The Hows and Whys of Semantic Distance

1
The Hows and Whys of Semantic Distance

• Vipul Kashyap, National Library of Medicine, NIH
• kashyap_at_nlm.nih.gov
• NIST Semantic Distance Workshop, Gaithersburg, MD
• November 11, 2003

2
Outline

• Why Semantic Distance ?
• Some applications of Semantic Distance
• The Semantic Conveyance problem
• Translations across multiple ontologies
• Role of Semantic Distance
• One approach of measuring Semantic Distance
• The Hows of Semantic Distance
• Types of Semantic Distance measures

3
Why do we need Semantic Distance (in Healthcare)?

• Healthcare Information is characterized by
  multiple terminologies,
• E.g., MeSH, CPT, LOINC, SnoMed, etc.
• Interoperability across terminologies is crucial
  to healthcare information system interoperability
• Which terminology do I interoperate with?
• What criteria/measure do we use?
• Application dependent v/s application specific
• Should the measure be machine understandable?
• Should the measure be human understandable?

4
Applications of Semantic Distance

• Information Retrieval
• To improve automated assignment of indexing based
  descriptors
• Semantic Vocabulary Integration
• To choose the closest related concepts while
  translating in and out of the multiple
  vocabularies
• Look at the Semantic Vocabulary Interoperability
  Project at
• http//cgsb2.nlm.nih.gov/kashyap/projects/SVIP

5
Semantic Conveyance
XXX Book
Sender
Receiver

• The Sender has his own ontology (The Red
  Ontology)
• The Receiver has his own (The Blue Ontology)
• For communication to take place,
• The receiver should translate the message
  (content) from the Red Ontology to the Blue
  Ontology
• Questions
• Is it always possible?
• How many candidate possibilities are there?
• How do we choose from them, Semantic Distance?

6
Terminology 1 The Red Terminology
Instructions
Reference-Manual
http//www.cogsci.princeton.edu/wn/w3wn.html
7
Terminology 2 The Blue Terminology
Conference
Agent
Person
Organization
Author
Publisher
University
Thesis
Periodical-Publication
http//www-ksl.stanford.edu/knowledge-sharing/onto
logies/html/bibliographic-data/
8
Inter-terminological relationships
Typically represented in the UMLS Metathesaurus

• Synonyms
• semantics preserving
• Hyponyms/Hypernyms
• semantics altering
• typically results in loss of information
• List of Hyponyms
• technical-manual hyponym manual
• book hyponym book
• proceedings hyponym book
• thesis hyponym book
• misc-publication hyponym book
• technical-reports hyponym book
• press hyponym periodical-publicatio
  n
• periodical hyponym periodical-publicatio
  n

9
Translations across multiple
terminologies
union(Book, Proceedings, ..., Misc-Publication),
document
Technical-Manual
GuideBook
10
Semantic Conveyance
XXX Book
Sender
Receiver
XXX Document XXX union(.)

• XXX Document
• XXX union(.)
• XXX .
• How do we chose between the various alternatives
• Semantic Distance to the rescue!!

11
Proposal for Semantic Distance Extensional
Measure
Loss in Precision
Loss in Recall
Ext(Term)
Ext(Translation)
Precision Ext(Term) ? Ext(Translation)
Ext(Translation)
Recall Ext(Term) ? Ext(Translation)
Ext(Term)
Percentage Loss Ext(Term) ?
Ext(Translation)
Ext(Term) Ext(Translation)
12
Choosing an optimal translationLocal v/s Global
Decision Making
Publication
Document
LOSS(Document, Book)
Document
Document
Journal
Publication
Book
Journal
Book
Journal
LOSS(Publication, Journal)
LOSS(Journal, Book)
LOSS(Document, Publication)

• Local Decision Making
• LOSS(Publication, Journal) LOSS(Document,
  Publication)
• Document is chosen as the translation
• But LOSS(Book, Document) LOSS(Book, Journal) !!
• Global Decision Making
• Both translations Document, Journal are passed
  on to the next level
• Journal is chosen as the appropriate translation

13
Using Subsumption for tighter bounds on Semantic
Distance

• Term subsumes Translation
• Ext(Translation) ? Ext(Term) ? Ext(Term) ?
  Ext(Translation) Ext(Translation)
• Precision 1,
• Recall Ext(Translation)
• Ext(Term)
• Should be able incorporate other
  application-specific measures to adapt distance
  measures
• Same terminological translation might be have
  different semantic distances based on application
  specific adaptations

14
Proposal for Semantic DistanceIntensional
Measure

• Difference in Translation
• Book ? union(Book, Thesis, Proceedings,
  Technical-Manual, Misc-Publication)
• Terminological Difference
• Book ? (AND Publication (ATLEAST 1 ISBN))
• Publication ? (AND document (ATLEAST 1
  PLACE-OF-PUBLICATION))
• Book ? (AND document (ATLEAST 1 ISBN) (ATLEAST 1
  PLACE-OF-PUBLICATION))
• Loss of Information
• (-) union(Trade-Book, Brochure, SongBook,
  PrayerBook, TextBook)
• information related to trade books, brochures,
  song books, prayer books and text books is lost
• () (AND (ATLEAST 1 ISBN) (ATLEAST 1
  PLACE-OF-PUBLICATION))
• spurious documents that dont have an ISBN number
  and a place of publication are gained

15
Measures for Semantic Distance Pros and Cons

• Intensional Measure
• May not make sense as it mixes two vocabularies,
• e.g., does Book - Book make any sense ?
• The problem becomes worse if the two
  terminologies are in different languages
• Makes it hard for the system to differentiate
  between the various alternatives
• Extensional Measure
• Based on Standard Information Retrieval Measures
  (F-measure)
• Can be tailored to reflect change in semantic
  distance for different applications
• However
• Probability distributions of various terms need
  to be estimated
• An information loss interval doesnt make much
  sense to the user.

16
How do we measure Semantic Distance?

• Extensional Approaches
• Intensional Approaches (based on concept
  definitions)
• Combination of the above approaches
• Combining the above using some weightage schemes
• Applying semantics of subsumption to the semantic
  distance measures
• Applying other constraints specified in the
  semantic network

17
Types of Semantic Distance Metrics Intensional

• Numerical
• Based on features (e.g., Tverskys measure)
• Based on traversal of specific conceptual
  relationships (is-a, part-of) and arbitrary
  domain specific relationships
• Non-numerical
• Based on semantic concept differences, e.g. a
  book without a publication date
• Important for human understandability

18
Proposal for Semantic DistanceTverskys measure
from Psycho-semantics

• S(a, b) A n B
• A n B a(a, b) A B (1 - a(a, b))B
  A
• S(a, b) is the similarity between two arbitrary
  objects, a,b
• A and B are feature sets of a, b respectively
• a is a real no. ? ? ? a ? 1

19
Types of Semantic Distance Metrics Extensional

• Numerical Based on estimation of underlying
  concept intensions
• Computation of joint and conditional probability
  distributions
• Computation of concept co-occurrences in
  documents
• Computation of cosine measures in a vector space
  mode

20
A Classification of Numerical (Intensional
Extensional Approaches)

• Traversal of graph-based information models
• Traversal of Hierarchical Relationships
• Intensional
• Feature contrast based approaches (e.g., Tversky)
• Intensional
• Probabilistic approaches (e.g., Precision,
  Recall, F-measure)
• Based on estimation of extensions/distributions
  of concepts
• Some combination of the above?

21
Counter-Example Hierarchical approaches
C4
C2
C3
C1

• Hiearchical approach
• semantic-distance(C1, C2) C4)
• However, if we look at concept extensions, it
  might be the case that
• only 10 of C2 is C1 and
• 90 of C3 is C1 and 20 of C4 is C2
• This implies, 18 of C4 is C1
• Thus, semantic-distance(C1, C2)
  semantic-distance(C1, C4)

22
Counter Example Probabilistic approaches
C2
C3
C1

• Probabilistic approach
• semantic-distance(C1, C2) and semantic-distance(C2
  , C3) are given
• Can we compute semantic-distance(C1, C3) ?
• It is quite possible that
• semantic-distance(C1, C3) C2) and
• semantic-distance(C1, C3) C3)
• Feature contrast based approach
• If the features of the concepts are known,
  Tverskys measures may be used
• However, the Semantic Proximity algorithm has to
  identify C3 as a concept that may have semantic
  resemblance with C1

23
Research areas dealing with Semantic Distances

• Knowledge Representation
• primarily non-numerical, intensional
• Statistical Clustering
• primarily numerical, vector space model
• Data Mining
• primarily numerical, probabilistic
• Machine Learning
• primarily numerical
• Information Retrieval
• primarily numerical, probabilistic
• Medical Informatics
• primarily numerical, intensional
• Natural Language Processing
• ??
• . any other field

24
Conclusions

• Semantic Distance measures need to be application
  specific
• Text Retrieval
• (Structured) Data Retrieval
• Domain and Context Specific
• Semantic Distance measures should be both human
  and machine processable
• They should be based on standard measures as far
  as possible
• E.g., F-measure from Information Retrieval
• There is a need for estimation of various
  distributions of medical concepts in a given
  population
• E.g. May need to mine CDC databases