Morphology & Lexical Analysis

Morphology: The Structure of Words

Morphology is the study of the internal structure of words — how words are built from smaller meaningful units called morphemes.

Morphemes

A morpheme is the smallest unit of language that carries meaning:

Type	Example	Breakdown
Free morpheme	"cat", "run"	Can stand alone as a word
Bound morpheme	"-ing", "un-", "-ness"	Must attach to another morpheme
Root	"happi" in "happiness"	Core meaning carrier
Prefix	"un-" in "unhappy"	Attaches before root
Suffix	"-ness" in "happiness"	Attaches after root
Infix	"-bloody-" (informal)	Inserted within a word

Inflectional vs Derivational Morphology

	Inflectional	Derivational
Purpose	Grammatical variation of same word	Creates new words with new meaning
Changes word class?	No	Often yes
Examples	walk → walks, walked, walking	happy → happiness (N), unhappy (Adj)
Number of affixes	Limited (8 in English)	Hundreds

English inflectional suffixes (all 8):

• Noun: -s (plural), -'s (possessive)
• Verb: -s (3rd person), -ed (past), -ing (progressive), -en (past participle)
• Adjective: -er (comparative), -est (superlative)

Why Morphology Matters for NLP

• Reduces vocabulary size: "run", "runs", "running", "ran" → same root
• Handles out-of-vocabulary words: decompose into known morphemes
• Language generation: inflect correctly based on number/tense/gender
• Cross-lingual NLP: morphologically rich languages (Finnish, Turkish, Arabic) have thousands of word forms per lemma

Part-of-Speech (POS) Tagging

POS Tag Sets

The Penn Treebank tagset (45 tags) is the most common for English:

Tag	Meaning	Example
NN	Noun, singular	"dog"
NNS	Noun, plural	"dogs"
VB	Verb, base form	"run"
VBD	Verb, past tense	"ran"
VBG	Verb, gerund	"running"
JJ	Adjective	"fast"
RB	Adverb	"quickly"
DT	Determiner	"the", "a"
IN	Preposition	"in", "on"
PRP	Personal pronoun	"he", "she"
CC	Coordinating conjunction	"and", "but"

Universal POS tags (17 tags) — language-agnostic, used by spaCy and Universal Dependencies: NOUN, VERB, ADJ, ADV, PRON, DET, ADP (prepositions), CONJ, NUM, PUNCT, ...

POS Tagging Approaches

1. Rule-based: handcrafted rules ("if ends in -ing after auxiliary verb → VBG"). Fast but brittle.

2. Hidden Markov Model (HMM):

$P(t_1...t_n | w_1...w_n) \propto \prod_{i=1}^{n} P(w_i | t_i) \cdot P(t_i | t_{i-1})$

• Emission probability $P(w|t)$: likelihood of word given tag
• Transition probability $P(t_i|t_{i-1})$: how likely one tag follows another
• Decoded with Viterbi algorithm in $O(n \cdot |T|^2)$

3. Maximum Entropy / CRF: discriminative models that use richer features (prefix, suffix, capitalization)

4. Neural (BiLSTM-CRF / Transformer): BERT-based POS taggers achieve >98% accuracy on standard benchmarks

Chunking (Shallow Parsing)

Chunking groups tokens into non-overlapping phrases (chunks) — a step beyond POS tagging, short of full parsing:

[NP The quick brown fox] [VP jumped over] [NP the lazy dog]

Common chunk types: NP (noun phrase), VP (verb phrase), PP (prepositional phrase).

Chunking uses IOB (Inside-Outside-Beginning) tags:

• B-NP: Beginning of noun phrase
• I-NP: Inside noun phrase
• O: Outside any chunk

Lexical Analysis and Lexical Resources

WordNet

WordNet is a large lexical database grouping English words into sets of cognitive synonyms called synsets, linked by semantic relations:

Relation	Direction	Example
Synonyms (same synset)	—	"car" ↔ "automobile"
Hypernym (is-a, more general)	Up	"car" → "vehicle" → "artifact"
Hyponym (is-a, more specific)	Down	"vehicle" → "car", "truck", "bus"
Holonym (part-of)	Up	"wheel" → "car"
Meronym (has-part)	Down	"car" → "wheel", "engine"
Antonym	Opposite	"good" ↔ "bad"

WordNet is the basis for Wu-Palmer Similarity, Path Similarity, and Lin Similarity between word meanings.

Named Entity Classes (Preview)

Lexical analysis also identifies named entities — real-world objects with proper names:

• PERSON: Barack Obama, Marie Curie
• ORG: Google, United Nations
• GPE (Geo-Political Entity): France, New York
• DATE: January 1, 2024
• MONEY: $500 million

Full NER coverage is in Chapter 9.

Sense Inventory and Polysemy

Polysemy: a single word with multiple related meanings:

• "bank" → financial institution / riverbank / to bank a turn

Homonymy: different words that happen to share the same form:

• "bat" → flying mammal / cricket bat

Word Sense Disambiguation (WSD) selects the correct sense from context — covered in Chapter 5.

import nltk
import spacy
from nltk.corpus import wordnet as wn
from nltk import pos_tag, word_tokenize, RegexpParser

nltk.download(['averaged_perceptron_tagger', 'wordnet', 'punkt', 'maxent_ne_chunker', 'words'], quiet=True)

# ─── 1. POS Tagging with NLTK ────────────────────────────────
sentence = "The quick brown fox jumps over the lazy dog"
tokens = word_tokenize(sentence)
pos_tags = pos_tag(tokens)
print("NLTK POS Tags:")
for word, tag in pos_tags:
    print(f"  {word:<12} {tag}")

# ─── 2. POS Tagging with spaCy (preferred) ───────────────────
nlp = spacy.load("en_core_web_sm")
doc = nlp("The quick brown fox jumps over the lazy dog near the river bank.")
print("\nspaCy POS Tags:")
print(f"{'Token':<15} {'POS':<8} {'Fine POS':<10} {'Dep':<10}")
print("-" * 45)
for token in doc:
    print(f"{token.text:<15} {token.pos_:<8} {token.tag_:<10} {token.dep_}")

# ─── 3. Chunking with NLTK RegexpParser ─────────────────────
# Define grammar rules for noun phrases
grammar = r"""
  NP: {<DT>?<JJ>*<NN>}   # Determiner + adjectives + noun
  VP: {<VB.*><NP|PP>*}   # Verb followed by NP or PP
  PP: {<IN><NP>}          # Preposition + NP
"""
parser = RegexpParser(grammar)
tree = parser.parse(pos_tags)
print("\nChunk Tree:")
for subtree in tree.subtrees():
    if subtree.label() in ['NP', 'VP', 'PP']:
        print(f"  [{subtree.label()}] {' '.join(word for word, tag in subtree.leaves())}")

# ─── 4. WordNet Exploration ──────────────────────────────────
print("\nWordNet for 'car':")
for ss in wn.synsets('car'):
    print(f"  {ss.name()}: {ss.definition()}")

car = wn.synset('car.n.01')
print(f"\nHypernyms (car is a...): {[h.name() for h in car.hypernyms()]}")
print(f"Hyponyms (types of car): {[h.name() for h in car.hyponyms()[:5]]}")
print(f"Meronyms (car has...): {[m.name() for m in car.part_meronyms()]}")

# Semantic similarity
dog = wn.synset('dog.n.01')
cat = wn.synset('cat.n.01')
wolf = wn.synset('wolf.n.01')
print(f"\nPath similarity:")
print(f"  dog ↔ cat:  {dog.path_similarity(cat):.3f}")
print(f"  dog ↔ wolf: {dog.path_similarity(wolf):.3f}")

# Wu-Palmer similarity
print(f"Wu-Palmer similarity:")
print(f"  dog ↔ cat:  {dog.wup_similarity(cat):.3f}")
print(f"  dog ↔ wolf: {dog.wup_similarity(wolf):.3f}")

# ─── 5. Morphological Analysis ───────────────────────────────
from nltk.stem import WordNetLemmatizer
lemmatizer = WordNetLemmatizer()

words = [("studies", "v"), ("better", "a"), ("geese", "n"), ("running", "v")]
print("\nMorphological analysis:")
for word, pos in words:
    lemma = lemmatizer.lemmatize(word, pos=pos)
    synsets = wn.synsets(lemma)
    print(f"  {word} → lemma: {lemma}, synsets: {len(synsets)}")

Summary

• Morphemes are the smallest meaningful units; free morphemes stand alone, bound morphemes attach to roots
• Inflectional morphology: grammatical variation (tense, number) — English has 8 inflectional suffixes
• Derivational morphology: creates new words with new meanings (happy → happiness)
• POS tagging: assigns grammatical categories; neural models (BERT) achieve >98% accuracy
• Chunking: groups POS-tagged tokens into shallow phrase structures (NP, VP, PP)
• WordNet: lexical database with synsets linked by hypernymy, hyponymy, meronymy, antonymy

Next: Syntactic Analysis & Parsing — understanding sentence structure and grammar.