Technologies for an Intelligent Web

1
Technologies for an Intelligent Web

• Francis Heylighen
• Center Leo Apostel
• Vrije Universiteit Brussel

2
What is intelligence?

• capacity for problem-solving in the widest sense
• problem difference between perceived and
  preferred
• input perception, output plan for action
• problem-solving
• efficiently exploring mental map
• includes interpretation, search, inference,
  decision-making, etc.
• selecting the adequate combination of resources
  to go from present state to desired state
• requires mental map or knowledge
• representation of problem states and resources

3
Collective intelligence

• synergy
• when the group can find more/better solutions
  than the sum of solutions found by all members
  individually
• requires Collective Mental Map
• integrated sum of all individual knowledge
• read/write access for all people
• no individual or computer can store a CMM for
  humanity
• externall, shared memory
• requires a distributed representation/search
• must self-organize no centralized control
  possible
• the web can be made to function as a CMM

4
Global network - Global Brain?
5
The web as as a collective mental map

• distributed knowledge system
• sum total of individual contributions
• coherent because of its interlinking
• global neural network
• Web pages as neurons
• hyperlinks as synapses
• problem-solving support
• helps the user collect the resources that solve
  their problem
• e.g. find me
• a second-hand video recorder
• the quickest way to travel from here to there
• the treatment that tackles symptoms
• information about growing blueberries

6
Hypertext network
7
Network of Nodes and Links
8
Web as network of resources

• Nodes are any resources that can help solve
  problems
• web documents
• computer programs or databases
• software agents
• products fridges, TVs, phones, ...
• people
• organizations, public or commercial
• Links are relations between resources
• hyperlinks
• people having access to other people/devices/organ
  izations..
• relations between databases or programs

9
Links as relations

• links can have types
• e.g. is author of, cites, lives in, works
  for, is a type of
• links can have weights
• link weights measure degree of association
• effort needed for the one to access or connect
  to the other
• e.g. order in which telephone numbers are listed
  in cellular phone memory
• first ones are easier to access

10
Metasystem Transitions in the Brain

• one-to-one communication
• direct transmission
• traditional media phone, post, ...
• many-to-many communication
• integrating and processing different signals
• this is the level of the present web
• learning
• creating/adapting connections from experience
• thought
• exploring combinations never experienced together
  
• discovery
• developing new concepts, rules and models

11
Learning Webs

• let the web learn from the way it is used
• optimize connection between initial and desired
  states
• assumption users go from a web page to relevant
  page
• when link between two pages is used, weight is
  increased
• unused links are correspondingly weakened
• indirect links too are reinforced
• user goes A ? B, and B ? C, then also A ? C
  is reinforced
• creates shortcuts for often travelled paths
• turns the web into an associative network
• the more associated the nodes, the stronger their
  connection
• organization similar to the brain

12
The Learning Web Experiment

• performed by Johan Bollen and myself
• 150 most frequent English nouns
• each word gets one web page
• each page is linked randomly to 10 other
  pages/words
• users are asked to choose the best association
  out of 10

13
The Learning Web Experiment
14
Results from the experiment
15
Associative Network from Experiment
16
Hebbian Rule for Web Learning

• Connection is strengthened proportional to joint
  activation
• Activation degree of usefulness for user
• explicit evaluation by user
• implicit evaluation derived from
• duration of visit
• bookmarking, saving, printing, ordering, etc.
• Joint activation usage by same user
• product of activation degrees
• activation can be negative -gt link weakened
• if user dislikes resource
• activation decays exponentially
• ? reinforcement decays with interval between
  usages

17
Spreading activation

• Associative networks can be explored in parallel
• users can only move sequentially between nodes
• input nodes can be activated simultaneously
• activation follows associative links to other
  nodes
• these are in turn activated, proportionally to
  link strength
• thus, activation spreads over a semantic
  neighborhood
• primitive form of thinking
• exploring different combinations of concepts

18
Spreading activation illustration
bird
seagull
bank
financial institution
sit
19
Spreading activation illustration
bird
seagull
bank
financial institution
sit
20
Spreading activation illustration
bird
seagull
bank
financial institution
sit
21
Spreading activation illustration
bird
seagull
bank
financial institution
sit
22
Spreading activation illustration
bird
seagull
bank
financial institution
sit
23
Personalized Recommendations

• agent collects appreciated items
• e.g. liked pages, music records, concepts
• by spreading activation from these elements, the
  agent tries to find associated items, e.g.
• related pages, similar records
• pages related to all concepts
• e.g. paper, work, room -gt office
• the agent recommends the most activated items
• these are most likely to please the user
• similar to collaborative filtering
• recommend items appreciated by people with
  similar tastes

24
Finding attractors

• If spreading is repeated many times, activation
  concentrates in attractors of the network
• densely connected clusters of nodes
• equivalent to calculating eigenvectors of linking
  matrix
• Application finding communities
• related pages on a subject
• e.g. Kleinberg, CLEVER project
• Application determining authority
• Googles PageRank algorithm
• most attractive pages are most authoritative

25
Spreading Authority
26
Ill-Structured Problems

• User in general cannot formulate
  problem/goal/preferences
• only vague associations
• e.g. diarrhoea, constipation, cramps, colon, gas,
  bloating...
• implicit problem How to cure Irritable Bowel
  Syndrome?
• activate symptom resources
• let activation spread
• find most authoritative documents that solve
  problem
• The web thinks ahead of the user
• takes into account implicit signs of interest
• suggests solutions to problems the user may not
  even be aware of

27
The Semantic Web

• Spreading activation diffuses or ends up in
  attractors
• loss of information with respect to initial state
• Constrained spreading activation inference
• follow only specific link or node types
• allows activation to spread in a much more
  focused way
• Answering structured queries
• E.g. lady works for client, lives in Washington,
  has son that goes to Princeton
• link types employed by, adress, child of,
  studies at, ...
• E.g. appointment with nearest plumber within free
  hours
• Requires consensual ontologies
• explicit taxonomies of types and their relations

28
Collective Development of Ontologies

• Ontological categories must be formal,
  unambiguous
• very hard to develop manually
• Clustering
• put similar items into same category
• from soft associations to hard categories
• Bootstrapping
• concepts defined by relations with other concepts
• represented as column vectors of association
  matrix
• concepts more similar if associations overlap
  more
• similarity s can be calculated as dot product of
  vectors

29
Knowledge Discovery

• Web can autonomously create new knowledge
• clustering ? new categories or concepts
• rule if (concept), then (other concept)
• e.g. if banana, then yellow if fire and gas,
  then explosion
• system of concepts and rules ? knowledge
• Ex. medical syndrome
• huge database of persons, symptoms, treatments,
  etc.
• clustering on the basis of symptoms ?
  distinguishing syndromes
• correlating syndromes, treatments and outcomes ?
  finding best treatment for given syndrome

30
Conclusion

• web can be seen as network of nodes and links
• nodes resources
• new links can be learned implicitly from usage
• makes the web more efficient, intuitive, dense,
  ...
• network can be explored through spreading
  activation
• allows vague, intuitive, unstructured queries
• ontologies can be used to structure web
• allows concrete, explicit queries
• new structures can be mined from implicit
  relations
• allows creation of ontologies, knowledge discovery