✅

📌
NLP
word2vec, cbow, Bagof words, skip-gram related mcq with answer , easy to
midium with answer analysis 40 mcq cover main knowledge, some with
situation base question, all pros cons

Here you go — 40 MCQs on Word2Vec, CBOW, Skip-gram, Bag-of-Words

(BoW) with answers + explanations, including situational questions, pros/
cons knowledge, and easy → medium difficulty.

40 MCQs on Word2Vec, CBOW, Skip-gram, BoW (with

answers & explanations)

Section A — Concept Basics (Easy)

1. What does Bag-of-Words (BoW) primarily ignore?

A. Word counts
B. Grammar
C. Vocabulary size
D. Tokenization
Answer: B
Explanation: BoW ignores grammar and order of words.

2. In BoW, what happens when vocabulary size increases?

A. Accuracy reduces
B. Vector size increases
C. Model becomes faster
D. More semantic info
Answer: B
Explanation: BoW creates a vector the size of vocabulary → sparse and high
dimensional.

3. Word2Vec produces outputs as:

A. One-hot vectors
B. Count vectors
C. Dense distributed vectors
D. TF-IDF vectors
Answer: C

4. Which Word2Vec model predicts target word from the context

words?
A. Skip-gram
B. CBOW
C. BoW
D. TF-IDF
Answer: B
Explanation: CBOW → Context → Target
5. Skip-gram predicts:
A. Target from context
B. Context from target
C. Bag-of-words
D. Probabilities only
Answer: B
Explanation: Target → Context

6. Which method is better for small datasets?

A. CBOW
B. Skip-gram
C. BoW
D. TF-IDF
Answer: B
Explanation: Skip-gram works well with smaller datasets and rare words.

7. In Word2Vec, vector dimension is typically:

A. 5–10
B. 100–300
C. 2000–5000
D. Vocabulary size
Answer: B

8. Word2Vec achieves semantic relationships using:

A. Co-occurrence matrices
B. Window-based prediction
C. Frequency counts
D. Stop-word removal
Answer: B

9. “King – Man + Woman = Queen” is an example of:

A. One-hot encoding
B. Vector arithmetic
C. TF-IDF
D. Skip-gram failure
Answer: B

10. What is a major disadvantage of BoW?

A. Simple to implement
B. Requires large memory
C. Supports semantic meaning
D. Good for small data
Answer: B
 Section B — Word2Vec Training Concepts (Medium)

11. Negative Sampling is used to:

A. Increase vocabulary
B. Reduce training time
C. Remove stop words
D. Improve grammar
Answer: B
Explanation: It reduces expensive softmax computation.

12. Hierarchical softmax converts probability distribution into:

A. Flat tree
B. Binary tree
C. Sparse matrix
D. One-hot encoding
Answer: B

13. CBOW is generally:

A. Faster
B. Slower
C. More accurate for rare words
D. More memory-heavy
Answer: A

14. Skip-gram is better for:

A. Rare words
B. Frequent words
C. Stop words
D. Random sampling
Answer: A

15. Word2Vec is trained using:

A. Supervised learning
B. Semi-supervised
C. Unsupervised learning
D. Reinforcement learning
Answer: C
Explanation: No labels required; learns from sequences.

16. What is the window size used for?

A. Learning grammar
B. Selecting context words
C. Choosing vector dimension
D. Assigning labels
Answer: B
17. Which one is NOT a Word2Vec training technique?
A. Skip-gram
B. CBOW
C. Negative sampling
D. Hashing Trick
Answer: D

18. Word embeddings are:

A. Sparse
B. Dense
C. Large
D. Zero vectors
Answer: B

19. Word2Vec learns word meaning based on:

A. Syntax rules
B. Co-occurrence patterns
C. Part-of-speech
D. Lemmatization
Answer: B

20. A key advantage of Word2Vec over BoW is:

A. Simpler model
B. Captures meaning
C. Smaller vocabulary
D. Fast with huge datasets
Answer: B

Section C — Situation Based MCQs

21. You have a small dataset but want to capture meaning of rare
words. Which model?
A. TF-IDF
B. CBOW
C. Skip-gram
D. BoW
Answer: C

22. You need high training speed; accuracy for frequent words is
enough. Choose—
A. Skip-gram
B. CBOW
C. BoW
D. N-gram model
Answer: B

23. You want a model that represents sentence length doesn’t

matter.
A. Word2Vec
B. CBOW
C. BoW
D. Skip-gram
Answer: C
Explanation: BoW loses order, length, and structure.

24. A model must identify synonyms well. Best choice—

A. BoW
B. One-hot
C. Word2Vec
D. TF-IDF
Answer: C

25. You have 10 million training words. Which model is too slow?
A. CBOW
B. BoW
C. Skip-gram
D. Negative sampling version
Answer: C
Explanation: Skip-gram is slow for huge datasets.

26. You want to reduce sparse features. Which method helps?

A. BoW
B. Word2Vec
C. TF-IDF
D. N-grams
Answer: B

27. Text classification simple baseline:

A. Word2Vec
B. BoW
C. Skip-gram
D. CBOW
Answer: B

28. Which gives fixed-length vectors irrespective of vocabulary

growth?
A. BoW
B. Word2Vec
C. TF-IDF
D. N-gram
Answer: B

29. Embeddings fail when—

A. Polysemy words like "bank"
B. Rare words
C. Small window size
D. Corpus is too large
Answer: A
Explanation: Classic Word2Vec → one vector per word, not context-dependent.

30. You build a search engine and want semantic matching of

queries.
A. Bag-of-words
B. One-hot
C. Word2Vec
D. Letter-level encoding
Answer: C

Section D — Pros/Cons Awareness MCQs

31. Major disadvantage of Skip-gram?

A. Computationally slow
B. Cannot learn semantics
C. No word window
D. Sparse vectors
Answer: A

32. Disadvantage of CBOW?

A. Slow training
B. Ignores rare words
C. Hard to scale
D. Uses dense vectors
Answer: B

33. Disadvantage of BoW?

A. No vocabulary
B. Dense vectors
C. No order/semantic info
D. Too accurate
Answer: C

34. Advantage of Word2Vec?

A. Sparse vectors
B. Captures semantic similarity
C. Hard to train
D. Large memory usage
Answer: B

35. TF-IDF improves BoW by—

A. Adding semantics
B. Reducing word importance
C. Down-weighting common words
D. Reducing vocabulary
Answer: C

36. Word2Vec problem if trained on biased data?

A. Small vectors
B. Cannot converge
C. Embeddings become biased
D. Needs huge RAM
Answer: C

37. Why is Word2Vec not good for OOV (out of vocabulary)?

A. It’s too slow
B. Cannot generate unseen word vectors
C. Ignores stop words
D. Uses too many dimensions
Answer: B

38. BoW still useful when—

A. Semantic meaning matters
B. Word order matters
C. Very small datasets
D. Vocabulary is complex
Answer: C

39. Skip-gram uses which mechanism for better speed?

A. Regular softmax
B. Negative sampling
C. Sparse matrix multiply
D. Window expansion
Answer: B

40. CBOW smoothing helps because—

A. Averages context words
B. Removes noise
C. Learns grammar
D. Reduces softmax
Answer: A
 If you want, I can also provide:
✔ A PDF
✔ A mind-map style summary
✔ A memory sheet (one page quick revision)
✔ A practice exam (80–100 questions)
Just tell me which format you want.

give a mind map for Word2Vec, CBOW, Skip-gram, Bag-of-Words (BoW),

ngrams

Here is a clean, exam-friendly mind map for Word2Vec, CBOW, Skip-gram,

Bag-of-Words, N-grams — structured in a way you can memorize quickly.

Mind Map: Word2Vec, CBOW, Skip-gram, BoW, N-grams

Bag-of-Words (BoW)
Concept:
● Converts text → vector of word counts
● Ignores grammar & word order
● Vocabulary size = vector size
Pros:
● Simple
● Fast
● Good for small datasets + classical ML
Cons:
● Sparse vectors
● No semantics
● No context
● Large dimensionality
Use cases:
● Baseline models
● Text classification
● TF-IDF improvement

N-grams
Definition: Sequence of N words
● Unigrams (1)
● Bigrams (2)
● Trigrams (3)
Why useful:
● Captures local word order
● Better than BoW for phrases
Pros:
● Handles some context
 ● Easy to understand
Cons:
● Explodes vocabulary size
● Still no deep semantics
● Memory heavy
Use cases:
● Language modelling
● Text classification
● Feature engineering

Word2Vec (General Concept)

Goal: Learn dense word embeddings (100–300 dims)
Approach: Predictive model based on context-window
Why powerful:
● Captures semantics
● Vector arithmetic (king – man + woman = queen)
● Dense, compact vectors
Training tricks:
● Negative sampling
● Hierarchical softmax
● Subsampling frequent words
Limitation:
● One vector per word → poor for polysemy (“bank”)
● OOV issue (out-of-vocab words not handled)

CBOW (Continuous Bag of Words)

Predict: Target word from context words
Flow:
Context → Hidden vector → Predicted target
Pros:
● Fast
● Stable predictions
● Good for frequent words
● Smaller data OK
Cons:
● Not good for rare words
● Averages context → loses order
Use cases:
● Large corpora requiring speed
● Real-time NLP pipelines

Skip-gram
Predict: Context words from target
Flow:
Target → Hidden → Multiple context predictions
Pros:
● Excellent for rare words
● Better quality embeddings
● Works well on small datasets
Cons:
● Slow to train
● Needs more data to converge
Use cases:
● Semantic tasks
● Small corpora
● Rare-word heavy domains

Combined Mind Map Flow

┌───────────────┐
│ NLP Text │
└──────┬────────┘
│

┌─────────────────────┼─────────
─────────────┐
│ │ │
┌─────────▼─────────┐
┌──────▼───────┐
┌───────▼────────┐
│ Bag-of-Words │ │ N-grams │ │ Word2Vec │
└─────────┬──────────┘
└──────┬────────┘
└────────┬──────┘
│ │ │
Sparse counts, no order Captures local order Dense
semantic vectors
High dimension, simple Big vocabulary issue Vector
arithmetic
│ │ │
│ │ │
┌───────────┴───────────────┐
┌────────┴───────────┐
│ TF-IDF │ │ Two architectures │
└────────────────────────────┘
└────────┬───────────┘
 │

┌────────────────────────────┴──
─────────────────────────┐
│ │
┌───────▼───────────┐
┌──────▼──────────┐
│ CBOW │ │ Skip-gram │
└────────────────────┘
└─────────────────┘
Predict target word Predict
context
Fast, good for frequent Slow, best
for rare

Want the Mind Map as:

✔ PDF
✔ One-page cheat sheet
✔ PowerPoint slide
✔ PNG image mind map (visual)
Tell me your preferred format!

just mapflow in text cat

Here is a clean text-only map-flow (CAT style) — compact, hierarchical, easy

to memorize.

TEXT MAP-FLOW (CAT STYLE)

(Word2Vec, CBOW, Skip-gram, BoW, N-grams)

NLP TEXT
├── Bag-of-Words (BoW)
│ ├── Count vectors
│ ├── Sparse, high dimension
│ ├── Ignores order & semantics
│ ├── Pros: Simple, fast
│ └── Cons: No meaning, huge vectors
│
├── N-grams
│ ├── Unigram, Bigram, Trigram
│ ├── Captures local order
│ ├── Pros: Better than BoW for phrases
│ └── Cons: Exploding vocabulary
│
└── Word2Vec
├── Dense semantic embeddings
├── Learns from context window
├── Two architectures:
│
│ ├── CBOW
│ │ ├── Context → target
│ │ ├── Fast, stable
│ │ ├── Good for frequent words
│ │ └── Weak for rare words
│
└── Skip-gram
├── Target → context
├── Slow but powerful
├── Best for rare words
└── Works well on small data

NLP • Text Representation Techniques • Bag of Words (BoW) • TF-IDF •

Word2Vec Text classification • Sentiment analysis • Language translation SMT,
NMT Generative Pre-trained Transformer Bidirectional Encoder Representations
from Transformers Transformer give me 100 mcq on these topic with answer
analysis with why other is not right answer

That’s a big and excellent NLP request — you want 100 high-quality MCQs
covering:
● Text representation (BoW, TF-IDF, Word2Vec)
● Text classification tasks (Sentiment, Translation)
● SMT vs NMT
● Transformers (GPT, BERT, Transformer architecture)
Each question with:
Correct answer
Short analysis/explanation
Why other options are incorrect
To keep the detail high-quality and readable, I’ll give them in sets of 25 MCQs
per batch (so you can review and memorize easily).
Would you like the first 25 to focus on:
(A) Text Representation (BoW, TF-IDF, Word2Vec),
or
(B) Transformers and NLP Models (SMT, NMT, GPT, BERT, Transformer)?
Once you pick, I’ll generate the first 25 with full analysis and reasoning.

Perfect — we’ll start with Set 1: Text Representation Techniques (BoW,

TF-IDF, Word2Vec).
Here are 25 MCQs with correct answers, detailed analysis, and why other
options are wrong.

SET 1: Text Representation Techniques

1.
Which of the following best describes the Bag of Words (BoW) model?
A. Considers word order and context
B. Ignores word order, represents text as word frequency
C. Uses neural embeddings for meaning
D. Generates probabilistic word sequences
Answer: B
Analysis: BoW counts word occurrences, ignoring order.
A: Word order ignored.
C: That’s Word2Vec.
D: That’s a language model concept.

2.
In BoW, two sentences with the same words but in different order will have:
A. Same vector representation
B. Different vector representation
C. No vector representation
D. Randomized vector representation
Answer: A
Analysis: Word order is ignored, so both have same vector.
B: Only true if model considers order.
C: Always has a vector.
D: Not random.

3.
Which of the following can cause sparsity in BoW?
A. Large vocabulary size
B. Stop word removal
C. Lemmatization
D. Low-dimensional embeddings
Answer: A
Analysis: Huge vocab = most entries zero → sparse.
B/C: Reduce sparsity.
D: Related to Word2Vec, not BoW.
4.
TF in TF-IDF stands for:
A. Term Frequency
B. Text Factor
C. Token Function
D. Term Formula
Answer: A
Analysis: TF measures how frequently a word appears in a document.
B/C/D: Incorrect terms.

5.
The “IDF” part of TF-IDF penalizes words that:
A. Are rare
B. Are common across many documents
C. Are not in dictionary
D. Have multiple meanings
Answer: B
Analysis: IDF reduces weight of common words (like “the”, “is”).
A: Rare words get higher weight.
C/D: Unrelated.

6.
Which formula best represents TF-IDF weighting?
A. TF × IDF
B. TF / IDF
C. TF + IDF
D. TF × log(IDF)
Answer: A
Analysis: TF-IDF = TF × IDF where IDF = log(N/df).
B/C/D: Not the standard formula.

7.
What is the main limitation of TF-IDF?
A. It cannot handle rare words
B. It ignores semantics and word order
C. It overfits data
D. It requires neural networks
Answer: B
Analysis: TF-IDF captures frequency, not meaning or context.
A: Actually gives high weight to rare words.
C: Not model-based.
D: No neural networks used.

8.
Which text representation captures semantic meaning best?
A. BoW
B. TF-IDF
C. Word2Vec
D. Count Vectorizer
Answer: C
Analysis: Word2Vec learns contextual embeddings.
A/B/D: Frequency-based, not semantic.

9.
Word2Vec uses which learning technique?
A. Supervised learning
B. Unsupervised learning
C. Reinforcement learning
D. Semi-supervised learning
Answer: B
Analysis: Learns word representations from raw text without labels.
A/C/D: Not applicable.

10.
In Word2Vec, CBOW model predicts:
A. Center word from context words
B. Context words from center word
C. Sentence sentiment
D. Next sentence in sequence
Answer: A
Analysis: CBOW = Continuous Bag of Words predicts target word from context.
B: That’s Skip-gram.
C/D: Different tasks.

11.
In Word2Vec, Skip-gram model predicts:
A. Center word from context
B. Context words from center word
C. Word frequency
D. Sentence embedding
Answer: B
Analysis: Skip-gram uses target word to predict surrounding context.
A: CBOW does that.
C/D: Not model output.

12.
Word2Vec representations are:
A. Sparse and high-dimensional
B. Dense and low-dimensional
C. Symbolic and discrete
D. Binary
Answer: B
Analysis: Dense embeddings capture meaning in few dimensions.
 A: BoW/TF-IDF are sparse.
C/D: Incorrect.

13.
In Word2Vec training, the objective is to:
A. Maximize reconstruction loss
B. Minimize cosine similarity
C. Maximize likelihood of nearby words
D. Predict next sentence
Answer: C
Analysis: Learns vectors that make context words probable.
A/B/D: Not applicable.

14.
Which vector similarity measure is often used with Word2Vec?
A. Euclidean distance
B. Cosine similarity
C. Manhattan distance
D. Jaccard similarity
Answer: B
Analysis: Cosine similarity measures angle → semantic closeness.
A/C: Distance-based, less meaningful for embeddings.
D: Set-based metric.

15.
BoW, TF-IDF, and Word2Vec are all examples of:
A. Feature extraction techniques
B. Supervised models
C. Classification algorithms
D. Tokenization methods
Answer: A
Analysis: They extract numerical features from text.
B/C/D: Not models or tokenizers.

16.
Which of these techniques would give different vectors for “bank” in “river
bank” and “money bank”?
A. BoW
B. TF-IDF
C. Word2Vec
D. Count Vectorizer
Answer: C
Analysis: Word2Vec captures context; meaning changes.
A/B/D: Ignore context.
17.
If two words have similar meanings, Word2Vec will produce:
A. Similar vectors
B. Dissimilar vectors
C. Zero vectors
D. Random vectors
Answer: A
Analysis: Semantically related words are close in vector space.
B/C/D: Not how embeddings work.

18.
TF-IDF assigns high weight to words that:
A. Appear often in a single document but rarely overall
B. Appear in all documents equally
C. Are stop words
D. Are unknown tokens
Answer: A
Analysis: Important distinguishing words get higher weight.
B/C/D: Lower or no weight.

19.
What does “embedding” mean in NLP?
A. Frequency count of words
B. Dense vector representation capturing meaning
C. Random encoding
D. Syntax parsing
Answer: B
Analysis: Embeddings = dense numerical meaning representation.
A/C/D: Wrong concepts.

20.
In BoW, what happens when vocabulary size increases?
A. Sparsity decreases
B. Model generalization improves
C. Sparsity increases
D. Semantic meaning captured
Answer: C
Analysis: Larger vocab = more zeros per vector.
A/B/D: Opposite or irrelevant.

21.
The process of reducing words to their root form before BoW is:
A. Lemmatization
B. Stop-word removal
C. Tokenization
D. Chunking
 Answer: A
Analysis: Lemmatization → reduces “running”, “ran” → “run”.
B/C/D: Different preprocessing steps.

22.
Why might Word2Vec outperform TF-IDF?
A. It captures order and semantics
B. It counts word frequencies
C. It ignores context
D. It uses symbolic representation
Answer: A
Analysis: Word2Vec models relationships and semantics.
B/C/D: TF-IDF traits.

23.
The training objective of Skip-gram involves:
A. Maximizing probability of context words given center
B. Minimizing word frequency
C. Counting bigrams
D. Random initialization
Answer: A
Analysis: Predicts context from center → probability-based training.
B/C/D: Not relevant.

24.
Which statement is true about TF-IDF compared to BoW?
A. TF-IDF penalizes frequent words, BoW doesn’t
B. BoW uses context, TF-IDF doesn’t
C. TF-IDF always produces dense vectors
D. BoW has fewer dimensions
Answer: A
Analysis: TF-IDF reduces common word impact.
B: Both ignore context.
C: Still sparse.
D: Same vocab size.

25.
What is the embedding dimension in Word2Vec?
A. Number of unique words
B. Number of context windows
C. Length of each word vector
D. Number of sentences
Answer: C
Analysis: Each word represented as fixed-length vector (e.g., 100D).
A/B/D: Not dimensions.

Set Summary:
 ● BoW → Frequency-based, ignores order.
● TF-IDF → Adds importance weighting.
● Word2Vec → Learns dense semantic meaning.

Would you like me to now continue with Set 2 (next 25 MCQs) on Text
Classification (Sentiment + Language Translation)?

ok pls

Great
Now continuing with SET 2: Text Classification & Language Translation
(Sentiment Analysis, Text Classification Concepts, Language Translation)
Below are 25 MCQs (26–50) with
Correct answer
Explanation
Why other options are wrong

SET 2: Text Classification & Language Translation

26.
Text classification mainly refers to:
A. Translating text between languages
B. Assigning predefined labels to text
C. Generating new text
D. Compressing text data
Answer: B
Analysis: Text classification assigns categories (spam/ham, positive/negative).
A: Translation task
C: Text generation
D: Data compression

27.
Which of the following is a text classification problem?
A. Summarization
B. Sentiment analysis
C. Machine translation
D. Question answering
Answer: B
Analysis: Sentiment analysis classifies text into sentiments.
A/C/D: Different NLP tasks.

28.
Sentiment analysis typically classifies text into:
A. Topics
B. Grammar rules
C. Emotional polarity
D. Word frequency
Answer: C
Analysis: Sentiment = positive, negative, neutral emotions.
A: Topic modeling
B/D: Not sentiment-related.

29.
Which sentiment classification is the simplest form?
A. Aspect-based sentiment
B. Fine-grained sentiment
C. Binary sentiment
D. Emotion detection
Answer: C
Analysis: Binary sentiment → positive vs negative.
A/B/D: More complex.

30.
Aspect-based sentiment analysis focuses on:
A. Entire document sentiment
B. Sentence length
C. Sentiment toward specific features
D. Language detection
Answer: C
Analysis: Example: “Battery is good, camera is bad.”
A: Document-level sentiment
B/D: Irrelevant.

31.
Which ML algorithm is commonly used for text classification?
A. K-Means
B. Naive Bayes
C. Apriori
D. KNN Regression
Answer: B
Analysis: Naive Bayes works well with text features.
A: Clustering
C: Association rule mining
D: Regression variant

32.
Why is Naive Bayes effective for sentiment analysis?
A. It captures word order
B. Assumes feature independence
C. Uses deep neural networks
D. Requires no text preprocessing
Answer: B
Analysis: Independence assumption works surprisingly well in text.
 A: It ignores order
C: Not deep learning
D: Preprocessing still needed

33.
Which feature representation is most suitable for classical sentiment analysis?
A. Raw text
B. TF-IDF
C. Image pixels
D. Audio signals
Answer: B
Analysis: TF-IDF captures word importance.
A: Needs vectorization
C/D: Different domains

34.
Which metric is best for imbalanced sentiment datasets?
A. Accuracy
B. Precision, Recall, F1-score
C. Mean Squared Error
D. R² Score
Answer: B
Analysis: Accuracy can be misleading in imbalance.
A: Not reliable
C/D: Regression metrics

35.
Text classification usually follows which pipeline?
A. Tokenization → Feature extraction → Model training
B. Translation → Classification → Tokenization
C. Parsing → Speech recognition
D. Image preprocessing → CNN
Answer: A
Analysis: Standard NLP pipeline.
B/C/D: Incorrect order or domain.

36.
Language translation in NLP means:
A. Converting speech to text
B. Translating text between languages
C. Summarizing documents
D. Classifying text topics
Answer: B
Analysis: Machine Translation converts one language to another.
A: Speech recognition
C: Summarization
D: Classification
37.
Which is an example of machine translation?
A. English → French
B. Spam detection
C. Emotion detection
D. Topic modeling
Answer: A
Analysis: Translation task.
B/C/D: Classification or unsupervised tasks.

38.
Which challenge is common in language translation?
A. Word ambiguity
B. Fixed grammar rules
C. Small vocabulary
D. Lack of syntax
Answer: A
Analysis: Words may have multiple meanings.
B/C/D: Incorrect assumptions.

39.
What does parallel corpus mean in translation?
A. Same text written twice
B. Aligned sentences in two languages
C. Random text collection
D. Audio-text pairs
Answer: B
Analysis: Used to train translation models.
A/C/D: Incorrect definition.

40.
Which representation is most useful for neural translation?
A. One-hot vectors
B. TF-IDF
C. Word embeddings
D. Count vectors
Answer: C
Analysis: Embeddings capture semantic relationships.
A/B/D: Sparse and limited.

41.
Text classification is usually:
A. Supervised learning
B. Unsupervised learning
C. Reinforcement learning
D. Clustering
Answer: A
Analysis: Requires labeled data.
B/D: No labels
C: Not typical.

42.
Which problem occurs when sentiment depends on context?
A. Polarity shift
B. Overfitting
C. Underfitting
D. Tokenization
Answer: A
Analysis: “Not bad” flips sentiment meaning.
B/C/D: Different issues.

43.
Why is sarcasm difficult in sentiment analysis?
A. Words contradict actual meaning
B. Vocabulary is small
C. Grammar is incorrect
D. Language is unknown
Answer: A
Analysis: Positive words may imply negative sentiment.
B/C/D: Not the reason.

44.
Which task predicts whether a review is positive or negative?
A. Topic modeling
B. Sentiment classification
C. Language modeling
D. Parsing
Answer: B
Analysis: Classic sentiment task.
A/C/D: Different NLP problems.

45.
In multi-class sentiment analysis, labels can be:
A. Positive, Negative
B. Happy, Sad, Angry, Neutral
C. Spam, Ham
D. English, French
Answer: B
Analysis: Multiple emotion categories.
A: Binary
C/D: Different tasks.
46.
Which approach uses neural networks for translation?
A. Rule-based MT
B. Statistical MT
C. Neural MT
D. Dictionary-based MT
Answer: C
Analysis: Neural Machine Translation uses deep learning.
A/B/D: Older methods.

47.
Sentiment analysis at document level means:
A. Classifying each word
B. Classifying entire document sentiment
C. Translating document
D. Extracting keywords
Answer: B
Analysis: Overall sentiment of the document.
A/C/D: Incorrect tasks.

48.
Which preprocessing step removes common words like “is”, “the”?
A. Stemming
B. Tokenization
C. Stop-word removal
D. Lemmatization
Answer: C
Analysis: Removes low-information words.
A/B/D: Different preprocessing steps.

49.
Why are embeddings preferred over one-hot encoding?
A. They are larger
B. They capture semantic similarity
C. They are random
D. They require more memory
Answer: B
Analysis: Similar words have similar vectors.
A/D: Actually smaller
C: Learned, not random.

50.
Language translation is best described as:
A. Classification problem
B. Sequence-to-sequence problem
C. Clustering problem
D. Regression problem
Answer: B
Analysis: Input and output are sequences.
A/C/D: Incorrect formulation.

Progress so far:
● 1–25: Text Representation
● 26–50: Text Classification & Translation
Shall I continue with SET 3 (51–75) on SMT vs NMT + Transformer basics
next?

ok

Great
Now continuing with SET 3: SMT vs NMT & Transformer Fundamentals (51–
75)
As before, each question includes:
Correct answer
Explanation
Why other options are incorrect

SET 3: SMT, NMT & Transformer Basics

51.
What does SMT stand for in NLP?
A. Semantic Machine Translation
B. Statistical Machine Translation
C. Supervised Machine Translation
D. Sequential Machine Translation
Answer: B
Analysis: SMT uses statistical models based on probability.
A/C/D: Incorrect expansions.

52.
SMT primarily relies on:
A. Hand-crafted grammar rules
B. Probability distributions from parallel corpora
C. Deep neural networks
D. Reinforcement learning
Answer: B
Analysis: SMT learns translation probabilities from aligned data.
A: Rule-based MT
C: NMT
D: Not used.
53.
Which component is NOT part of traditional SMT?
A. Language model
B. Translation model
C. Attention mechanism
D. Decoder
Answer: C
Analysis: Attention is used in NMT, not SMT.
A/B/D: Core SMT components.

54.
One major limitation of SMT is:
A. High memory efficiency
B. Poor handling of long-distance dependencies
C. Requires no data
D. Perfect fluency
Answer: B
Analysis: SMT struggles with context and long sentences.
A: Opposite
C: Needs large corpora
D: Fluency is often poor.

55.
What does NMT stand for?
A. Neural Machine Translation
B. Natural Machine Translation
C. Networked Machine Translation
D. Numeric Machine Translation
Answer: A
Analysis: NMT uses neural networks end-to-end.
B/C/D: Incorrect terms.

56.
NMT treats translation as:
A. Word-to-word mapping
B. Sequence-to-sequence learning
C. Classification problem
D. Rule-based conversion
Answer: B
Analysis: Input and output are sequences.
A: SMT-like
C/D: Incorrect.

57.
Which neural architecture was first widely used in NMT?
A. CNN
B. RNN (Encoder–Decoder)
C. Decision Tree
D. SVM
Answer: B
Analysis: Early NMT used RNN encoder–decoder models.
A: Later usage
C/D: Not suitable.

58.
Why was the attention mechanism introduced in NMT?
A. To increase vocabulary size
B. To focus on relevant source words during translation
C. To reduce dataset size
D. To remove encoder
Answer: B
Analysis: Attention aligns source and target words dynamically.
A/C/D: Incorrect.

59.
Compared to SMT, NMT generally produces translations that are:
A. Less fluent
B. More fluent and natural
C. Word-by-word
D. Grammar-free
Answer: B
Analysis: NMT captures context better.
A/C/D: Opposite or incorrect.

60.
One disadvantage of NMT is:
A. Requires no training data
B. Computationally expensive
C. Cannot handle unknown words
D. Produces deterministic outputs only
Answer: B
Analysis: NMT requires GPUs and large data.
A: Needs lots of data
C: Subword techniques help
D: Outputs are probabilistic.

61.
The Transformer model was introduced to:
A. Replace BoW
B. Remove recurrence and convolution
C. Improve word counting
D. Reduce vocabulary
Answer: B
Analysis: Transformer uses self-attention instead of RNN/CNN.
A/C/D: Incorrect.

62.
Which paper introduced the Transformer architecture?
A. Word2Vec
B. Attention Is All You Need
C. BERT
D. GPT
Answer: B
Analysis: Vaswani et al., 2017.
A/C/D: Different works.

63.
The core building block of Transformer is:
A. Recurrent unit
B. Convolution layer
C. Self-attention mechanism
D. Hidden Markov Model
Answer: C
Analysis: Self-attention enables parallel processing.
A/B/D: Not Transformer-based.

64.
Self-attention allows the model to:
A. Process words sequentially
B. Focus on different parts of the sentence simultaneously
C. Ignore word relationships
D. Use fixed context window
Answer: B
Analysis: Captures global dependencies.
A: RNN behavior
C/D: Incorrect.

65.
Which problem of RNNs does Transformer solve?
A. Overfitting
B. Vanishing gradients & lack of parallelism
C. Tokenization
D. Vocabulary mismatch
Answer: B
Analysis: Transformer removes recurrence.
A/C/D: Not primary issues.

66.
Transformer encoder consists of:
A. Self-attention + Feed-forward network
B. CNN + Pooling
C. RNN + Attention
D. Only embedding layer
Answer: A
Analysis: Standard encoder block.
B/C/D: Incorrect.

67.
What is multi-head attention?
A. Multiple decoders
B. Attention applied at different representation subspaces
C. Multiple vocabularies
D. Parallel datasets
Answer: B
Analysis: Enables richer representation.
A/C/D: Incorrect.

68.
Why is positional encoding needed in Transformer?
A. To remove attention
B. To provide word order information
C. To reduce dimensions
D. To replace embeddings
Answer: B
Analysis: Transformer has no inherent sequence order.
A/C/D: Incorrect.

69.
Which component helps Transformer learn non-linear patterns?
A. Self-attention
B. Feed-forward network
C. Tokenizer
D. Softmax
Answer: B
Analysis: FFN adds non-linearity.
A/C/D: Different roles.

70.
Transformer processes sequences:
A. One token at a time
B. Sequentially only
C. In parallel
D. Backwards only
Answer: C
Analysis: Parallelism is a major advantage.
A/B/D: RNN-like.
71.
Which model uses only the Transformer encoder?
A. GPT
B. BERT
C. T5
D. Transformer (original)
Answer: B
Analysis: BERT is encoder-only.
A: Decoder-only
C: Encoder-decoder
D: Full architecture.

72.
Which model uses only the Transformer decoder?
A. BERT
B. GPT
C. Transformer encoder
D. ELMo
Answer: B
Analysis: GPT is autoregressive decoder-only.
A/C/D: Different architectures.

73.
Original Transformer was designed mainly for:
A. Sentiment analysis
B. Language translation
C. Text summarization
D. Image classification
Answer: B
Analysis: Proposed for NMT.
A/C/D: Later applications.

74.
Compared to RNN-based NMT, Transformer:
A. Is slower
B. Cannot model long sequences
C. Handles long-range dependencies better
D. Uses less memory always
Answer: C
Analysis: Self-attention captures global context.
A/B/D: Incorrect.

75.
SMT vs NMT: which statement is correct?
A. SMT uses neural networks
B. NMT translates word-by-word
C. SMT relies on probability tables
D. NMT requires no data
Answer: C
Analysis: SMT is statistical, table-based.
A: NMT does
B: SMT-like
D: Needs lots of data.

Progress:
● 1–25: Text Representation
● 26–50: Text Classification & Translation
● 51–75: SMT, NMT & Transformer Basics
Next and final SET 4 (76–100): GPT, BERT & Advanced Transformer
MCQs
Shall I continue?