Office 95's natural-language help system employs Bayes' Rule, an equation often used in artificial intelligence systems (including speech and pattern recognition, decision analysis, and expert systems) to infer probabilities.
You start with known probabilities and causal relationships. For instance, let's say someone has a headache and you want your system to diagnose the reason for the malady. The patient is a hopeless hypochondriac and thinks his headache indicates a brain tumor. You know some basic facts about these conditions. For instance, the probability that a person with a brain tumor will have a headache is 90 percent, or in proper notation:
P(headache | brain tumor) = .9
You also know the random probability of a person having
a headache (let's say 10 percent) and of a person having a brain tumor (one in 1000). From these two unconditional probabilities and the causal relationship, Bayes' formula derives an unknown probability; in this case, the probability of a person with a headache having a brain tumor:
P(brain tumor | headache) = (P(headache | brain
tumor)*P(brain tumor))/P(headache), or
P(brain tumor | headache) = (.9*.001)/.1=.009
Using Bayes' formula, we now know that nine out of every 1000 people afflicted with a headache actually have a brain tumor. More to the point, we have derived an unknown probability from two known independent probabilities and a causal relationship (or joint probability).
Bayesian updating
lets us add new information dynamically (one piece at a time) until we reach an acceptable level of certainty. The system can then determine if new information is needed or if a reasonable decision can be made with the current information.
Inside Help
