
Answer: A. Reset gate
24. Compared to a simple RNN, the GRU can:
A. Better capture long-term dependencies
B. Reduce the need for regularization
C. Lower the vanishing gradient effect
D. Both A and C
Answer: D. Both A and C