Ontology Learning

1
Ontology Learning

• For the Semantic Web

2
The Paper Itself

• Based around two products OntoEdit and
  Text-to-Onto.
• A rather foundational approach to the problems
  surrounding information extraction.
• Occasionally, some really weak sentence
  structure.
• Problems with inconsistent example use.
• A frustrating exercise in presenting questions
  and not the answers.

3
The Web(our semantic battleground)

• The Web was created as a free-form information
  space.
• Made for human comprehension, not machine
  understanding.
• From experience with the web there seems to be
  some inherent aversion to correct speeling or
  gRamar.

4
Machine Semantics

• Computers, while originally designed to
  understand a series of electrical pulses, have
  had that same vocabulary expanded to be able to
  also evaluate letters, booleans, and numbers.
• -Brian Goodrich

5
Ontologies

• Ontologies are metadata schemas.
• Controlled vocabulary of concepts
• Machine understandable semantics
• Define shared domain conceptualizations
• (e.g. website to website, people to machines,
  etc.)

6
The Assumption

• If every internet webpage had an associated
  perfect ontology that was just as accessible as
  the selfsame webpage, the creation of the
  semantic web would be only as far away as the
  creation of a browser that can find and interpret
  those ontologies and extract information based
  upon those models.
• -Brian Goodrich

7
The Knowledge Bottleneck

• manual acquisition of ontologies still remains
  a tedious cumbersome task resulting in a
  knowledge acquisition bottleneck.
• Steffen Staab

8
The Challenge

• In Overcoming this knowledge acquisition
  bottleneck the authors took a three-fold
  approach
• Time (Can you develop an ontology fast?)
• Difficulty (Is it difficult to build an
  ontology?)
• Confidence (How do you know that youve got the
  ontology right?)

9
OntoEdit

• OntoEdit supports the development and maintenance
  of ontologies using graphical means. It supports
  RDF-Schema, DAML-ONT, OIL and F-Logic.
• Has many of the same features as our Ontology
  Editor.
• Cardinality Restrictions
• Keyword Associations
• Value Phrase Restrictions

10
Assaulting the walls of Jericho

• Multi-disciplinary approach
• Machine learning (human assisted)
• Five Phase approach
• Import and Re-use existing ontologies
• Data Extraction uses machine learning to sculpt
  major sections of the target ontology.
• Target ontology is pruned
• Refinement(?) (automatically and incrementally
  maintained by evaluating quality of proposals)
  Hahn Schnattinger
• Validation using prime target application as a
  measure for success of the ontology.

11
Another Wonderful Graph

• Import/ReUse
• Extract
• Prune
• Refine
• Validate
• Legacy data reference to archaic databasing
  techniques
• Text-to-Onto

12
Components for Learning Ontologiesby Staab and
Maedche

• Management Component
• Resource Processing Component
• Algorithm Library
• GUI for Manual Engineering

13
Ontology Primitives

• a set of strings that describe lexical entries L
  for concepts and relations
• a set of concepts2 C
• a taxonomy of concepts with multiple inheritance
  (heterarchy) HC
• a set of non-taxonomic relations R described
  by their domain and range restrictions
• a heterarchy of relations, i.e. a set of
  taxonomic relations HR
• relations F and G that relate concepts and
  relations with their lexical entries,
  respectively
• a set of axioms A that describe additional
  constraints on the ontology and allow to make
  implicit facts explicit

14
Management Component

• OntoEngineer uses to select desired XML/HTML
  pages, document type definitions, databases, or
  pre-existing ontologies
• Selects methods for the Resource Processing
  Component and Algorithms for the Library
  Component
• Also includes a crawler that can find legacy data
  relevant to creation of the ontology on the web.
  (used for training data)

15
Resource Processing Component

• HTML documents may be indexed and reduced to free
  text.
• Semi-structured documents, like dictionaries,
  may be transformed into a predefined relational
  structure.
• Semi-structured and structured schema data (like
  DTDs, structured database schemata, and existing
  ontologies) are handled following different
  strategies that may (or may not) be discussed
  later.
• For processing free natural text our system
  accesses the natural language processing system
  SMES (Saarbrucken Message Extraction System), a
  shallow text processor for German. SMES comprises
  a tokenizer based on regular expressions, a
  lexical analysis component including various word
  lexicons, a morphological analysis module, a
  named entity recognizer, a part-of-speech tagger
  and a chunk parser.

16
Algorithm Library Component

• This is the actual ontology builder and where we
  revisit our previous model
• Import/ReUse
• Extraction
• Pruning
• Refining
• And then almost introduce some actual algorithms
  these phases use.

17
Import/Reuse

• Recovering Conceptualizations
• First, schema structures are identified and
  imported separately. This may be done manually
  or using reverse engineering tools.
• Second, merging and aligning.
• This is a HUGE body of research that is largely
  ignored by this document
• While the general research issue concerning
  merging and aligning is still an open problem,
  recent proposals (e.g., 8) have shown how to
  improve the manual process of merging/aligning.

18
Extraction

• Lexical Entry Concept Extraction
• Hierarchical Concept Clustering
• Dictionary Parsing
• Association Rules

19
Lexical Entry Concept Extraction

• Uses statistical technique (N-grams) similar to
  the product from Cuis presentation on the
  BioMedicine data extractor to group multi-word
  nouns together and associate them with their
  corresponding verbs
• Every time a new lexical entry is introduced to L
  the OntoEngineer must decide whether to include
  the entry in an existing concept domain or to
  introduce a new one.

20
Hierarchical Concept Clustering

• A useful way of creating a taxonomic
  classification of concepts.
• Done automatically Text-to-Onto clusters concepts
  by adjacency of terms and syntactical
  relationships.
• Done by a cooperative machine learning system,
  ASIUM, presented by Faure Nedellec. Uses the
  verb to noun and noun to verb association method.
  
• Thus, they cooperatively extend the lexicon, the
  set of concepts, and the concept heterarchy. (L,
  C, HC)

21
Dictionary Parsing

• This is really only one step further than what we
  are doing with the lexicons in our own Ontology
  Editor. The identified dictionary words are used
  with the concept clustered verb and noun
  associations to infer relationships between
  lexical entries.

22
Association Rules

• These algorithms are usually used for data
  mining.
• Works by using the taxonomy heterarchy to
  generalize the lexical entries and thereby draw
  conclusions about their use.
• Snacks are purchased together with drinks
  -instead of-
  Lays chips are purchased with Sprite.

23
Example Output from Text-to-Onto
24
Completeness vs. Scarcity (Pruning)

• Pruning the Ontology
• It is a widely held belief that targeting
  completeness for the domain model on the one hand
  appears to be practically unmanageable and
  computationally intractable, and targeting the
  scarcest model on the other hand is overly
  limiting with regard to expressiveness. Hence,
  what we strive for is the balance between these
  two, which is really working.
• Staab and Maedche

25

• Import and ReUse, as well as the different
  Extraction methods weve discussed all tend to
  introduce unfocused elements into the ontology,
  as more general rules satisfy the conditional
  statements much more often.
• Pruning is the art of diminishing the ontology to
  more specific rules.
• First, must evaluate how removal of item from C
  (the set of concepts) will affect the rest of the
  ontology. (Petersen 9, no dangling or broken
  links)
• Second, based on absolute or relative counts of
  frequency determine which ontology items are to
  be either kept or pruned. (Kietz 13)

26
Refine

• Hahn and Schnattinger
• Incremental approach to updating an ontology
  centered around linguistic and conceptual
  quality of various forms of evidence i.e.
  conflicting and analogous semantic structures
  underlying the generation and refinement of
  concept hypothesis.

27
Conclusions

• Ontology learning a significant leverage to
  Semantic Web.
• Propels propagation of ontologies
• Multi-disciplinary approach to the problem

28
Further Challenges in Learning

• XML namespace mechanisms will turn the web into
  an amoeba-like structure, with ontologies
  supporting and referring to each other (ReUse and
  Import) Not clear yet on what will be the
  semantic result of this evolution.
• This examination has been restricted almost
  entirely to RDF-Schema. Additional layers of RDF
  (future OIL or DAML-ONT) will require new means
  for improved Ontology engineering.

29
Questions?
Dabu?