Module 1: Introduction to NLP & Ambiguity

Definition: NLP in Plain English

Natural Language Processing (NLP) is a branch of Artificial Intelligence that enables computers to understand, interpret, and generate human language — the language we speak and write every day.

Think about it: computers natively understand only numbers and binary (0s and 1s). They don't "know" that "apple" is a fruit, or that "love" is an emotion, or that "Can you pass the salt?" is a request, not a question about ability. NLP is the bridge between human language and computer understanding.

What NLP Systems Actually Do

NLP systems perform three broad categories of tasks:

• Understanding (Analysis): Reading text and extracting meaning. Examples: classifying emails as spam, identifying people's names in news articles, determining if a review is positive or negative.
• Transformation (Conversion): Changing text from one form to another. Examples: translating English to Hindi, converting speech to text, summarizing a 10-page article into 3 sentences.
• Generation (Creation): Producing new text from scratch. Examples: ChatGPT writing an essay, Gmail suggesting the next sentence in your email, AI writing code comments.

A Brief History of NLP

NLP has evolved through three major eras:

• 1950s-1990s: Rule-Based NLP. Linguists wrote thousands of hand-crafted rules. "If the sentence starts with 'Can you', it's likely a request." This approach was brittle — new rules had to be written for every new situation. The famous ELIZA chatbot (1966) used simple pattern matching to simulate a therapist.
• 1990s-2010s: Statistical NLP. Instead of rules, researchers used probability and statistics. "How likely is 'bank' to mean 'financial institution' when the previous word is 'river'?" N-gram models, Hidden Markov Models, and Naive Bayes classifiers dominated this era.
• 2010s-Present: Deep Learning & LLMs. Neural networks, especially the Transformer architecture (2017), revolutionized NLP. Models like BERT, GPT-4, and Claude read billions of text pages and learn patterns automatically. This is the era of ChatGPT and modern AI assistants.

NLP is not an academic exercise — it powers products that billions of people use every single day. Let's explore the major application areas with specific examples and the companies behind them.

1. Search & Information Retrieval

Every time you search on Google, Bing, or DuckDuckGo, NLP is working behind the scenes. Modern search engines don't just match keywords — they understand your intent.

Example: You search for "apple release date." The search engine knows you mean the company Apple, not the fruit, because "release date" is semantically associated with products, not produce. This is entity disambiguation using NLP.

Key Techniques: TF-IDF, BM25, Word2Vec, BERT embeddings, semantic search, query expansion.

Companies: Google, Bing, Elastic, Algolia.

2. Text Classification

Assigning predefined categories to text. This is one of the most common and commercially valuable NLP tasks.

• Spam Detection: Gmail processes 300 billion emails daily. Its spam filter uses NLP to classify emails with >99.9% accuracy.
• Sentiment Analysis: Companies like Zomato and Swiggy analyze customer reviews to understand satisfaction. Stock traders analyze Twitter sentiment to predict market movements.
• Topic Classification: News aggregators like Google News automatically categorize articles into "Politics," "Sports," "Technology."
• Support Ticket Routing: When you email customer support, NLP automatically routes your ticket to the right department (billing, technical, returns).

Key Techniques: Naive Bayes, Logistic Regression, SVM, LSTM, BERT fine-tuning.

3. Named Entity Recognition (NER) & Information Extraction

Identifying and extracting specific information from text — names, dates, locations, organizations, amounts.

Example: Consider this news headline: "Apple CEO Tim Cook announced a new $1 billion factory in Bangalore on January 15, 2026." An NER system would extract:

• Organization: Apple
• Person: Tim Cook
• Title: CEO
• Money: $1 billion
• Location: Bangalore
• Date: January 15, 2026

Real-World Use: Banks extract transaction details from emails. Hospitals extract diagnoses and medications from doctor's notes. Law firms extract case details from legal documents.

4. Machine Translation

Converting text from one language to another. This is what powers Google Translate, DeepL, and Microsoft Translator.

Evolution: Early systems used rule-based translation (word-by-word with grammar rules). Statistical systems (2010s) used parallel corpora. Modern Neural Machine Translation (NMT) uses sequence-to-sequence models with attention, producing fluent, natural-sounding translations.

Key Challenge: Languages are not one-to-one. Hindi "जा रहा हूँ" has no direct English equivalent — it combines "go + ing + I am." The model must learn these structural differences.

5. Text Summarization

Condensing long documents into shorter versions while preserving key information.

• Extractive: Selecting the most important sentences from the original text. Fast but less natural.
• Abstractive: Generating new sentences that capture the meaning. Uses LLMs like GPT. More natural but can hallucinate facts.

Example: TLDR; browser extensions use abstractive summarization to summarize web articles. News services like Inshorts use extractive summarization for quick reads.

6. Question Answering & Chatbots

Systems that answer questions based on documents or general knowledge.

Types:

• Closed-domain QA: Answers from a specific document (e.g., "What is the refund policy?" answered from a company's FAQ page)
• Open-domain QA: Answers from general knowledge (e.g., "Who wrote the Indian Constitution?")
• Conversational AI: ChatGPT, Claude, Gemini — can maintain context across multiple turns of conversation

7. Speech Recognition & Synthesis

Converting speech to text (ASR) and text to speech (TTS). Siri, Alexa, Google Assistant all use NLP pipelines that start with speech recognition.

Key Challenge: Accents, background noise, homophones ("write" vs "right"), and code-switching (Hindi + English in the same sentence, common in India).

What is Lexical Ambiguity?

Lexical ambiguity occurs when a single word (or "lexeme") has more than one meaning. In linguistics, these different meanings are called senses or word senses.

According to the Oxford English Dictionary, the average English word has about 2.5 meanings. Some words have dozens! The word "run" has over 600 senses in English.

Types of Lexical Ambiguity

• Homonymy: Words that share the same spelling and pronunciation but have completely unrelated meanings. "Bank" (financial institution vs. river edge) and "Bat" (cricket equipment vs. flying mammal).
• Polysemy: Words with related but distinct meanings. "Foot" can mean the body part, the bottom of a mountain, or a unit of measurement (12 inches). These senses are conceptually related.
• Homophony: Words that sound the same but are spelled differently. "Right" (correct) vs. "Write" (put words on paper) vs. "Rite" (ceremony). Speech recognition systems must handle this.

Real Examples That Break NLP Systems

How Do Humans Disambiguate?

Humans use several cues automatically:

• Contextual words: If you hear "bank" near "money," "deposit," or "account," you think financial. Near "river," "water," or "fishing," you think river edge.
• World knowledge: You know fish don't live in financial buildings. You know people don't swim in banks.
• Grammatical role: "Book" as a verb ("Book a table") vs. noun ("Read a book") has different grammatical positions.

How Do Computers Disambiguate?

There are several computational approaches:

• Word Sense Disambiguation (WSD): Using machine learning to classify which sense is intended. Features include surrounding words, part-of-speech, and word embeddings.
• Contextual Embeddings (Modern): BERT and GPT models generate different vector representations for the same word depending on context. "Bank" in "river bank" gets a different embedding than "bank" in "bank account." This is the state-of-the-art approach.
• Knowledge Bases (Classical): WordNet is a lexical database that organizes words into synsets (sets of synonyms). "Bank" has separate synsets for financial institution and river edge.
• Supervised Learning: Train a classifier on labeled examples where each occurrence of "bank" is tagged with the correct sense. Features: surrounding words, POS tags, etc.

What is Syntactic Ambiguity?

Syntactic ambiguity (also called structural ambiguity) occurs when a sentence has more than one valid grammatical structure, leading to different interpretations. Unlike lexical ambiguity (where a single word is ambiguous), syntactic ambiguity involves the entire sentence structure.

Classic Examples

Example 1: Prepositional Phrase Attachment

"I saw the man with binoculars."

Two possible parses:

• Parse A: I used binoculars to see the man. ["with binoculars" modifies "saw"]
• Parse B: I saw a man who had binoculars. ["with binoculars" modifies "man"]

This is the most common type of syntactic ambiguity — deciding which word a prepositional phrase attaches to.

Example 2: Coordination Ambiguity

"Old men and women sat on the bench."

• Parse A: [Old men] and [women] sat on the bench. (Only the men are old)
• Parse B: Old [men and women] sat on the bench. (Both are old)

Example 3: Scope Ambiguity

"Everyone loves someone."

• Reading A: For every person, there exists someone they love. (Everyone loves at least one person — possibly different people)
• Reading B: There exists one person who is loved by everyone. (One universal beloved — like a celebrity)

Example 4: VP Attachment

"Visiting relatives can be boring."

• Reading A: [Visiting relatives] can be boring. (The act of visiting relatives is boring — "visiting" is a gerund/noun)
• Reading B: Visiting [relatives can be boring]. (Relatives who visit can be boring — "visiting" is a participle/adjective)

Parse Trees: Visualizing Syntactic Structure

A parse tree (or syntax tree) represents the grammatical structure of a sentence. When a sentence is syntactically ambiguous, it has multiple valid parse trees.

Types of Syntactic Ambiguity

How NLP Systems Handle Syntactic Ambiguity

• Probabilistic Context-Free Grammars (PCFG): Assign probabilities to each grammar rule. The parse with the highest probability is chosen. "The man with binoculars" is more likely to modify "man" than "saw" because people carrying binoculars is common.
• Dependency Parsing: Instead of full parse trees, identify direct relationships between words (who did what to whom). Modern parsers like SpaCy's use neural networks to predict dependencies.
• Neural Parsing (Modern): Transformers like BERT can implicitly resolve syntactic ambiguity through contextual understanding. The model learns that "with binoculars" modifying "man" is more likely in certain contexts.

What is Semantic Ambiguity?

Semantic ambiguity occurs when a sentence has a clear grammatical structure, but its meaning is unclear because it requires knowledge about the world, context, or the speaker's intent.

Unlike lexical ambiguity (ambiguous word) or syntactic ambiguity (ambiguous grammar), semantic ambiguity is about meaning — and meaning is deeply tied to what we know about the world.

Types of Semantic Ambiguity

1. Idioms and Figurative Language

Idioms cannot be understood by analyzing the individual words. The meaning is conventional — you must know the idiom as a whole.

• "He kicked the bucket." → He died (not that he literally kicked a bucket)
• "It's raining cats and dogs." → It's raining heavily (not literal animals)
• "Break a leg!" → Good luck (especially in theatre)
• "The ball is in your court." → It's your turn to act

Challenge for NLP: A literal NLP system might try to find "bucket" in the "kick the bucket" sentence, or look for falling animals in "raining cats and dogs." Only systems trained on idiomatic expressions can handle this.

2. Metaphor and Analogy

"Time is money." → Time is valuable and should be spent wisely. This is not a literal statement about time being a currency.

"The CEO is a shark." → The CEO is aggressive in business, not that they are a literal marine animal.

Challenge for NLP: Metaphor detection requires understanding that "shark" in a business context has a different meaning from "shark" in an ocean context. LLMs handle this reasonably well because they've seen millions of metaphorical usages in training data.

3. Sarcasm and Irony

Sarcasm is when the literal meaning is the opposite of the intended meaning. This is one of the hardest problems in NLP.

• "Great job missing the deadline." → The speaker is NOT praising the person. They are criticizing them.
• "Oh, I LOVE waiting in traffic for 2 hours." → The speaker hates traffic.
• "Yeah, that's exactly what I needed today." (after spilling coffee on laptop) → The speaker is frustrated, not satisfied.

Why it's hard: Sarcasm requires understanding the speaker's intent, emotional state, and the situational context. Even humans sometimes miss sarcasm in text (without tone of voice, it's harder). State-of-the-art sarcasm detection systems achieve only ~70-80% accuracy.

4. Presupposition and Entailment

Some sentences assume background knowledge that isn't explicitly stated.

• "John stopped smoking." → Presupposes that John used to smoke. If he never smoked, the sentence is odd.
• "Have you stopped beating your dog?" → This is a classic loaded question. It presupposes you were beating your dog.

5. Entity Ambiguity (Named Entity Disambiguation)

"Apple released a new product." → "Apple" could be the company (Apple Inc.) or the fruit. In a tech news context, it's the company. In a cooking blog, it might be the fruit.

How NLP Systems Handle Semantic Ambiguity

• WordNet and Knowledge Graphs: WordNet groups words by meaning. "Bank" as a financial institution is in a different synset from "bank" as a river edge. Knowledge graphs like Wikidata encode relationships between entities.
• Contextual Embeddings (BERT, GPT): These models learn semantic meaning from massive training data. "Apple" in "Apple released iPhone" gets a different embedding than "Apple" in "Apple pie recipe." The model learns from seeing millions of examples.
• Sentiment Analysis as a Proxy: For sarcasm detection, some systems check if the literal sentiment (positive) contradicts the expected sentiment (negative given context). If the text is positive but the situation is clearly negative, flag as sarcasm.
• Multi-Modal Approaches: Sarcasm is easier to detect when you have audio (tone of voice) or video (facial expressions). Text-only sarcasm detection is much harder.

What is Pragmatics?

Pragmatics is the study of how context contributes to meaning. It's not about what words mean in a dictionary (that's semantics), or how they're arranged grammatically (that's syntax). It's about what speakers INTEND to communicate in specific situations.

In other words: semantics tells you what a sentence means; pragmatics tells you what the speaker means by it.

The Four Levels of NLP Analysis

To understand pragmatics, you need to see where it fits in the broader NLP stack:

Key Concepts in Pragmatics

1. Speech Acts (What the speaker is DOING)

Every utterance is an action. Linguist J.L. Austin identified three types:

• Locutionary act: The literal meaning. "Can you pass the salt?" = asking about ability to pass salt.
• Illocutionary act: The intended action. "Can you pass the salt?" = REQUESTING salt (not asking about ability).
• Perlocutionary act: The actual effect. The listener passes the salt.

Why this matters for NLP: A chatbot that answers "Yes, I can pass the salt" to "Can you pass the salt?" has failed at the illocutionary level. It understood the words but missed the intent.

2. Conversational Implicature (Reading between the lines)

Philosopher H.P. Grice proposed that speakers follow a "Cooperative Principle" — they make contributions that are:

• Relevant: On-topic
• Informative: Saying enough, but not too much
• Truthful: Not saying false things
• Clear: Not obscure or ambiguous

When speakers violate these maxims, they create implicature:

"Your code is... interesting." → If the speaker says this with a weird tone, they might mean "Your code is terrible" but are being polite. The implicature is the opposite of the literal meaning.

3. Anaphora Resolution (What does "it" refer to?)

Pronouns and references require tracking across sentences:

"Rahul told Arjun that he failed the exam." → Who failed? Rahul or Arjun? This is anaphora resolution, and it's a core NLP task.

"The dog chased the cat. It barked." → "It" refers to the dog (the cat meows, doesn't bark). World knowledge helps resolve this.

4. Discourse Coherence (How sentences connect)

Individual sentences make sense, but discourse requires understanding how they relate:

"John loves movies. He went to the cinema." → Cause-effect? The second sentence explains what John did because he loves movies.

"John loves movies. He hates popcorn." → Contrast? These are independent facts about John.

How Modern NLP Handles Pragmatics

• Intent Classification: In chatbots, the first step is classifying user intent: "book_flight," "check_status," "complaint," "greeting." This is a pragmatic task.
• Coreference Resolution: Specialized models (like SpanBERT) identify what pronouns and references point to. Essential for document understanding and question answering.
• Dialogue State Tracking: In multi-turn conversations, track what the user has said, what they want, and what information has been provided. "What time?" — the system must remember they were booking a flight.
• Context Windows (LLMs): GPT-4, Claude, and Gemini maintain context across thousands of tokens. This lets them resolve anaphora, track discourse, and infer intent across long conversations.

The NLP Task Hierarchy

NLP tasks are organized by level of linguistic analysis and by output type. Here's the complete taxonomy:

1. Word-Level Tasks

Tasks that analyze individual words:

• Tokenization: Splitting text into words, subwords, or characters. The foundation of everything.
• Stemming: Reducing words to their root form ("running" → "run"). Crude but fast.
• Lemmatization: Reducing to dictionary form ("better" → "good"). Uses POS tags and dictionaries.
• Part-of-Speech Tagging (POS): Labeling each word with its grammatical role (noun, verb, adjective).
• Morphological Analysis: Analyzing word structure — prefixes, suffixes, roots ("unhappiness" = un + happy + ness).
• Word Sense Disambiguation (WSD): Determining which meaning of a word is intended ("bank" = financial or river).

2. Sequence-Level Tasks

Tasks that analyze sequences of words:

• Named Entity Recognition (NER): Identifying entities like people, organizations, locations, dates. "Barack Obama visited Paris" → PER: Barack Obama, LOC: Paris.
• Chunking / Shallow Parsing: Grouping words into phrases (noun phrases, verb phrases). "[The quick brown fox] [jumped] [over the lazy dog]."
• Dependency Parsing: Finding grammatical relationships between words. "fox" → subject of "jumped"; "dog" → object of "over."
• Constituency Parsing: Building full parse trees showing nested phrase structure.

3. Document-Level Tasks

Tasks that analyze entire documents or sentences:

• Text Classification: Assigning a category. Sentiment analysis (positive/negative), spam detection, topic classification.
• Text Similarity: Measuring how similar two texts are. Cosine similarity, semantic similarity, paraphrase detection.
• Text Summarization: Extractive (selecting sentences) or abstractive (generating new text).
• Question Answering: Extracting answers from documents or generating them from knowledge.
• Machine Translation: Converting text between languages.
• Text Generation: Creating new text from prompts. Creative writing, code generation, dialogue.

4. Dialogue and Interactive Tasks

• Intent Classification: Determining what the user wants.
• Slot Filling / Entity Extraction: Extracting parameters for the intent ("Book a flight from Mumbai to Delhi on March 15" → origin, destination, date).
• Dialogue State Tracking: Maintaining context across conversation turns.
• Response Generation: Generating appropriate replies.

5. Knowledge-Level Tasks

• Relation Extraction: Finding relationships between entities. "Steve Jobs founded Apple" → (Steve Jobs, founded, Apple).
• Knowledge Graph Construction: Building structured knowledge from text.
• Entailment / Natural Language Inference:: Determining if one sentence logically follows from another. "All cats are mammals. Garfield is a cat." → "Garfield is a mammal."

The Three Pillars of Modern AI

Machine Learning has three major subfields that dominate industry applications:

Deep Dive: NLP

What makes NLP unique:

• Discrete input space: Text is discrete (words are categorical), not continuous like pixels. This requires embedding layers to convert words to vectors.
• Variable-length input: A sentence can be 3 words or 300 words. Models must handle variable-length sequences.
• Long-range dependencies: "The cat, which was very hungry and had been searching for hours, finally ate the fish." The subject "cat" and verb "ate" are far apart. Transformers handle this with self-attention.
• Compositionality: Meaning is built from smaller parts (words → phrases → sentences → paragraphs). Understanding requires hierarchical processing.

Data characteristics: Text data is everywhere — books, websites, social media. But it's noisy, unstructured, and often unlabeled. High-quality labeled data is expensive to create.

Deep Dive: Computer Vision

What makes CV unique:

• Continuous input space: Pixels are continuous values (0-255). This makes convolution operations natural and effective.
• Spatial structure: Nearby pixels are related. CNNs exploit this with local receptive fields and weight sharing.
• Translation invariance: A cat in the top-left corner is still a cat. CNNs are naturally translation-invariant.
• High dimensionality: A 1080p image has 2,073,600 pixels. This requires efficient architectures.

Data characteristics: Images are structured and dense, but labeling is expensive (need bounding boxes, segmentation masks). Data augmentation (rotation, flipping, cropping) is very effective for CV.

Deep Dive: Reinforcement Learning

What makes RL unique:

• No labeled dataset: The agent learns by interacting with an environment and receiving rewards or penalties.
• Delayed rewards: A chess move might not show its value until 20 moves later. Credit assignment is hard.
• Exploration vs exploitation: Should the agent try a new action (explore) or stick with what works (exploit)? This is the core tradeoff.
• Environment dynamics: The agent's actions change the environment, which affects future decisions.

Data characteristics: RL generates its own data through interaction. No need for pre-labeled datasets, but training is slow and unstable. Simulators are often used (e.g., game environments, robot simulations).

Where They Overlap: Multi-Modal AI

Modern AI systems combine all three:

• Vision-Language Models (VLM): GPT-4V, Claude, Gemini can look at images and describe them in text. They combine CV encoders (process image) with NLP decoders (generate text).
• Self-Driving Cars: CV (recognize objects) + RL (make driving decisions) + NLP (understand voice commands, read signs).
• Robotics: CV (see the world) + RL (learn to grasp/manipulate) + NLP (follow instructions like "pick up the red block").
• AI Agents: LLM (reasoning and planning) + tools (search, code execution) + RL (learn from feedback). This is the hottest area in AI right now.

How to Choose the Right Approach