NLP Techniques: Parsing, POS Tagging, NER & Semantic Analysis

TAKEAWAYS:

• Parsing breaks sentences into syntactic trees (e.g., constituency vs. dependency parsing) to enable machine understanding of grammar.
• POS tagging assigns grammatical labels (noun, verb, adjective) to words, critical for disambiguation in Nepali/English (e.g., "run" as noun vs. verb).
• Named Entity Recognition (NER) extracts real-world entities (people, dates, locations) from text, powering search and recommendation systems like Daraz or Ncell.
• Semantic analysis resolves ambiguity (e.g., "bank" as financial vs. river) using ontologies (OWL) and word embeddings (Word2Vec, GloVe).
• Morphology (word structure) and lexicon (vocabulary) are the building blocks for accurate NLP pipelines in Nepali (e.g., handling sandhi rules).
• Real-world tie-ins: WhatsApp’s message parsing, eSewa’s NER for transaction validation, and Google’s semantic search rely on these techniques.

Core NLP Techniques: How Machines Understand Language

1. Morphology and Lexicon: The Building Blocks

NLP starts with morphology (studying word forms) and lexicon (vocabulary). In Nepali, morphology is complex due to:

• Sandhi rules: Word combinations change forms (e.g., "राम + लाई" → "रामलाई").
• Inflection: Verbs change based on tense/person (e.g., "खानु" → "खाइस", "खान्छ").

Worked Example: Nepali Verb Conjugation Sentence: "मेरो पुस्तक खान्छ" (My book eats) → Incorrect. Correct: "मेरो पुस्तक खाइन्छ" (My book is eaten). Trace:

1. Identify root: "खानु" (to eat).
2. Apply passive rule: "खाइन्छ" (is eaten).
3. Assign POS: "खाइन्छ" → verb (passive, present).

graph LR
    A["खानु (root)"] --> B["खान्छ (active)"] --> C["खाइन्छ (passive)"]
    A --> D["खाइस (past)"] --> E["खाइएको (past participle)"]

2. Part-of-Speech (POS) Tagging: Labeling Words

POS tagging assigns grammatical labels to words. Example for English: Sentence: "The quick brown fox jumps over the lazy dog." Tags: DET ADJ ADJ NOUN VERB PREP DET ADJ NOUN .

Nepali Example: Sentence: "रामले पुस्तक पढ्यो।" Tags: PROPER VERB NOUN VERB . (रामले = subject, पुस्तक = object, पढ्यो = verb).

Why It Matters:

• Disambiguates homographs (e.g., "run" as noun vs. verb).
• Enables syntax parsing (next step).

Worked Example: Nepali POS Tagging Sentence: "सुरुमा मलाई थाहा थिएन।"

1. Tokenize: ["सुरुमा", "मलाई", "थाहा", "थिएन", "।"]
2. Tag:
   • सुरुमा → ADV (adverb)
   • मलाई → PRON (pronoun)
   • थाहा → NOUN (noun)
   • थिएन → VERB (past negative)
   • । → PUNCT

graph TD
    A["सुरुमा\nADV"] --> B["मलाई\nPRON"]
    B --> C["थाहा\nNOUN"]
    C --> D["थिएन\nVERB"]
    D --> E["।\nPUNCT"]

3. Parsing: Building Syntax Trees

Parsing converts sentences into structured trees to understand relationships. Two main types:

1. Constituency Parsing: Hierarchical phrase structure (like a family tree). Example: "रामले पुस्तक पढ्यो" →

   S
   ├── NP: रामले
   └── VP
       ├── V: पढ्यो
       └── NP: पुस्तक
   

2. Dependency Parsing: Word-to-word relationships (who does what to whom). Example:

   पढ्यो (root)
   ├── subject: रामले
   └── object: पुस्तक
   

Worked Example: Nepali Dependency Parse Sentence: "मेरो गाडी खराब भएको छ।"

1. Tokenize: ["मेरो", "गाडी", "खराब", "भएको", "छ", "।"]
2. Dependencies:
   • भएको (root)
     • subject: गाडी
     • modifier: मेरो
     • copula: छ
     • adjective: खराब

graph TD
    A["भएको\n(root)"] --> B["गाडी\n(subject)"]
    A --> C["छ\n(copula)"]
    B --> D["मेरो\n(modifier)"]
    A --> E["खराब\n(adjective)"]

4. Named Entity Recognition (NER): Extracting Real-World Entities

NER identifies and classifies entities like:

• Person: राम, Barack Obama
• Location: काठमाडौं, New York
• Organization: NTC, Google
• Date: २०८० चैत १, 2023-10-05
• Money: ₹५००, $100

Real-World Use in Nepal:

• eSewa: Extracts user names, transaction IDs, and amounts from chat inputs.
• Ncell: Identifies phone numbers in customer queries for routing.
• Daraz: Pulls product names, prices, and seller info from reviews.

Worked Example: Nepali NER Sentence: "रामले २०८० चैत १ मा काठमाडौंमा ५००० रुपैयाँ जमा गर्यो।" Entities:

5. Semantic Analysis: Beyond Words to Meaning

Semantics resolves ambiguity and links words to real-world knowledge. Key techniques:

1. Word Sense Disambiguation (WSD):
   • "Bank" → financial (NTC) vs. river (Ganga).
   • Tools: WordNet, BabelNet.
2. Ontologies (OWL):
   • Links "लेखक" (author) to "व्यक्ति" (person) in Nepali ontologies.
3. Word Embeddings:
   • Word2Vec, GloVe: Represent words as vectors (e.g., "काठमाडौं" close to "नेपाल").

Worked Example: Nepali WSD Sentence: "मेरो बैंकमा पैसा छ।"

• Possible senses for "बैंक":
  1. Financial institution (NTC).
  2. River bank (Ganga).
• Context clues: "पैसा" (money) → NTC (financial).

Visualizing Word Embeddings:

graph TD
    A["नेपाल"] --> B["काठमाडौं"]
    A --> C["पोखरा"]
    B --> D["ललितपुर"]
    C --> D

6. NLP Pipelines: Putting It All Together

A typical NLP pipeline for Nepali text processing:

flowchart LR
    A["Text Input"] --> B["Tokenization"]
    B --> C["POS Tagging"]
    C --> D["Parsing"]
    D --> E["NER"]
    E --> F["Semantic Analysis"]
    F --> G["Output: Structured Data"]

Real-World Pipeline: WhatsApp Chat Analysis

1. Input: "रामले २०८० चैत १ मा ५००० जमा गर्यो।"
2. Tokenization: Split into words.
3. POS Tagging: Identify verbs, dates.
4. NER: Extract राम, २०८० चैत १, ५०००.
5. Semantic Analysis: Link "जमा गर्यो" to financial transaction.
6. Output: Structured data for eSewa/Khalti validation.

In the Real World

1. eSewa/Khalti:

   • NER + Semantic Analysis: Extracts user names, transaction amounts, and service codes (e.g., "बिजुली बिल") from chat inputs to auto-fill forms.
   • Example: Input "रामले २००० बिजुली बिल भुक्तानी गर्नुहोस्" → System identifies:
     • User: राम
     • Amount: २०००
     • Service: बिजुली बिल

2. Daraz Customer Support:

   • Parsing + POS Tagging: Routes queries like "मेरो order ID DZ123456 भएको छ, कहाँ छ?" by extracting:
     • Order ID: DZ123456 (via NER).
     • Intent: status inquiry (via POS + parsing).

3. Ncell Chatbot:

   • Morphology Handling: Corrects Nepali grammar errors (e.g., "मेरो नम्बर कस्तो छ" → "मेरो नम्बर के हो?") before processing.
   • Semantic Search: Matches "तपाईंको नम्बर के हो" to "What’s my number?" using word embeddings.

4. Nepali Wikipedia/BabelNet:

   • Ontology Integration: Links Nepali terms (e.g., "लिच्छवी") to English ("Licchavi") and historical context using OWL-based ontologies.

Comparisons: NLP Techniques vs. Semantic Web

Exam Tip

1. For Short Questions (1+4 marks):

   • Parsing: Always draw a dependency tree for the given sentence. Label nodes clearly (e.g., "subject", "object").
   • POS Tagging: Show tokenization + tags in a table. Example:

     Word	POS
     रामले	PROPN
     पुस्तक	NOUN
     पढ्यो	VERB

2. For Long Questions (5+5 marks):

   • Structure: Use the pipeline flowchart (tokenization → POS → parsing → NER → semantics).
   • Real-World Tie-Ins: Pick one Nepali app (eSewa, Daraz) and explain how two techniques (e.g., NER + semantics) work together.
   • Nepali Examples: Always use Nepali sentences for parsing/POS tagging. Examiners favor localized examples.

3. Common Pitfalls:

   • Forgetting morphology for Nepali (e.g., sandhi rules in POS tagging).
   • Confusing constituency vs. dependency parsing. Always ask: "Does it show phrases or word relationships?"
   • Overlooking semantics in NER (e.g., "bank" can be both financial and geographical).

4. Visuals in Exams:

   • If asked to "explain parsing," draw a dependency tree (even if rough).
   • For NER, use a table with columns: Entity Type | Text | Confidence.
   • For semantics, sketch a simple ontology graph (e.g., "लेखक" → "व्यक्ति").

Pro Tip: Memorize one Nepali sentence for each technique (e.g., POS tagging, parsing) and practice tracing it step-by-step. Examiners often reuse similar examples!