Thesaurus Creation/ Term Clustering

1
Thesaurus Creation/Term Clustering

• Two major applications
• Query expansion
• fleshing out sparse queries with related words
• improves recall
• (at possible expense of reduced precision)
• Termset dimensionality reduction
• Similar outcome with smaller model

2
Term Dimensionality Reduction
Query Water Spaniel Diseases
Spaniel Spaniels disease diseases collie illness
ant
Original vector
VQ
Reduced vector
VQ
DOG ILL
INSECT
0
1
1

• Collie illnesses
• Poodle sickness

Problem Reduced flexibility in partial
weighting of synonym set ?
Synonyms got as much weight as the original

• Equivalent to query expansion
• when ?i for all synonyms is 1

3
Query Expansion
Query Water Spaniel Diseases
Original
syn
Expanded
1 1 1 ?1 ?2 ?3
?4 0
Water Spaniel diseases Spaniels diseases
illness collie ant
stem
? ? ?i ? semantic dist(wi ,t)
Relate Document set D1 Water Spaniels D2
Water Spaniel illnesses D3 Collie diseases
stem
syn
syn
4
Query Expansion
Query Water Spaniel diseases document1
water spaniels ..
5
Simplest Term Clustering
? Stemming is a clustering method
stemming
Original Term Set Clustered
Term Set
computing flies computers houses computation flown
compute flew house
comput fly
6
Another simple clustering method
Pre-existing thesauri(e.g. Rogets)
different pos

• illness ? disease, sickness, unwell, sick, ill,
• PhD ? Ph.D, PhD, Phd, Ph.D., .
• Term equivalence classes
• Loosely related topic sets
• DOG ? Spaniel, Collie, Schunauzer, bulldog,
  Poodle, .

same part of speech(pos)
7
Term Clustering
Non-hierarchical methods single pass(Salton,
71)

• Given clustering threshold/target size and
  similarity function sim(i , j )
• ? Pick random document Dj
• Assign a document di sim(Dj , dj) lt ? to
  cluster Cj and recalc centroid
• else create a new cluster Ck with centroid dk
• ? Exclude di from document list
• ? Repeat until document list empty

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
?
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
D2
.
.
?
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
D1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
?
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
D3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8
Types of Clustering Behavior/Criterion

• Sim(ti , tj)
• Document level co-occurrence in same document
• Verb-object
• Syntagmatic similarity sim(drink, wine)
• appears together sim(eat, meat)
• in region sim(drink, water)
• Paradigmatic similarity sim(wine, water)
• appears as objects sim(wine, drink)
• of the same verb sim(wine, meat)

based on object of drink or of all verbs
?
?
9
N-gram Syntagmatic similarity sim(Hong,
Kong) occur together sim(soap,
opera) sim(soap, suds) Paradigmatic
similarity sim(opera, suds) occur in same
sim(tall, short) context sim(long,
short) sim(Hong, Kong)
?
soap opera suds residue
Ivory soap Dial Lye