NLP Core: PEAS, Parsing, Semantics, Pragmatics & Challenges

What is NLP?

Natural Language Processing (NLP) is the field of AI that enables computers to understand, generate, and manipulate human language. It bridges linguistics and computer science to build systems that can:

• Interpret user queries (e.g., "Book a flight to Kathmandu")
• Generate human-like responses (e.g., chatbots)
• Translate languages (e.g., Google Translate)
• Extract structured data from text (e.g., eSewa parsing bills)

Why Machines Need NLP

Humans communicate primarily through language (90%+ of data is unstructured text). Machines need NLP to:

1. Interact naturally (e.g., voice assistants like Google Assistant).
2. Automate tasks (e.g., eSewa processing payments from text).
3. Extract insights (e.g., Ncell analyzing customer complaints).

PEAS Framework for Intelligent Agents

The Performance, Environment, Actuators, Sensors framework defines an agent’s operational context. Let’s analyze two agents:

Example 1: Internet Shopping Assistant (e.g., Daraz)

• Performance: Must recommend relevant products (90%+ accuracy) and process orders in <2s.
• Environment: Dynamic (prices change, trends shift).
• Actuators: Sends recommendations, confirms orders.
• Sensors: User search queries, product catalog, reviews.

Example 2: English Language Tutor (e.g., Duolingo)

• Performance: Corrects 95% of grammar errors, teaches 1000+ words/year.
• Environment: Personalized (adapts to user mistakes).
• Actuators: Provides hints, quizzes.
• Sensors: Records user responses, tracks mistakes.

In the Real World

1. eSewa: Uses NLP to parse payment instructions from SMS/text (e.g., "Pay Rs. 500 to ABC for electricity"). The semantic analysis step extracts the amount, recipient, and service type.
2. Khalti: When you say, "Pay Rs. 1000 to my brother," the pragmatic analysis resolves "brother" to a contact in your phone’s address book.
3. Pathao Driver App: The syntactic parser breaks down "Take me to Thamel" into subject (user), verb (go), and object (Thamel), then maps it to GPS coordinates.

NLP Pipeline: From Text to Meaning

The NLP pipeline converts raw text into structured data. Here’s how it works for the query: "Book a flight to Kathmandu for Rs. 15,000 on 15th June."

Step 1: Tokenization

Split text into words/tokens: ["Book", "a", "flight", "to", "Kathmandu", "for", "Rs.", "15,000", "on", "15th", "June"]

Step 2: Morphological Analysis

Break words into stems and identify parts of speech (POS):

Step 3: Syntactic Analysis (Parsing)

Build a parse tree to understand sentence structure:

graph TD
    S["Sentence"] --> NP1["NP: Book"]
    S --> VP["VP: flight to Kathmandu"]
    VP --> V["V: book"]
    VP --> PP1["PP: to Kathmandu"]
    PP1 --> P["P: to"]
    PP1 --> NP2["NP: Kathmandu"]
    VP --> PP2["PP: for Rs. 15,000 on 15th June"]
    PP2 --> P2["P: for"]
    PP2 --> NP3["NP: Rs. 15,000"]
    PP2 --> PP3["PP: on 15th June"]

Key Insight: The verb "book" governs the flight (direct object) and the time/money (prepositional phrases).

Step 4: Semantic Analysis

Extract meaning using a knowledge base (e.g., flight database):

• Entities:
  • Action: book_flight
  • Destination: Kathmandu
  • Price: 15000 NPR
  • Date: 2024-06-15
• Relations:
  • book_flight(destination=Kathmandu, price=15000, date=2024-06-15)

Step 5: Pragmatic Analysis

Resolve ambiguity and context:

• If you’re logged into Yeti Airlines, the system assumes you want a Yeti flight.
• If you say "Book a flight to Kathmandu for my mom," the system may ask: "Which ticket: adult or child?"

Challenges in NLP

Real Picture: NLP in Action

A screenshot of Google Translate converting "मेरो नाम सूर्य हो" to "My name is Surya." Caption: Semantic and syntactic analysis at work: the system maps Nepali words to English equivalents while preserving grammar.

Worked Example: Parsing a Nepali Sentence

Input: "मेरो बाइकले पेट्रोल भर्नु पर्छ।" (My bike needs petrol.)

Step 1: Tokenization

["मेरो", "बाइकले", "पेट्रोल", "भर्नु", "पर्छ", "।"]

Step 2: POS Tagging

Step 3: Dependency Parse Tree

Translation: need(petrol, my_bike)

Syntactic vs. Semantic Analysis

Natural Language Generation (NLG) vs. Understanding

Morphological Analysis in Nepali

Nepali is highly inflected, meaning words change based on context. Example:

• "बाटो" (bāṭo) can mean:
  • Noun: "road" (बाटो = road)
  • Verb: "to go" (बाटो = from, postposition)
  • Adjective: "from which" (बाटो = postposition)

Worked Example: Analyzing "बाटोमा"

1. Token: "बाटोमा"
2. Breakdown:
   • Stem: "बाटो" (root: "बाट" = path)
   • Suffix: "मा" (locative case marker)
3. Meaning: "on/at the road"

Visual:

Exam Tip

1. PEAS Framework: Always define Performance (metrics), Environment (context), Actuators (outputs), and Sensors (inputs). Use real examples (e.g., eSewa for actuators = payment confirmation).
2. NLP Pipeline: Memorize the 5 steps (tokenization → morphology → syntax → semantics → pragmatics) and one example per step. For syntax, draw a parse tree (even if simple).
3. Challenges: Link challenges to real-world failures:
   • Ambiguity → "I shot an elephant in my pajamas" (was the elephant in pajamas?).
   • Context → "The stock market crashed. He died." (Who is "he"?).
4. Comparisons: For NLG vs. NLU, use a table and one sentence per cell.
5. Nepali Focus: Highlight morphological complexity (e.g., case markers, verb conjugations) and low-resource challenges (e.g., lack of labeled data).

Summary Checklist

• Can define PEAS for two agents (e.g., Daraz + tutor).
• Can draw a parse tree for a simple sentence.
• Knows 3 NLP challenges + solutions.
• Can tokenize and POS-tag a Nepali sentence.
• Understands NLG vs. NLU with examples.