Automating Ontology Development

Next

Source:  http://www.jfsowa.com/pubs/autotalk.htm

Abstract:  Large ontologies are required for two different, but related purposes: understanding unrestricted natural language text; and merging and aligning independently developed knowledge bases and databases. The largest ontologies currently available for these purposes, Cyc, WordNet, and EDR, have been developed at enormous expense by organizing and encoding the most critical information by hand. The high cost and slow pace of the development indicate that handcoding techniques are obsolete, inflexible, and inappropriate for large-scale ontology development. Although fully-automated ontology development is not yet feasible, many automated and semiauomated techniques have been implemented for performing some of the required subtasks. This talk surveys some of those techniques and proposes a framework for integrating them into a system of tools that could support more efficient and flexible methods of ontology development and customization.

Questions to Consider

• How can we build large ontologies?

• How can we accommodate existing ontologies, databases, knowledge bases, and web sites?

• How do we relate them to natural languages?

• How can we use automated or semiautomated tools to build them, modify them, extend them, or apply them?

Large Hand-Coded Ontologies

Three large ontologies:

• Cyc: 100,000 concept types with over a million axioms.

• Electronic Dictionary Research (EDR): 400,000 concept types with mappings to English and Japanese words.

• WordNet: 166,000 English word senses with related projects for other languages.

Building these things requires a great deal of time and money.

Small Hand-Coded Ontologies

Purpose of This Talk

• Propose a more flexible structure.

• Show how it can coexist with various kinds of systems, large and small.

• Show how it can support natural language processing.

• Show how it can support detailed deductions for knowledge-based systems.

• Show how automated and semiautomated tools can help the above tasks.

2N+2 Hierarchies

Two hierarchies for each natural language:

• Words.  A hierarchy of words and word senses, similar to WordNet and EDR.

• Canonical graphs.  A partial ordering of conceptual graphs that express the lexical patterns associated with each word of each language.

Two language-independent hierarchies:

• Types.  A lattice of all the concept and relation types that are used in the theories and in the canonical graphs.

• Theories.  An open-ended, potentially infinite, lattice of all possible theories, each corresponding to a possible language game.

Some Automated Techniques

A small sample of many techniques — similar, related, or radically different:

• Word hierarchies.  Example:  MindNet by Richardson et al. (1998), Dolan et al. (2000).

• Canonical graph hierarchies.  Example:  Ariosto by Basili et al. (1993, 1994, 1996, 1999).

• Type hierarchies.  Example:  Formal Concept Analysis by Ganter & Wille (1999).

• Lattices of theories.  Example:  BIAIT by Donald Burnstine (1979).

MindNet

Automated development of semantics for the MS-NLP project

• Processes dictionaries (LDOCE) and corpora (Encarta Encyclopedia).

• Derives word hierarchies automatically (7 hours on a P266).

• Finds relationships by analyzing corpora (34 hours for 500,000 sentences on a P266).

Ariosto-Lex and Trevi

Ongoing project to develop integrated tools for NLP engineering

• Includes modules for lexical aquisition, linguistic processing, and applications.

• Lexical acquistion starts with a generic word hierarchy with generic canonical graphs and uses automated means to specialize them to particular domains.

Formal Concept Analysis

Techniques for analyzing concepts and creating lattices

• Analyze a given set of concepts to determine relevant attributes.

• Present concepts and attributes in a table.

• Construct a lattice of concept types automaticaly.

• Basis for tools that can be used in semiautomated and collaborative development.

Table of Attributes and Categories

A Lattice of Beverages

Revised Lattice of Beverages

Disagreements Lead to Distinctions

• Socrates' principle:

  Whenever two philosophers — human or machine — disagree, draw a distinction.

• Anyone may discover a conflict.

• Anyone may suggest a distinction.

• Machine recomputes the lattice.

• Repeat until everybody is happy.

Collaborative Development

Semiautomated development:

Collaboration between one machine and one human.

Collaborative development:

• Collaboration among multiple humans and multiple machines distributed across the Internet.

• Anyone may draw attention to an error, conflict, or inconsistency.

• Owner of the data has the right to accept or reject a suggested change.

BIAIT

Business Information Analysis and Integration Technique — based on seven binary distinctions:

1. Bill.  Does the supplier bill the customer, or does the customer pay by cash?

2. Future.  Does the supplier deliver the product at some time in the future, or does the customer take the order from stock?

3. Profile.  Does the supplier keep a profile of the customer, or is every transaction a surprise?

4. Negotiate.  Is the price negotiated or fixed?

5. Rent.  Is the product rented or purchased?

6. Track.  Does the supplier keep track of the product after it is sold or not?

7. Make to order.  Is the product made to order, or prefabricated?

Building Theories with BIAIT

Combinatorial construction of theories from conjunctions of axioms:

• Each BIAIT question is based on a binary distinction.

• Each choice contributes one axiom.

• Result of seven choices is a theory with seven axioms.

• Total number of theories is 2⁷ = 128, which describe 128 kinds of businesses classified according to their record-keeping needs.

The Lattice of Theories

An infinite lattice of all possible theories, also called a Lindenbaum lattice:

• Only a finite number of theories are stored at any given time.

• But there are enough slots in the lattice to accommodate any theory that anyone could ever conceive.

• Ordered by generalization and specialization.

• Belief-revision operators are used to relate theories: contraction, expansion, and analogy.

Navigating the Lattice of Theories

Special-Purpose Theories

Summary

• A single, consistent, all-encompassing theory is impossible.

• The enormous number of special-purpose solutions aren't going to disappear overnight, if ever.

• The infinite hierarchy of theories can accommodate everything, general or special-purpose, scruffy or neat.

• Automated and semiautomated techniques are essential for developing, extending, and using the four kinds of hierarchies.

References

For the slides used in this talk, see

http://www.jfsowa.com/pubs/autotalk.htm

For further discussion of the hierarchies, see

http://www.jfsowa.com/pubs/signtalk.htm

For even more detail, see the [unfinished] paper:

http://www.jfsowa.com/pubs/signproc.htm

All other references are [or will be] in the bibliography:

http://www.jfsowa.com/bib.htm