How Computers Calculate Probabilities in Large Knowledge Bases

A method for finding answers in a database of uncertain facts by ignoring probabilities to find a solution first, then calculating how likely that solution is based on the underlying evidence.

What it covers

This patent describes a way to query a knowledge base where facts (axioms) come with a probability score. Instead of trying to calculate complex probabilities for the entire database at once, the system first ignores the probability scores to find a logical solution using standard reasoning. Once it identifies the 'justification'—the minimal set of facts needed to prove that answer—it then applies the probability scores only to that specific subset. This makes it computationally efficient to determine the 'net probability' of a conclusion in massive datasets that would otherwise be too slow to process.

What it doesn't cover

• —Does not cover calculating probabilities for the entire knowledge base simultaneously without first finding a logical justification.
• —Does not cover systems that lack a defined set of axioms or logical rules for deriving answers.
• —Does not cover non-probabilistic reasoning systems that treat all facts as 100% certain.

The clever bit

The innovation is the 'divide and conquer' strategy: by decoupling the search for a logical answer from the mathematical calculation of its certainty, the system avoids the exponential complexity of calculating probabilities for every possible combination of facts.

Why it matters

As AI and expert systems deal with increasingly messy, real-world data, they often encounter conflicting or uncertain information. This approach allows developers to build systems that can provide 'likely' answers without getting bogged down in the massive mathematical overhead of traditional probabilistic modeling. It is a key technique for scaling semantic web technologies and enterprise knowledge graphs.

Real-world examples

• 1.Enterprise knowledge graphs used for automated decision support
• 2.Semantic search engines that rank results based on probabilistic confidence
• 3.Medical diagnostic software evaluating symptoms with varying levels of clinical certainty