Emergent Semantics

1
Emergent Semantics

• Trends Controversies
• Satya Sanket Sahoo
• LSDIS Lab,
• University Of Georgia

2
Outline

• Overview
• Case 1 Image Retrieval
• (Simone Santini, University of California, San
  Diego)
• Case 2 Multiagent system communication
• (Luc Steels, University of Brussels AI lab and
  Sony Computer Science Lab, Paris)
• Case 3 Emergent Semantics for ontologies
• (Alexander Maedche, University of Karlsruhe)
• Conclusions

3
Overview of Emergent Semantics

• We are drowning in information starving for
  knowledge! Web, Knowledge management etc.
• Wittgenstein I World is composed of complex
  facts and we analyze into less complex facts till
  we arrive at atomic facts
• Wittgenstein II The meaning of an expression
  must be understood in terms of its communication
  context i.e. a posteriori.

4
Overview

• How about combining WI WII synergistically?
• Problems with current semantic-based systems
• Knowledge acquisition bottleneck shortage of
  up-to-date semantic description
• Interactions between users and systems go unused
  can help semantic structure emerge
• 
  
• Emergent Semantics Let the semantics for a
  system emerge through the interactions of human
  machine agents!

5
Case 1 Image Retrieval

• Problem Retrieve images from a database based on
  their content.
• Content ? images meaning perhaps?
• Idea - extract features from image and cast in a
  metric space so that similar images have
  similar meaning.
• But, problem in Image database because of this
  very Semantic presupposition

6
Images meaning

• Images are terms predicated by an external
  discourse
• Images in isolation have no meaning, rely on an
  external context to predicate their content.
• Every decoding is another encoding! context
  makes the meaning clear.
• Ex UFO Artwork
• (Year 1660)

7
Semantics in Database

• Possibly, design a schema using some data types
  and data algebras that encapsulate the desired
  semantics.
• Database discourse The set of constructs that
  determine the semantics of records.
• So, the database schema determines the discourse
  a priori completely formalized.
• But, in Image databases, representing discourse
  is not possible.

8
Solution concentrate on these points

• Parts of an images semantic is derived from its
  relation with other images so, query should
  manipulate similarity function
• Semantics of images descriptor (features) should
  be specified through a discourse (logic,
  algebraic or functional aspects)
• Image always has meaning relative to practices
  and social codes of specific user.

9
Goals

• The interaction between user and database should
  not be so much as to retrieve image based on
  pre-existing semantics but, to create image
  semantics
• Interaction between user and database should be a
  navigation, not a query, so that user dictates
  similarities association between images
  through this re-organize the database to reflect
  the desired semantics.

10
Case 2 Language Games for Emergent Semantics

• To process information, the information must be
  represented there is no computation without
  representation.
• Can computer systems develop and adapt
  representations?
• Origins of representations can be classified into
  two approaches
• Induction
• Selection

11
Induction

• In statistical-pattern recognition, symbolic
  machine learning and neural-network research
• Can be supervised the system receives
  feedback on what it needs to learn
• Can be unsupervised the system attempts to
  detect the natural classes or regularities in the
  data.

12
Induction - limitations

• For supervised learning great deal of human
  intervention (assemble adequate set of training
  data, choose outline of representation to bias
  learning algorithm..)
• For unsupervised learning systems come up with
  concepts that are irrelevant to task at hand.
• Ex a series of images may cluster based on
• the of day they were taken rather than on
• objects contained in the image.

13
Selection

• Based on Darwinian model of genetic algorithm.
• One process generates at random a variety of
  possible representations and the other process
  (selection process) picks out the representation
  best suited for the given task.
• But, limitations w.r.t Human intervention and
  risk associated with search like ending up with
  local minima or long search time.

14
Construction

• Proposed by the author Luc Steels
• Ex Moving into a new house, so order everyones
  shoes we need a shoes to be categorized and an
  external representation of the categorization.
• Need three activities
• Interaction
• Construction
• Communication

15
Construction

• Interaction involves task that generates need
  for new meaning and we need a multiagent system
  that interacts with the real world to evolve a
  representation
• Construction these agents impose new categories
  (based on function of the task). It does not
  necessarily have to gradual development.

16
Construction

• Communication agents need to communicate among
  themselves develop symbols (on their own) to
  externalize their concepts

17
Goals

• Co-evolution of language (external
  representation) and meaning memetically?
• This is a collective and incremental process of
  development of conceptual framework.
• can be applied to formation and adaptation of
  ontologies for Web-based agents or evolving
  dialogs with humanoid robots.
• This approach to Emergent Semantics is
  complementary to Induction and Selection.

18
Case 3 Emergent Semantics for Ontologies

• The proposed semantic web will use formal
  languages like OWL to represent ontologies.
• But, Semantic Web should bring value to human
  users and not just enable machine understanding
  of content.
• Communication process between humans and machines
  cannot be ensured by these formal semantics alone.

19
Emergent Semantics for Ontologies

• We need to analyze, design and develop computing
  systems using and evaluating Semiotic theories.
• In general, there are three levels of
  understanding
• Syntactic Natural language primitives
• Semantic Meaning of the primitives
• Pragmatic Interpretation of primitives by humans.

20
Ontology development

• Ontology tries to formalize natural language to
  enable machine processable and understandable
  data.
• There are three aspects to be covered
• Formal semantics imposes machine processable
  semantics
• Human cognitive structures the language
  primitives in use
• How humans use these structures to communicate

21
Ontology learning

• Ontology learning should be a bottom-up
  approach.
• Given a set of data to reflect human
  communication and interactive process discover
  semantics implicitly contained in it
• Emergent Semantic require support from ontology
  learning approaches
• Extracting ontology from scratch
• Ontology evolution from analyzing legacy and
  application data.

22
Emergent Semantics - Conclusions

• Instead of imposing a priori concepts on
  observations (Top-Down).
• Let semantics emerge from simple observations
  derived from interactions (Bottom-up).
• This approach is proposed to provide precise
  query, retrieval, communication or translation
  for a wide variety of applications.

23
References

• Steffen Staab , Alexander Mädche , Frank Nack ,
  Simone Santini , Luc Steels Emergent Semantics.
• IEEE Intelligent Systems, Trends
  Controversies, 17(1), Jan/Feb 2002, pp. 78-86.,
  Publishing Year 2002
• W.I. Gorky, D.V. Sreenath, and F. Fotouhi
  Emergent Semantics and the Multimedia Semantic
  Web.
• Ramesh Jain, University of California, San Diego
  and PRAJA Inc. Emergent Semantics and
  Experiential Computing. jain_at_praja.com