Mockpreps
🧑💼
🪴
Login now and master your exams with Mockpreps’ mock tests!
Top Interview Questions
Natural Language Processing (NLP) is a field of computer science and artificial intelligence that focuses on enabling machines to understand, interpret, and generate human language. It bridges the gap between human communication and computer understanding, allowing machines to process text and speech in a meaningful way.
At its core, Natural Language Processing (NLP) is about teaching computers to work with human language—whether it’s written text or spoken words. Humans communicate in complex, nuanced ways, full of ambiguity, context, and emotion. NLP aims to make machines capable of handling this complexity.
For example, when you ask a voice assistant a question, use autocomplete while typing, or translate text between languages, NLP is working behind the scenes.
Human language is unstructured and highly variable. Unlike programming languages, it doesn’t follow strict rules. NLP is important because it allows computers to:
Understand human input (text or speech)
Extract useful information from large volumes of data
Automate communication tasks
Improve human-computer interaction
With the explosion of digital data (emails, social media, documents), NLP helps organizations make sense of vast amounts of text.
NLP involves several stages and techniques:
Before analyzing text, it must be cleaned and prepared.
Common steps include:
Tokenization (breaking text into words or sentences)
Removing stop words (e.g., "is", "the")
Stemming and lemmatization (reducing words to base forms)
Syntax refers to the grammatical structure of a sentence.
Techniques include:
Part-of-speech tagging (noun, verb, adjective, etc.)
Parsing (analyzing sentence structure)
This focuses on understanding the meaning of words and sentences.
Examples:
Word sense disambiguation (understanding context)
Named entity recognition (identifying names, places, dates)
Pragmatics deals with understanding context beyond literal meaning.
For example:
“Can you open the window?” is a request, not a question about ability.
Understanding how sentences relate to each other in a conversation or paragraph.
Splitting text into smaller units like words or phrases.
Represents text as a collection of word frequencies.
Measures importance of a word in a document relative to a dataset.
Converts words into numerical vectors that capture meaning.
Popular models:
Word2Vec
GloVe
Modern NLP relies heavily on neural networks.
Examples include:
Recurrent Neural Networks (RNNs)
Transformers
A major breakthrough came with transformer-based models like Transformer architecture, which power systems such as GPT and BERT.
NLP is widely used in many real-world applications:
Systems like chatbots and assistants can understand and respond to user queries.
Examples:
Customer support bots
Voice assistants
Automatically translating text from one language to another.
Example:
Google Translate
Determining whether text expresses positive, negative, or neutral sentiment.
Used in:
Product reviews
Social media monitoring
Condensing large documents into shorter summaries.
Converting spoken language into text.
Improving search results by understanding user queries.
Filtering unwanted emails or messages.
While NLP has made significant progress, human language understanding is still far more advanced.
Challenges include:
Ambiguity (words with multiple meanings)
Sarcasm and humor
Cultural context
Idioms and slang
For example:
“It’s raining cats and dogs” cannot be interpreted literally.
Words and sentences can have multiple meanings.
Meaning often depends on context.
NLP models require large amounts of training data.
Different languages have different grammar and structure.
Models may inherit biases from training data.
NLP heavily relies on machine learning and deep learning techniques.
Traditional NLP used rule-based systems, but modern NLP uses:
Statistical methods
Neural networks
Large language models
These models learn patterns from data rather than relying on predefined rules.
Early NLP relied on hand-written rules.
Used probabilities and data-driven approaches.
Introduction of neural networks improved accuracy.
Models like ChatGPT and BERT revolutionized NLP by handling context more effectively.
Automates text processing
Improves customer experience
Handles large-scale data efficiently
Enables intelligent applications
Struggles with nuanced language
Requires significant computational power
May produce incorrect or biased results
The future of NLP is promising, with advancements in:
Conversational AI
Multilingual models
Emotion detection
Real-time translation
NLP is becoming more accurate and context-aware, moving closer to human-like understanding.
Natural Language Processing (NLP) is a powerful and rapidly evolving field that enables machines to understand and interact with human language. From chatbots and translation tools to search engines and sentiment analysis, NLP is transforming how humans interact with technology.
As advancements in AI and deep learning continue, NLP will play an even greater role in shaping the future of communication, making interactions between humans and machines more natural, intuitive, and efficient.
Answer:
Natural Language Processing (NLP) is a branch of AI that enables computers to understand, interpret, and generate human language.
Applications:
Chatbots
Spam detection
Sentiment analysis
Machine translation
Answer:
NLU (Natural Language Understanding): Understands human text/speech. Example: Intent detection in chatbots.
NLG (Natural Language Generation): Generates human-like text. Example: AI content generators.
Answer:
Text classification
Sentiment analysis
Named Entity Recognition (NER)
POS tagging
Machine translation
Question answering
Answer:
NLP: Focused on language understanding and generation.
Text Analytics: Focused on extracting insights/statistics from text.
Answer:
Stop words are common words (like “is”, “the”, “and”) that are often removed to reduce noise in text processing.
Example:
Original: “The cat is on the mat”
After removing stop words: “cat mat”
Answer:
Tokenization is splitting text into tokens (words, sentences, or subwords).
Example:
Sentence: “I love NLP”
Word tokens: [“I”, “love”, “NLP”]
Answer:
Stemming reduces words to their root form.
Example:
“Running” → “run”
“Studies” → “studi”
Common Algorithms: Porter Stemmer, Snowball Stemmer
Answer:
Lemmatization reduces words to their dictionary/base form.
Example:
“Running” → “run”
“Better” → “good”
Difference: Lemmatization is more accurate than stemming.
Answer:
Part-of-Speech tagging assigns grammatical tags to each word.
Example:
Sentence: “I love NLP”
POS tags: [(“I”, PRON), (“love”, VERB), (“NLP”, NOUN)]
Answer:
NER identifies entities like names, dates, locations, or organizations in text.
Example:
“Apple is launching a new iPhone in California”
Entities: [(“Apple”, ORG), (“iPhone”, PRODUCT), (“California”, LOC)]
Answer:
Converting text to a standard form for NLP tasks. Includes:
Lowercasing
Removing punctuation
Expanding contractions
Removing special characters
Answer:
BoW represents text as a vector of word counts, ignoring grammar and word order.
Example:
Texts: [“I love NLP”, “NLP is great”]
Vocabulary: [“I”, “love”, “NLP”, “is”, “great”]
Vectors: [1,1,1,0,0], [0,0,1,1,1]
Answer:
TF-IDF (Term Frequency–Inverse Document Frequency) weights words based on importance.
TF: Frequency of word in document
IDF: Rare words across corpus get higher weight
Formula:
TF-IDF = TF × log(N / DF)
Answer:
N-gram is a sequence of n words.
Unigram → 1 word
Bigram → 2 words
Trigram → 3 words
Example:
Sentence: “I love NLP”
Bigrams: [“I love”, “love NLP”]
Answer:
Word embeddings represent words as dense vectors capturing semantic meaning.
Examples:
Word2Vec
GloVe
FastText
| Feature | One-Hot | Word Embedding |
|---|---|---|
| Dimensionality | Vocabulary size | Fixed, smaller |
| Semantic info | No | Yes |
| Sparsity | Sparse | Dense |
Answer:
Word2Vec is a neural network-based embedding model.
CBOW → Predicts a word from context
Skip-gram → Predicts context from word
Answer:
GloVe (Global Vectors) is a count-based word embedding trained on co-occurrence statistics.
Answer:
FastText represents words as subword (character n-grams) embeddings.
Handles rare words and misspellings better
Answer:
Represents document by averaging word embeddings of its words.
Answer:
Sequence modeling predicts or generates sequences of data (like text).
Examples:
RNN (Recurrent Neural Network)
LSTM (Long Short-Term Memory)
GRU (Gated Recurrent Unit)
Answer:
Attention allows models to focus on important words in input while generating output.
Example:
Machine translation: Focus on “Paris” when translating French sentence to English
Answer:
Transformer is a deep learning architecture using attention instead of recurrence.
Components: Encoder, Decoder, Multi-head Attention, Feed-forward layers
Example models: BERT, GPT, T5
Answer:
BERT (Bidirectional Encoder Representations from Transformers) is pre-trained on large text corpus using Masked Language Modeling.
Good for NLU tasks: Q&A, classification, NER
Answer:
GPT (Generative Pre-trained Transformer) is a decoder-based transformer for text generation.
Generates human-like text
Unidirectional
| Feature | BERT | GPT |
|---|---|---|
| Type | Encoder | Decoder |
| Direction | Bidirectional | Left-to-right |
| Task | Understanding | Generation |
Answer:
Models that convert input sequence to output sequence.
Example: Machine translation, summarization
Uses Encoder-Decoder architecture
Answer:
Predicts masked words in a sentence.
Example:
Input: “I love [MASK]” → Model predicts “NLP”
Answer:
NLTK → Tokenization, POS tagging, corpus
spaCy → Industrial NLP, NER
Gensim → Word2Vec, topic modeling
Hugging Face Transformers → Pre-trained models
Answer:
Natural Language Toolkit (NLTK) is a Python library for text processing.
Tokenization, stemming, lemmatization, parsing
Answer:
spaCy is a fast NLP library with features:
Tokenization, NER, dependency parsing
Supports pre-trained embeddings
Optimized for production
Answer:
Gensim is used for topic modeling and word embeddings.
Word2Vec, Doc2Vec, LDA
Answer:
Library for pre-trained transformer models: BERT, GPT, T5, etc.
Features:
Fine-tuning
Tokenization
Easy deployment
Answer:
Measures how similar generated text is to reference text.
Common in machine translation
Answer:
Measures overlap of n-grams between generated summary and reference summary.
Answer:
Measures how well a language model predicts a sample.
Lower perplexity → better model
Answer:
Measures similarity between two vector embeddings.
Value: -1 (opposite) → 1 (same)
Answer:
Tokenization → Split words/sentences
Stop word removal → Reduce noise
Lemmatization/stemming → Normalize text
Vectorization → BoW, TF-IDF, embeddings
Answer:
Use subword embeddings (FastText, BPE)
Map unknown words to <UNK> token
Answer:
Use streaming or batching
Use efficient embeddings (Word2Vec, FastText)
Avoid loading entire corpus in memory
Answer:
Classical: Naive Bayes, SVM with BoW/TF-IDF
Deep Learning: LSTM, BERT
Answer:
BERT, RoBERTa, DistilBERT
Answer:
Seq2Seq with attention
Transformer models like T5, BART
| Feature | Text Classification | Sequence Labeling |
|---|---|---|
| Output | Single label | Label per token |
| Example | Spam detection | POS, NER |
Answer:
Tokenization → Splits text into words
Subword tokenization → Splits words into smaller units (handles OOV)
Understand concept before memorizing
Give examples from real life (emails, chatbots, social media)
Use diagrams for model architecture if asked
Show familiarity with Python libraries
Text preprocessing (tokenization, stopwords, stemming, lemmatization)
Vectorization (BoW, TF-IDF, embeddings)
Common models (RNN, LSTM, Transformers)
Evaluation metrics (BLEU, ROUGE, perplexity)
Python libraries (NLTK, spaCy, Gensim, HuggingFace)
Answer:
NLP is a branch of AI that enables machines to understand, interpret, and generate human language.
Applications:
Chatbots (e.g., customer support)
Sentiment analysis
Machine translation (e.g., Google Translate)
Text summarization
Information extraction
| Term | Focus |
|---|---|
| NLP | Processing text |
| NLU | Understanding meaning (intent, entities) |
| NLG | Generating text (summaries, responses) |
Tokenization: Splitting text into words/sentences.
Stemming: Reducing words to root (e.g., “running” → “run”)
Lemmatization: Reducing to dictionary form considering POS (e.g., “better” → “good”)
Trade-off: Stemming is faster but less accurate; Lemmatization is accurate but slower.
Bag-of-Words: Simple count of word occurrences, ignores word order.
TF-IDF: Weights words by frequency and rarity; reduces common words’ importance.
Use case: TF-IDF for text classification or retrieval; BoW for simple models.
Word2Vec: Predicts context (CBOW, Skip-gram)
GloVe: Matrix factorization of co-occurrence counts
Key: Captures semantic similarity (“king”–“man” + “woman” ≈ “queen”)
| Type | Example | Advantage |
|---|---|---|
| Count-based | TF-IDF | Simple, interpretable |
| Prediction-based | Word2Vec | Captures semantic relationships |
Binary vector per word
Limitation: high-dimensional, no semantic meaning, sparse
Predicts next word given previous words.
Types: N-gram, RNN-based, Transformer-based (BERT, GPT)
| Model | Key Feature |
|---|---|
| RNN | Sequential, suffers from vanishing gradients |
| LSTM | Memory cells + gates, handles long dependencies |
| GRU | Simplified LSTM, faster training |
RNN → small datasets, sequential tasks
Transformer → large datasets, long sequences, attention mechanism
Weights important words when encoding/decoding.
Solves long-term dependency problem in RNNs.
Key in Transformers (BERT, GPT).
Bidirectional Encoder Representations from Transformers
Pre-trained on masked language modeling + next sentence prediction
Fine-tuned for tasks: QA, NER, sentiment
RoBERTa → optimized BERT (longer pretraining, no NSP)
DistilBERT → smaller, faster, 40% fewer parameters, retains ~97% performance
GPT → Autoregressive, generates text, left-to-right
BERT → Bidirectional, mainly understanding, not generation
Contextual embeddings: each word vector depends on sentence context.
Unlike Word2Vec, same word in different sentences has different embeddings.
Subword tokenization (BPE, WordPiece)
Unknown token <UNK>
Character-level embeddings
Adds order information since Transformers have no recurrence
Sinusoidal functions or learned embeddings
MLM → BERT, masks random words, bidirectional
Causal → GPT, predicts next word, left-to-right
Pretrain → Large corpus
Fine-tune → Task-specific data
Benefits: less labeled data needed, better performance
Identify entities like Person, Location, Organization
Models: CRF, BiLSTM-CRF, Transformers (BERT-based)
Evaluation: F1-score
Tags words with grammatical role
Models: HMM, CRF, BiLSTM
Example: spam detection, sentiment analysis
Models: Logistic Regression, CNN, BERT
Features: TF-IDF, embeddings
Translation, summarization, question answering
Model: Encoder-Decoder RNN or Transformer
| Type | Explanation | Example |
|---|---|---|
| Extractive | Select sentences | TextRank |
| Abstractive | Generate new sentences | BART, T5 |
Determine correct sense of a word in context
Approaches: Knowledge-based (WordNet), Supervised, Contextual embeddings
Rule-based or ML-based
Challenges: sarcasm, negation
Transformers → state-of-the-art
Extractive → span selection (BERT)
Generative → generate answer (GPT)
Determine which pronouns refer to which entities
Models: rule-based, neural networks, SpanBERT
Cosine similarity over embeddings
Useful in recommendation, search engines
Unsupervised: LDA, NMF
Discover latent topics in text corpus
N-gram based
FastText or CLD3
Oversampling / undersampling
Focal loss for deep learning
Weighted loss functions
| Task | Metric |
|---|---|
| Classification | Accuracy, F1-score, Precision, Recall |
| Generation | BLEU, ROUGE, METEOR |
| NER | F1-score |
| Ranking/Search | MAP, NDCG |
Text preprocessing (cleaning, tokenization, stopwords)
Feature extraction (BoW, embeddings)
Modeling (classification, seq2seq)
Evaluation
Deployment
Streaming data instead of loading all in memory
Using TFRecords / memory-mapped files
Efficient tokenization (HuggingFace Tokenizers)
Use subword tokenizers
Add domain-specific vocabulary
Fallback to <UNK> or character embeddings
Multilingual embeddings: mBERT, XLM-R
Translation → monolingual models
Challenges: scripts, tokenization differences
Freeze lower layers
Gradual unfreezing
Use mixed precision (FP16)
Small learning rate for pre-trained layers
REST API (Flask/FastAPI)
Batch processing for large corpus
Optimizations:
ONNX / TorchScript
Quantization
GPU/CPU deployment strategies
Steps:
Intent recognition (classification)
Entity extraction (NER)
Response generation (retrieval-based / generative)
Context management (conversation history)
Evaluation (user satisfaction, BLEU for responses)
| Feature | Traditional ML | Deep Learning |
|---|---|---|
| Features | Manual (TF-IDF) | Embeddings (Word2Vec, BERT) |
| Performance | Good for small data | Excellent for large data |
| Training | Fast | Slow, requires GPU |
Noisy text (typos, slang)
Domain adaptation
Model drift over time
Latency for large Transformer models
Attention visualization
SHAP / LIME for feature importance
Embedding visualization (t-SNE, PCA)
Generative → models joint distribution, can generate text
Discriminative → models conditional probability, classification
Use Spark NLP
Parallel tokenization, embedding computation
Suitable for millions of documents
Knowledge distillation (e.g., DistilBERT)
Quantization (INT8)
Pruning unimportant weights
Sliding window
Longformer / BigBird architectures
Chunking + attention masking
| Term | Definition |
|---|---|
| Zero-shot | No labeled data, use pretrained LM |
| Few-shot | Small labeled examples |
| Fine-tuning | Full task-specific training |
Convert entities to canonical forms
Example: “NYC”, “New York City” → “New York City”
Extract entities and relations → build graph
Applications: QA, recommendation, reasoning
For 4+ years experience, interviewers expect you to demonstrate:
Real-world experience with preprocessing pipelines, embeddings, Transformers
Ability to handle production issues like latency, deployment, OOV
Deep understanding of model trade-offs, metrics, evaluation
Experience with transfer learning, fine-tuning, and multilingual NLP
