Knowledge Acquisition in Expert Systems

Tyler Young

Overview

• Science Fiction and Reality
• What is Learning?
• Introduction to Expert Systems
• Expert Systems Today: WordNet and Its Descendents
• Inferring Taxonomies from Text
• Discovering Other Relations
• Applying Knowledge to the Right Concepts (Disambiguation)
• Coping with Ambiguity

Science Fiction and Reality

Science Fiction and Reality

This isn't likely any time soon.

• Natural language is ambiguous.

• It's difficult to properly parse the relationships between elements in natural language.

• Most existing systems glean only a small of the available information from their "reading."

What is Learning?

What is Learning?

For a computer?

• Recall versus higher levels of knowledge
• No ability to identify referents

Not...

• Word associations
• Mental imagery
• Affective responses
• Physical metaphors in terms of which concepts are understood

Expert Systems

Then:

• Utilize stored knowledge base to solve problems at the level of a human "expert"
• Domain-specific
• Require a great deal of work on the part of both a human expert and a programmer

Now, and in the future:

• Ideally, build the knowledge base themselves, making them both more powerful and more cost effective.
• Use WordNet as a starting point
• Reason based on taxonomies

WordNet

Advantages to a Taxonomy-Based Knowledge Base

• Allows for inferences from the general to the specific

• Saves a great deal of time re-learning similar facts

• Allows for analysis in terms of a concept's parts and conceptual neighbors

Inferring Taxonomies from Text

• Search for domain independent patterns
• • E.g., American universities, such as Harvard, Yale, and Brown, . . .
  • NP₀, such as {NP₁, . . ., NP_n−1 (and/or) NP_n} . . .,
  • Obviously, quite limited in application.
• Automated search based on seed concepts
• • E.g., seed concepts interest rate, stock market, inflation, and economic growth
  • Computer searches corpus for these terms
  • E.g., discovering the term mortgage interest rate after reading “When the US economy enters a boom, mortgage interest rates rise . . ..”

Making the Best of Pattern-Based Search

• Goal: process 1 TB of data in a reasonable time (~4 months)
• Discover new patterns using a minimum edit distance algorithm
• • E.g., known relationship is a metal indicated by pattern “NP₀ is a metal.”
  • Find an edit distance of 1 for sentence “Platinum is a precious metal.”
  • Thus, judge with high probability that platinum has the property is a metal
• 30x to 1000x faster than previous methods

Discovering Other Relations

• New, popular approach uses NELL, the Never Ending Language Learner
• Uses WordNet as a starting point
• Many subsystems analyze a corpus
• • Some "read" (extract knowledge)
  • Some "learn" (find new patterns for the next pass)
  • • E.g., discovering new 2-argument relations

    • RiverFlowsThroughCity( someRiver, someCity )

    • Discovers pattern “r, in the heart of c, . . .,” where r is a known river and c a known city

Discovering Other Relations

• New, popular approach uses NELL, the Never Ending Language Learner
• Uses WordNet as a starting point
• Many subsystems analyze a corpus
• After finishing this analysis, another pass analyzing the same corpus begins
• Biggest benefit: Can extend new relationships to categories already stored in the knowledge base

Applying Knowledge to the Right Concepts

NELL's method for ambiguity resolution

• E.g., given the known relation

  ceoOfCompany(steve_jobs, apple)

  there are multiple potential conceptual referents for the word "apple"
• • Already knows relations like

    companyProducingComputers(apple)

    which refer to Apple, Inc.
  • Knows of no such company relations for the fruit
  • Thus, chooses Apple, Inc. as the referent

Applying Knowledge to the Right Concepts

NELL's method for ambiguity resolution

• Advantage over purely statistical methods: can cope with concepts that are used together infrequently, but still related conceptually
  • E.g., "Apple trees are free of mercury and arsenic."
  • Statistical analysis indicates Apple, Inc. is mentioned frequently with "mercury-free," "arsenic-free"
    • Maybe the "tree" is some new product by Apple, Inc.?
• Similar to the way humans resolve ambiguity
• • Word frequency influences speed of processing
  • Content knowledge can act as a final arbiter between competing interpretations

Applying Knowledge to the Right Concepts, cont'd.

• Similar method can be used to resolve synonyms in the knowledge base
• Given two orthographically similar terms that share a relation
• • E.g., relation

    ceoOfCompany(steve_jobs, company)

    shared by

    apple_computer_corp

    and

    apple

• . . . the system can infer that the terms refer to the same concept.

Requirements for Disambiguation

Whosoever hath, to him shall be given . . .

• Use previous knowledge both to disambiguate text and to rate the reliability of new information
• Apart from a large, relatively trustworthy knowledge base, the system has little means of evaluating the quality of its work.
• Becomes more important as more sources of knowledge are used

Coping with Ambiguity

Fuzzy logic

• Fuzzy sets: classify items without using clearly defined boundaries
• Contrast with "crisp sets," where an element is unambiguously either in or not in the set
• E.g., "My father is tall."
• • How tall is he?
  • In comparison to what?
• Humans automatically ignore these ambiguities.
• Using fuzzy rules to encode the rules, a machine can, too.
• • In tests, show greater accuracy than statistical methods

Using Fuzzy Logic, cont'd.

• Most useful in linguistic granulation tasks
• • Process of choosing some finite number of categories to classify objects
  • E.g., classifying ages: young, middle-aged, and old
• Meaning of each term determined using fuzzy logic

Looking to the Future

Present limitations

• Only extract relational information from text written to be informative
• • Obviously, this is a very small subset of all the text available!
  • Simplifies processing: can ignore many pragmatic considerations
• Only extract a small amount of information from analyzed text
• • May rely on quick-and-dirty pattern matches, like "NP₀, such as NP₁. . ."
  • Even in more sophisticated systems (e.g., NELL), only a small fraction of sentences are parsed fully

Looking to the Future

Future goals

• Understand greater breadth and depth in relations
• Understand greater variety of sentence structures
• "Read" a given corpus faster