Smeagol:

1
Smeagol A Goal Directed Learning Agent for the
Semantic Web

• Motivation
• The Semantic Web benefits from learning
• Users need personalised views of information.
• Ontologies may not provide relevant properties
  for your application.
• Machine Learning techniques can help identify
  this missing information!
• Machine Learning benefits from the Semantic Web
• Learning methods often require large amounts of
  domain specific background knowledge and rules.
• Previously this knowledge has been manually
  acquired from domain-experts and coded in the
  representation language of the applications.
• Now RDF provides a standard representation
  language and the Semantic Web provides a
  world-wide web of knowledge backed by ontologies!
• However
• In the Semantic Web context we have to confront
  issues such as scalability and data sparseness.

• Goal Directed Learning
• What types of Goals are there?
• Set-Difference
• e.g. distinguish between movies from the 80s and
  90s.
• Identify-Sets
• e.g. differentiate Star Wars Trilogy from LOTR
  trilogy.
• Knowledge Reformulation
• e.g. map instances from one ontology to another.
• How do learning goals arise?
• How do learning goals affect the learning
  process?
• How do learning-goals relate to each other?
• How are learning goals represented?

• FOAF Experiments
• We have previously conducted learning experiments
  using FOAF data.
• 6.5 Million Triples
• ILP and clustering used to identify unspecified
  
• conceptualisations, and learn descriptions of
  these.
• Extremely resource-hungry! Gigabytes of memory
• and days to perform experiments.
• member(A) -
• foaf_groupHomepage(A, http//www.aktors.org).
• member(A) -
• contact_nearestAirport(A, airports?ABZ).

FOAF Experiments
Goal Directed Learning
Motivation
We use an RDF representation, where a goal has
the following properties GoalType, Parameters,
Contexts, Datasources.
ltSmeagolGoalgt lttype rdfresourcesmeagolSetD
ifference /gt ltset1gt ltQuerygtltrdqlgtSELECT ?x
WHERE (?x, ltimdbdirectorgt,
lthttp//imdb.com/name/nm0040/gt)lt/rdqlgtlt/Querygt
lt/set1gt ltexclude-predicatesgt ltrdfSeqgt
ltrdfli rdfresourceimdbdirector /gt
lt/rdfSeqgt lt/exclude-predicatesgt ltdatasource
rdfresourcedataIMDbMovies /gt ltdatasource
rdfresourcedataIMDbPeople /gt ltcontext
rdfresourceimdbArty /gt

• Smeagol
• Implemented in Python
• ILP Srinavasans Aleph
• Clustering Hierarchical Agglomerative Clusterer
• Rule learning Cohens Slipper

Smeagol
Data-Acquisition
Learning

• Sub-Goals
• Pick a context
• Learn a missing predicate
• Not all datasources will include all predicates.
• Include additional resources
• Fetch ontologies, traverse graph further.
• Sub-divide learning space
• e.g. map year to decade.

ILP
Clustering
Rule-learner
Create Sub-goal
Success!
Result
Sci-Fi movies have a special effects producer
and are directed by Stanley Kubrick or George
Lucas.
Failure
Moviehack Research Tool Built to allow easy
experimentation with and evaluation of Smeagol.

• RDF Data
• Extracted from Internet Movie Database (IMDB).
• Homemade ontology using FOAF.
• Top 250 movies ? 30k people.
• 1.5 Million Triples.
• Trivia information extracted from text
  descriptions.
• ltimdbMovie rdfIDhttp//imdb/title/tt0068646/gt
  
• ltdctitlegtThe Godfatherlt/dctitlegt
• ltimdbyeargt1972lt/imdbyeargt
• ltimdbdirector rdfresource/nm00338 /gt
• ltimdbgenre rdfresourceimdbDrama /gt
• ltimdbcastgt
• ltimdbRolegt
• ltfoafnamegtDon Vito Corleonelt/foafnamegt
• ltimdbactorgt
• ltfoafPerson rdfresourcegt
• ltfoafnamegtMarlon Brando .

• User can choose predicates to include/exclude
  from learnt rules.

• RDQL is used to specify sets.
• Web-Interface written in PHP using RDF API for
  PHP (RAP)

• Results presented as Prolog rules.

Demo Available!
Gunnar Aastrand Grimnes, Alun Preece Pete
Edwards Computing Science Department, University
of Aberdeen, UK
Contact ggrimnes_at_csd.abdn.ac.uk