wayground logo

Free Printable Worksheets

Font size

S
M
L
XL
Worksheets

Natural Language Processing Quiz

Total questions: 126

Worksheet time: 1hrs 3mins

Name
Class
Date
1.

Identify an application of Natural Language Processing (NLP).

a)

Voice assistant

b)

Image segmentation

c)

Circuit simulation

d)

Network routing

2.

Distinguish between natural language and formal language in terms of their typical use.

a)

Natural language is used for human communication, while formal language is used in mathematics and programming.

b)

Both natural and formal languages are only used by computers.

c)

Formal language includes spoken and written communication between humans.

d)

Natural language is used for creating algorithms and data structures.

3.

A student is building a smart reply feature for an email application. The model needs to predict the next word by analyzing the current and previous word in the sentence. Choose the appropriate N-Gram model to meet this requirement.

a)

Bigram Model

b)

Unigram Model

c)

Decision Tree Classifier

d)

Support Vector Machine

4.

A language model developed for autocomplete is producing incomplete or blank suggestions because certain n-gram combinations are missing from the training data. Choose the appropriate technique that assigns non-zero probabilities to these unseen word sequences.

a)

Add-One (Laplace) Smoothing

b)

Data Normalization

c)

One-Hot Encoding

d)

TF-IDF Weighting

5.

A developer is evaluating the poor performance of a transformer model fine-tuned for next sentence prediction. On inspecting the input encoding, they discover that Type IDs for both sentences are incorrectly set to 1: Input: [CLS] The weather is nice today. [SEP] Let’s go for a walk. [SEP] Type IDs: 1 1 1 1 1 1 1 1 1 1 1 1 1 Analyse the impact of this Type ID configuration on model behavior.

a)

The model may incorrectly treat both sentences as part of the same segment, making it difficult to learn inter-sentence relationships.

b)

The model will apply a bidirectional attention mask, causing generation errors.

c)

The positional embeddings will overwrite the segment embeddings during fine-tuning.

d)

The model will use zero padding instead of segment embeddings, resulting in loss of context.

6.

A developer is preparing input for a transformer model to perform a next sentence prediction task. The input contains two sentences: Input: [CLS] AI is evolving fast [SEP] It impacts many industries [SEP] Calculate the token structure and identify the correct Position IDs and Type IDs for this sequence.

a)

Position IDs: 0 1 2 3 4 5 6 7 8 9; Type IDs: 0 0 0 0 0 1 1 1 1 1

b)

Position IDs: 1 2 3 4 5 6 7 8 9 10; Type IDs: all 0

c)

Position IDs: 0 1 2 3 4 5 6 7 8 9; Type IDs: 1 1 1 1 1 0 0 0 0 0

d)

Position IDs: 0 1 2 3 4 5 6 7 8 9; Type IDs: 0 0 0 0 0 0 1 1 1 0

7.

A student applies a 1-skip-2-gram model to the sentence: "Transformers understand complex patterns" They generate the following skip-gram pairs: (Transformers, understand), (understand, complex), (complex, patterns) Select the skip-gram output and identify the error in the student's approach.

a)

The student missed valid skip pairs like (Transformers, complex) and (understand, patterns), which should be included in a 1-skip model.

b)

The student incorrectly included pairs more than 2 positions apart, violating the n-gram rule.

c)

The student included punctuation tokens that should have been filtered out.

d)

The student used a bidirectional model instead of a unidirectional skip-gram model.

8.

Name the category of models that integrate NLP methods with a structured knowledge base to perform reasoning in specialized domains like medicine or law.

a)

Knowledge-based models

b)

Statistical models

c)

N-Gram models

d)

Recurrent neural networks

9.

Identify the programming language most commonly used for Natural Language Processing (NLP) tasks.

a)

Python

b)

Java

c)

C++

d)

R

10.

Recall the deep learning libraries commonly used for developing NLP models.

a)

TensorFlow and PyTorch

b)

NumPy and SciPy

c)

Hadoop and Spark

d)

Eclipse and NetBeans

11.

Differentiate between a feedforward neural language model and an RNN-based language model.

a)

Feedforward models accept fixed input size; RNNs handle variable-length input sequences.

b)

Both feedforward and RNN models process input with backward connections.

c)

RNNs use one-hot vectors, while feedforward models use continuous vectors only.

d)

Feedforward models are bidirectional by design, while RNNs are not.

12.

Identify the reason that causes a neural language model using only one preceding word as input to have limited accuracy.

a)

It lacks broader context to understand relationships among multiple words.

b)

It processes tokens in reverse order, leading to confusion.

c)

It randomly selects the next word, reducing prediction confidence.

d)

It requires recurrent feedback to work, which is unavailable.

13.

A student is tasked with developing a chatbot that can remember context across several messages. Choose the neural model architecture that would be most appropriate for this.

a)

Recurrent Neural Network

b)

N-Gram Model

c)

Feedforward Network

d)

Naive Bayes

14.

A data scientist wants to analyze and tokenize raw text data for further processing in a deep learning pipeline. Choose the NLP toolkit that should be used to handle pre-processing tasks.

a)

NLTK

b)

Scikit-learn

c)

NumPy

d)

Matplotlib

15.

During evaluation, a poor performance is obtained when the model processes longer sequences. Select the option that mitigates this.

a)

Switch to an RNN or LSTM architecture

b)

Apply dropout regularization

c)

Reduce training epochs

d)

Increase vocabulary size

16.

A user finds that a model trained on GPT generates off-topic text. Select the method that helps to improve its performance.

a)

Better prompt engineering

b)

Increasing hidden layers

c)

Using a smaller training set

d)

Switching to a rule-based system

17.

State the acronym of BERT.

a)

Bidirectional Encoder Representations from Transformers

b)

Basic Encoder Representations from Transformers

c)

Binary Encoding for Recurrent Transformers

d)

Bi-layered Encoder with Recurrent Transformers

18.

Differentiate the transformer model from previous models like LSTM.

a)

It uses attention mechanisms to analyze all words simultaneously

b)

It only works with labeled datasets

c)

It ignores the relationship between words

d)

It depends on hand-crafted features

19.

A developer wants to classify sentiment from customer reviews. Predict the model that would be most appropriate for understanding full sentence context.

a)

BERT

b)

GPT

c)

Naive Bayes

d)

Unigram

20.

Identify the reason that makes BERT generally more effective than GPT for sentiment classification tasks.

a)

BERT is bidirectional, allowing better understanding of sentence context.

b)

GPT uses outdated attention mechanisms.

c)

GPT lacks pre-training capabilities.

d)

BERT generates longer sequences, improving accuracy.

21.

State the AI model referred by Sam Altman when discussing about prompt engineering.

a)

ChatGPT

b)

Siri

c)

Bixby

d)

Alexa

22.

Identify the reason that makes prompt engineering to be considered as an essential skill in working with AI models like GPT-4.

a)

It allows for more relevant and coherent model outputs

b)

It helps reduce the size of language models

c)

It avoids the need for training data

d)

It eliminates human supervision completely

23.

A developer wants to guide DALL-E to create a futuristic cityscape. Choose the approach that should be taken by them.

a)

Provide clear, detailed prompt instructions specifying the desired elements

b)

Use generic phrases like "draw something"

c)

Rely on default generation without any input

d)

Ask the model to fetch training data

24.

A prompt engineer tests the same instruction across GPT-4 and PaLM but gets varied results. Select the reason for this variation.

a)

Different models interpret prompts based on their own training and architecture

b)

Prompt engineers must train each model manually

c)

All models share a fixed response library

d)

GPT-4 does not support prompt engineering

25.

A student wants to guide a language model to give answers in a political tone. Choose the LangChain class that should be used.

a)

SystemMessage

b)

HumanMessage

c)

PromptTemplate

d)

ChatOpenAI

26.

A student debugs a response issue in LangChain where user context is ignored. Incorrect message sequencing is suspected. Select the template that is most likely misconfigured.

a)

ChatPromptTemplate

b)

PromptTemplate

c)

JinjaTemplate

d)

OpenAIChain

27.

A developer notices repeated variable prompts during a session. They suspect poor prompt composition. Select the LangChain strategy that could improve this.

a)

Use a PipelinePromptTemplate to modularize prompt creation

b)

Replace LLM with a faster model

c)

Disable token streaming

d)

Use default temperature settings

28.

Identify the method that is used in LangChain to return a message after processing a list of messages.

a)

predict_messages

b)

process_input

c)

generate_message

d)

message_prompt

29.

Name the library that should be installed to use Jinja2 templates in LangChain.

a)

Jinja2

b)

Flask

c)

React

d)

Django

30.

State the primary cognitive mechanism in an LLMChain.

a)

LLM

b)

PromptTemplate

c)

OutputParser

d)

Tokenizer

31.

Select the role of PromptTemplate.from_template() function in LangChain.

a)

It creates a formatted prompt using a predefined structure.

b)

It evaluates model performance.

c)

It visualizes LLM output.

d)

It generates a new model.

32.

Identify the role of partial prompting in supporting LangChain workflows.

a)

It allows developers to fill in variables step-by-step before execution.

b)

It reduces the size of the final response.

c)

It encrypts prompts before sending them to the model.

d)

It limits token generation by default.

33.

A student is building a chatbot that replies with customized information based on user roles, instructions, and examples. Choose the most suitable template type.

a)

PipelinePromptTemplate

b)

SystemMessage

c)

PromptTemplate

d)

PartialPrompt

34.

Choose the class used to build an LLMChain that mimics Morpheus' responses using examples.

a)

FewShotPromptTemplate

b)

StaticPromptTemplate

c)

JSONPromptBuilder

d)

MessageFormatter

35.

A model's output is logical but lacks the conversational flow seen in earlier chats. Select a solution for this.

a)

Check whether ChatPromptTemplate is being used correctly

b)

Check if the OpenAI key is expired

c)

Check the number of function arguments in the script

d)

Check if the prompt ends with a period

36.

A developer's LLMChain outputs inconsistent answers despite a consistent prompt. Select the appropriate reason for this.

a)

The temperature parameter might be too high

b)

The chat model isn’t installed correctly

c)

The output parser is malfunctioning

d)

The prompt lacks placeholders

37.

Name the prompt class that defines a fixed string format in LangChain.

a)

StringPromptTemplate

b)

TextPrompt

c)

StaticPromptBuilder

d)

BasePromptFormat

38.

Identify the role of prefix in a FewShotPromptTemplate.

a)

It provides introductory context before examples

b)

It separates each example

c)

It finalizes the question-answer sequence

d)

It defines user input schema

39.

Choose the appropriate message format to create an interactive chatbot using ChatPromptTemplate.

a)

System and Human messages

b)

User and Assistant labels

c)

Input/output transformers

d)

DynamicPromptText

40.

Identify the use of .run() method of an LLMChain.

a)

Executes the model with the provided inputs and returns a response

b)

Trains the model with fine-tuning data

c)

Initializes all components in memory

d)

Parses the raw tokens from the user prompt

41.

What does BERT stand for?

a)

Bidirectional Encoder Representations from Transformers

b)

Basic Encoder Representations from Transformers

c)

Binary Encoding for Recurrent Transformers

d)

Bi-layered Encoder with Recurrent Transformers

42.

How does a transformer differ from LSTM?

a)

It uses attention mechanisms to analyze all words simultaneously.

b)

It only works with labeled datasets.

c)

It ignores word relationships.

d)

It depends on hand-crafted features.

43.

Why is BERT more effective than GPT for sentiment classification?

a)

BERT is bidirectional, allowing better understanding of context.

b)

GPT uses outdated attention mechanisms.

c)

GPT lacks pre-training.

d)

BERT generates longer sequences.

44.

Why is prompt engineering an essential skill in GPT-4?

a)

It allows for more relevant and coherent outputs.

b)

It reduces model size.

c)

It avoids training data.

d)

It eliminates supervision.

45.

A prompt engineer tests the same instruction across GPT-4 and PaLM but gets varied results. Why?

a)

Different models interpret prompts based on training and architecture.

b)

Prompt engineers must train each manually.

c)

All models share a fixed response library.

d)

GPT-4 doesn't support prompt engineering.

46.

Which statistical model predicts a continuous value such as salary from years of experience?

a)

Logistic Regression

b)

Linear Regression

c)

Naive Bayes

d)

Decision Tree

47.

Which of the following is an example of a unigram model?

a)

Predicting a word using the previous two words

b)

Predicting a word using only itself (current word)

c)

Predicting a word using the last three words

d)

Predicting a word with embeddings

48.

In an N-Gram model, what assumption allows prediction of the next word?

a)

Independence assumption

b)

Maximum likelihood assumption

c)

Markov assumption (limited history)

d)

Gradient assumption

49.

Which smoothing technique adds 1 to every count to avoid zero probability?

a)

Good-Turing

b)

Backoff

c)

Add-One (Laplace)

d)

Kneser-Ney

50.

Which smoothing redistributes probability mass from frequent to unseen n-grams?

a)

Add-k

b)

Good-Turing Discounting

c)

Laplace

d)

One-hot

51.

What is the advantage of N-Gram models?

a)

They capture deep semantics

b)

They handle long dependencies

c)

They are simple and fast to implement

d)

They require no training corpus

52.

What is the limitation of N-Gram models?

a)

Too easy to train

b)

They suffer from data sparsity

c)

They always overfit

d)

They ignore frequencies

53.

Which corpus is an example of data source for training NLP models?

a)

A random generator

b)

A large structured set of texts stored electronically

c)

Image dataset

d)

Audio recordings only

54.

In contextual models, which ID helps understand the order of tokens?

a)

Position ID

b)

Type ID

c)

Class ID

d)

Mask ID

55.

In contextual models, which ID distinguishes between different sentences?

a)

Position ID

b)

Type ID (Segment ID)

c)

Word ID

d)

Context ID

56.

Which contextual model developed by OpenAI uses a unidirectional transformer?

a)

GPT

b)

BERT

c)

RoBERTa

d)

ELMo

57.

Which model was trained on BooksCorpus and Wikipedia?

a)

GPT

b)

BERT

c)

Word2Vec

d)

TF-IDF

58.

Which contextual model uses BiLSTM architecture?

a)

BERT

b)

GPT

c)

ELMo

d)

RoBERTa

59.

Which contextual model is known as "Robust BERT"?

a)

ELMo

b)

GPT

c)

RoBERTa

d)

Word2Vec

60.

Which is an advantage of contextual models?

a)

They require little data

b)

They capture long-range dependencies

c)

They ignore polysemy

d)

They rely only on n-gram counts

61.

Which is a limitation of contextual models?

a)

Cannot handle polysemy

b)

Work only with small datasets

c)

Require large training data and compute resources

d)

Cannot be fine-tuned

62.

What is the key difference between static embeddings and contextual embeddings?

a)

Contextual embeddings ignore word order

b)

Static embeddings capture polysemy

c)

Contextual embeddings change meaning based on sentence context

d)

Static embeddings are always dense

63.

Which neural model represents the simplest neural language model?

a)

RNN

b)

Feedforward Neural Network

c)

Transformer

d)

Naive Bayes

64.

Which neural model can handle variable-length sequences effectively?

a)

RNN

b)

Feedforward

c)

Logistic Regression

d)

Decision Tree

65.

What improvement do feedforward models gain when using multiple input words?

a)

They become similar to fixed-size N-Gram models

b)

They capture infinite context

c)

They lose sequence meaning

d)

They ignore embeddings

66.

Which is a key limitation of feedforward networks in NLP?

a)

They always overfit

b)

They cannot accept variable-length inputs

c)

They cannot use embeddings

d)

They require convolutional filters

67.

Which language is most popular for NLP implementations?

a)

Java

b)

R

c)

C++

d)

Python

68.

Which library was one of the first NLP libraries written in Python?

a)

PyTorch

b)

TensorFlow

c)

NLTK

d)

Hugging Face

69.

Which platform provides 135+ state-of-the-art open-source NLP models?

a)

Scikit-learn

b)

Hugging Face

c)

Spark

d)

Eclipse

70.

Which deep learning libraries are most commonly used for building NLP models?

a)

TensorFlow and PyTorch

b)

NumPy and SciPy

c)

Hadoop and Spark

d)

MATLAB and R

71.

Which vectorization method represents each word as a binary sparse vector?

a)

One-hot encoding

b)

Word2Vec

c)

TF-IDF

d)

Embeddings

72.

What is the main disadvantage of one-hot encoding?

a)

It captures semantics

b)

It is low-dimensional

c)

It produces very high-dimensional sparse vectors

d)

It requires embeddings

73.

Which technique maps words into dense, low-dimensional vectors?

a)

TF-IDF

b)

Word Embeddings (Word2Vec, GloVe)

c)

One-hot encoding

d)

Bag-of-Words

74.

Which property makes embeddings superior to one-hot vectors?

a)

They are sparse

b)

They capture semantic similarity

c)

They ignore meaning

d)

They require no training

75.

Which vectorization method would place "king" and "queen" close in vector space?

a)

One-hot encoding

b)

Bag-of-Words

c)

Word Embedding

d)

TF-IDF

76.

Which is an example of tokenization?

a)

Assigning probabilities to words

b)

Splitting text into words like "Natural language processing is fun" → ["Natural", "language", ...]

c)

Creating embeddings

d)

Removing stop words

77.

Which library is best suited for preprocessing tasks like stemming and parsing?

a)

NumPy

b)

NLTK

c)

Matplotlib

d)

Spark

78.

Which model produces probabilities that sum to 1 across classes?

a)

Word2Vec

b)

One-hot

c)

Softmax

d)

LSTM

79.

Why is the Softmax function widely used in NLP classifiers?

a)

It tokenizes text

b)

It normalizes embeddings

c)

It converts logits into probability distributions

d)

It removes stop words

80.

Which statement best summarizes LSTM compared to RNNs?

a)

Both forget equally fast

b)

LSTM remembers important details longer than RNNs

c)

RNNs remember better

d)

LSTM is simpler than RNN

81.

Which mechanism enables Transformers to replace RNNs in NLP?

a)

Pooling

b)

Self-Attention

c)

Dropout

d)

Convolution

82.

Why are transformers better than RNNs for parallelization?

a)

They ignore context

b)

They process all tokens simultaneously

c)

They need fewer layers

d)

They use embeddings incorrectly

83.

Which embedding method is dynamic and changes depending on sentence meaning?

a)

Word2Vec

b)

Contextual embeddings (BERT, ELMo)

c)

One-hot encoding

d)

Bag-of-Words

84.

Why are contextual embeddings considered more powerful than static ones?

a)

They ignore context

b)

They require fewer parameters

c)

They reflect meaning changes depending on context

d)

They are always sparse

85.

Which technique reduces zero-probability issues in unseen N-Grams?

a)

Tokenization

b)

Smoothing

c)

Dropout

d)

Padding

86.

Which smoothing method adds a small constant instead of 1?

a)

Add-k Smoothing

b)

Laplace Smoothing

c)

Good-Turing

d)

Backoff

87.

Which advanced smoothing technique adjusts both counts and context diversity?

a)

Kneser-Ney Smoothing

b)

Add-One Smoothing

c)

Backoff

d)

Good-Turing

88.

Which model backs off to lower-order n-grams when higher ones are missing?

a)

Backoff Models

b)

Laplace Models

c)

Decision Trees

d)

Word2Vec

89.

What is a disadvantage of N-Gram models when n is large?

a)

They ignore context

b)

They require exponential memory growth

c)

They train too fast

d)

They never smooth

90.

Which is an example of a corpus in NLP?

a)

ImageNet

b)

Wikipedia Text Dump

c)

CIFAR-10

d)

MNIST

91.

What does a contextual model use to understand polysemy?

a)

Fixed embeddings

b)

Dynamic embeddings based on context

c)

Sparse vectors

d)

TF-IDF

92.

Which model first introduced the concept of contextual embeddings?

a)

GPT

b)

ELMo

c)

Word2Vec

d)

Logistic Regression

93.

Which model uses masked language modeling as a pretraining objective?

a)

GPT

b)

BERT

c)

Naive Bayes

d)

RNN

94.

Which model uses next-token prediction as its main training method?

a)

GPT

b)

BERT

c)

RoBERTa

d)

ELMo

95.

Which contextual model was trained on larger and more diverse corpora than BERT?

a)

RoBERTa

b)

Word2Vec

c)

N-Gram

d)

TF-IDF

96.

What is the primary limitation of contextual models like BERT?

a)

They ignore word order

b)

They cannot handle polysemy

c)

They are slower to train and require large data

d)

They rely only on frequency counts

97.

Which advantage do contextual models have over N-Gram models?

a)

Faster computation

b)

Ability to capture full-sentence meaning

c)

Require smaller data

d)

Rely only on counts

98.

Which component of a Transformer allows it to capture relationships between all tokens?

a)

Pooling

b)

Self-Attention

c)

One-hot encoding

d)

Dropout

99.

What is the role of positional encoding in Transformers?

a)

Reduces vocabulary size

b)

Normalizes embeddings

c)

Adds sequence order information

d)

Converts to probability

100.

Which type of embedding is static and does not change with context?

a)

Word2Vec

b)

BERT

c)

ELMo

d)

RoBERTa

101.

Which type of embedding changes with sentence usage?

a)

Word2Vec

b)

Contextual embeddings

c)

One-hot vectors

d)

Bag-of-Words

102.

Which challenge do RNNs face with long-term dependencies?

a)

Large memory usage

b)

Vanishing gradients

c)

Too many embeddings

d)

No activation functions

103.

What mechanism in LSTMs allows remembering information for longer?

a)

Softmax

b)

Memory cell with gates

c)

One-hot vectors

d)

Dropout

104.

Which gate in LSTMs prevents irrelevant information from being stored?

a)

Forget gate

b)

Output gate

c)

Input gate

d)

Embedding gate

105.

Which is a benefit of using LSTMs for sequence tasks?

a)

They ignore context

b)

They retain order and context in long sequences

c)

They require no embeddings

d)

They train without labels

106.

Which is an example where LSTM outperforms simple RNN?

a)

Sorting arrays

b)

Language translation

c)

Tokenization

d)

Word segmentation

107.

What is an example of an application using LSTM in text generation?

a)

Predicting the next word in a sentence

b)

Image segmentation

c)

Data compression

d)

Speech-to-text alignment

108.

Which advantage do LSTMs have in speech recognition?

a)

Faster inference

b)

Ability to remember context over time

c)

No need for training

d)

No embeddings required

109.

Which model architecture is designed to handle sequential dependencies more effectively than RNNs?

a)

CNN

b)

LSTM

c)

Logistic Regression

d)

Naive Bayes

110.

Identify which probability-based model is commonly used for text classification like spam filtering.

a)

Logistic Regression

b)

Naive Bayes

c)

Decision Trees

d)

KNN

111.

Compare Bag-of-Words with Word Embeddings. Which is true?

a)

Both preserve word order

b)

Bag-of-Words ignores semantics; embeddings capture meaning

c)

Embeddings are always sparse

d)

Bag-of-Words reduces dimensions

112.

Explain why TF-IDF is often preferred over raw frequency counts.

a)

It ignores rare words

b)

It reduces weight of common words and emphasizes important ones

c)

It produces dense vectors

d)

It uses contextual embeddings

113.

Apply the concept of trigram modeling. Which sequence represents a trigram?

a)

("The", "sky")

b)

("The", "sky", "is")

c)

("blue")

d)

("sky", "blue", "river")

114.

Analyse why smoothing is crucial in N-Gram models.

a)

It increases sequence length

b)

It removes tokens

c)

It avoids assigning zero probability to unseen n-grams

d)

It normalizes embeddings

115.

Identify which model is pre-trained on Common Crawl data.

a)

GPT

b)

BERT

c)

ELMo

d)

Naive Bayes

116.

Explain why contextual embeddings are useful for polysemous words.

a)

They ignore meaning

b)

They assign different vectors based on context

c)

They reduce vocabulary size

d)

They always use static vectors

117.

Apply positional encoding in Transformers. What does it achieve?

a)

Introduces order information into input tokens

b)

Normalizes embeddings

c)

Reduces computation

d)

Performs classification

118.

Compare GPT and BERT. Which is true?

a)

GPT is bidirectional; BERT is unidirectional

b)

Both are unidirectional

c)

GPT is unidirectional; BERT is bidirectional

d)

Both use CNNs

119.

Identify which model uses BiLSTMs to generate embeddings.

a)

GPT

b)

BERT

c)

ELMo

d)

RoBERTa

120.

Analyse the key reason why transformers outperform RNNs in scalability.

a)

They ignore dependencies

b)

They allow parallel processing of tokens

c)

They always train faster

d)

They use smaller datasets

121.

Apply the concept of attention. What does self-attention calculate?

a)

Dropout rates

b)

Relationships between each token and all others in the sequence

c)

Padding lengths

d)

Output probabilities only

122.

Evaluate why RoBERTa improves over BERT.

a)

It ignores embeddings

b)

It is trained on larger and more diverse corpora

c)

It eliminates attention layers

d)

It uses unigram probabilities

123.

Identify the main drawback of feedforward neural language models.

a)

They require embeddings

b)

They cannot handle variable sequence lengths

c)

They always overfit

d)

They ignore context

124.

Explain why RNNs are better than feedforward networks for sequences.

a)

They use convolution

b)

They can process variable-length inputs

c)

They always use embeddings

d)

They remove context

125.

Apply embeddings: which method ensures "Paris -- France + Italy ≈ Rome"?

a)

One-hot

b)

TF-IDF

c)

Word2Vec

d)

Bag-of-Words

126.

Compare one-hot encoding with embeddings. Which is correct?

a)

Both are dense

b)

One-hot captures semantics

c)

Embeddings capture semantic similarity; one-hot does not

d)

Both are sparse