NLP: Syntax, Semantics, Pragmatics & Applications

Natural Language Processing (NLP): Core Concepts and Applications

1. What is NLP?

Natural Language Processing (NLP) is a subfield of AI that enables computers to understand, interpret, and generate human language. It bridges linguistics, computer science, and AI to process text or speech in a way that is meaningful to humans.

Why is NLP Important?

• Human-Computer Interaction: Powers chatbots, voice assistants (e.g., Google Assistant, Siri), and translation tools.
• Automation: Extracts insights from unstructured data (e.g., customer reviews, legal documents).
• Accessibility: Helps people with disabilities interact with technology via speech.

2. Key Components of NLP

NLP breaks down language processing into three layers:

A. Syntax (Structure of Language)

Syntax focuses on grammar and sentence structure. It answers:

• "How are words arranged?"
• "Is the sentence grammatically correct?"

Example:

• Input: "The cat sat on the mat."
• Syntax checks: Subject (The cat) + Verb (sat) + Prepositional phrase (on the mat).

Visual: Syntax Tree

graph TD
    S["Sentence"] --> NP1["NP: The cat"]
    S --> VP["VP: sat on the mat"]
    NP1 --> DT["Det: The"]
    NP1 --> N["Noun: cat"]
    VP --> V["Verb: sat"]
    VP --> PP["PP: on the mat"]
    PP --> P["Prep: on"]
    PP --> NP2["NP: the mat"]
    NP2 --> DT2["Det: the"]
    NP2 --> N2["Noun: mat"]

B. Semantics (Meaning of Language)

Semantics deals with meaning. It answers:

• "What does the sentence mean?"
• "Is the meaning logical?"

Example:

• Input: "The cat is on the mat."
• Semantics extracts: A feline (cat) is physically located on a piece of furniture (mat).

Visual: Semantic Graph

graph TD
    A["cat"] -->|"is_a"| B["animal"]
    A -->|"location"| C["mat"]
    C -->|"is_a"| D["furniture"]

C. Pragmatics (Context and Intent)

Pragmatics considers context, speaker intent, and real-world knowledge. It answers:

• "What does the speaker really mean?"
• "Is there an implied meaning?"

Example:

• Input: "It’s cold in here."
  • Literal meaning: Temperature is low.
  • Pragmatic meaning: "Close the window."

Why Pragmatics Matters in NLP:

• Helps AI understand sarcasm, humor, or indirect requests.
• Critical for customer service bots (e.g., Daraz’s chatbot resolving complaints).

3. Real-World Applications of NLP

A. In Nepal

B. Global Examples

Worked Example: Kathmandu Traffic Route Suggestion

• Input: "How to reach Thamel from my location?"
• NLP Pipeline:
  1. Tokenization: Split into words: "How", "to", "reach", "Thamel", "from", "my", "location?"
  2. Intent Classification: Detects "route_query" intent.
  3. Entity Extraction: "Thamel" (destination), "my location" (source).
  4. Pragmatic Response: Uses Google Maps API to suggest the fastest path (avoiding traffic jams).

4. NLP Techniques and Tools

A. Tokenization

Splitting text into words, phrases, or tokens. Example: Input: "I love NLP!" Output: ["I", "love", "NLP", "!"]

B. Stemming & Lemmatization

• Stemming: Reduces words to root form (crude). "running" → "run"
• Lemmatization: Uses vocabulary (more accurate). "better" → "good"

Visual: Stemming vs. Lemmatization

graph LR
    A["running"] -->|"Stemmer"| B["run"]
    A -->|"Lemmatizer"| C["run"]
    D["better"] -->|"Stemmer"| E["good"]
    D -->|"Lemmatizer"| F["good"]

C. Part-of-Speech (POS) Tagging

Labels words by grammatical role. Example: "The cat sat on the mat." → ["Det", "Noun", "Verb", "Prep", "Det", "Noun"]

D. Named Entity Recognition (NER)

Identifies people, places, organizations. Example: "Ramesh works at Ncell in Kathmandu." → ["Person: Ramesh", "Organization: Ncell", "Location: Kathmandu"]

E. Sentiment Analysis

Classifies text as positive, negative, or neutral. Example:

• "This phone is amazing!" → Positive
• "The service was terrible." → Negative

Visual: Sentiment Analysis Pipeline

flowchart LR
    A["Raw Text"] --> B["Preprocessing"]
    B --> C["Tokenization"]
    C --> D["Feature Extraction"]
    D --> E["Machine Learning Model"]
    E --> F["Sentiment Score"]
    F --> G["Positive/Negative/Neutral"]

5. Challenges in NLP

6. NLP in Machine Translation

A. Rule-Based vs. Statistical vs. Neural MT

Worked Example: Nepali → English Translation

• Input: "मेरो नाम रमेश हो। म काठमाडौँमा बस्छु।"
• Neural MT Steps:
  1. Tokenize: ["मेरो", "नाम", "रमेश", "हो", "म", "काठमाडौँमा", "बस्छु", "."]
  2. Embed words into vectors.
  3. Transformer model predicts: "My name is Ramesh. I live in Kathmandu."

7. NLP in Chatbots and Virtual Assistants

A. How Chatbots Work

flowchart LR
    A["User Input"] --> B["Preprocess"]
    B --> C["Intent Recognition"]
    C --> D["Entity Extraction"]
    D --> E["Dialogue Manager"]
    E --> F["Response Generation"]
    F --> G["User"]

Example: Daraz Customer Support Bot

• User: "My order #12345 is delayed."
• Bot Steps:
  1. Intent: "order_status_query"
  2. Entity: "order_id: 12345"
  3. Response: "Your order is out for delivery. ETA: 2 hours."

B. Limitations

• Lack of Context: Struggles with multi-turn conversations.
• Cultural Nuances: Nepali humor/slang may confuse global models.

8. Future of NLP

• Multimodal NLP: Combines text + images + speech (e.g., describing a photo).
• Zero-Shot Learning: Models that understand new languages/tasks without training.
• Ethical NLP: Reducing bias in chatbots (e.g., gender-neutral responses).

In the Real World

1. eSewa’s Automated Responses

   • Idea Used: Intent classification + NER
   • How: When you type "How to pay traffic fine?", eSewa’s bot extracts:
     • Intent: "payment_inquiry"
     • Entity: "traffic fine"
     • Response: Guides you to the payment portal.

2. Khalti’s Fraud Detection

   • Idea Used: Sentiment analysis + anomaly detection
   • How: Flags suspicious transactions by analyzing:
     • Unusual phrasing in chat messages (e.g., "Send money to this number").
     • Sudden large transfers from a user’s account.

3. Nepali Newspaper Summarization (e.g., Kantipur, Republica)

   • Idea Used: Text summarization (extractive/abstractive)
   • How: Converts a 500-word article into 5 bullet points using:
     • Keyword extraction (e.g., "lockdown", "supply chain").
     • Sentence importance scoring.

Exam Tip

How This Unit is Tested

1. Definitions (10%)

   • Expect: "Define NLP. Explain syntax vs. semantics vs. pragmatics."
   • Key Points to Remember:
     • Syntax = grammar.
     • Semantics = meaning.
     • Pragmatics = context + intent.

2. Applications (30%)

   • Common Questions:
     • "How does WhatsApp Business use NLP?" → Intent recognition + NER.
     • "Explain how Google Translate works." → Neural MT + attention mechanisms.
   • Tip: Always relate to Nepali examples (eSewa, Khalti, Daraz).

3. Techniques (40%)

   • Must-Know Steps:
     • Tokenization → POS Tagging → NER → Sentiment Analysis.
   • Worked Example: Given a sentence, show tokenization + POS tagging.

4. Challenges (20%)

   • Expected Answer:
     • Ambiguity → Use context.
     • Slang → Train on social media data.
     • Multilingual → Use language-specific models.

Model Answer for Past Exam Question

Question: "Describe about natural language processing. Why is pragmatic analysis important in language processing?"

Answer: Natural Language Processing (NLP) is an AI subfield that enables computers to understand, interpret, and generate human language. It involves three key layers:

1. Syntax: Analyzes sentence structure (e.g., "The cat sat on the mat" → Subject-Verb-Object).
2. Semantics: Extracts meaning (e.g., "cat" refers to a feline).
3. Pragmatics: Interprets context and intent (e.g., "It’s cold!" may mean "Close the window").

Why Pragmatic Analysis is Important:

• Real-World Context: Humans rarely speak literally. Pragmatics helps AI understand:
  • Implied meanings (e.g., "You’re late" → "Hurry up").
  • Politeness (e.g., "Could you help?" vs. "Help me!").
• Chatbot Effectiveness: Without pragmatics, a bot might fail to:
  • Handle sarcasm (e.g., "Oh great, another power cut.").
  • Resolve ambiguity (e.g., "I saw the man with binoculars" → Who has them?).
• Nepali Applications: In eSewa chatbots, pragmatic analysis helps:
  • Detect urgent requests (e.g., "My fine is overdue!").
  • Provide empathetic responses (e.g., "I’m sorry for the delay.").

Example:

• Input: "The meeting is postponed."
  • Syntax: Subject (meeting) + Verb (is postponed).
  • Semantics: The meeting will happen later.
  • Pragmatics: "Please reschedule your plans."

Final Checklist for Full Marks

✅ Define NLP (1 mark). ✅ Explain syntax, semantics, pragmatics (3 marks). ✅ Give 2 real-world examples (eSewa/Khalti) (4 marks). ✅ Discuss challenges (ambiguity, slang) (3 marks). ✅ Relate to Nepali context (1 mark).

A diagram of a Transformer’s encoder-decoder layers with attention mechanisms. (Image: Yuening Jia, CC BY-SA 3.0, via Wikimedia Commons)