CYC Cycorp

Dec 15, 2008 10:12 PM

“learning only occurs at the fringes of what you already

know,”   lk

Over the last twenty years, many human

“Cyclists” have been engaged in this painstaking

representation process, and have represented over two

million facts and rules about more than 200,000 entities

and types of entities.

CYC Inference Engine supports  DEDUCTIVE, ABDUCTIVE, and INDUCTIVE INFERENCE.

NL translation to go from  English to  CycL  (the Cyc formal language)

Ultimately, we found that

everything the reviewers wanted to express about a single

sentence could be captured by five characteristics:

· Comprehensibility: is a generated sentence easily

understood by a human knowledge worker? In

most cases, when a sentence is not clear, either the

natural-language generation is faulty or the

sentence makes use of a predicate that is intended

to express a non-natural structural or internal

concept.

· Appropriateness: is a sentence something that

should have been found to be invalid during the

type-checking phase of generation? For example, it

would be inappropriate, if mildly entertaining, to

suggest that “Osama bin Laden is a triangle” or

that “Al-Qaida is affiliated with inference

parameters.”

· Truth: is the information content of the sentence

sound? Many of the generated sentences, while

reasonable hypotheses, are simply false.

Interest value: is the sentence saying something

that is worth representing in the system for further

reasoning, as opposed to an irrelevance, such as

“George W. Bush is sitting down”?

· Plausibility: is the sentence something that seems

likely to be true, based on common-sense

evaluation rather than domain expertise? An

example of an implausible sentence might be, “Mel

Gibson is affiliated with Al-Qaida” – this might be

true, but seems improbable. Evaluating the

plausibility of a sentence is both subjective and

contextual; it remains to be seen, therefore,

whether evaluating the plausibility of sentences is

useful.

klklk

Attachments