Learning commonsense knowledge from the interpretation of individual experiences

Abstract

To understand human intelligence, we need to discover how we learn commonsense knowledge. We humans are able to infer generalizations of knowledge, make predictions about the future, and answer questions based on what we've experienced. In this thesis I present a method for performing these commonsense tasks in an artificial intelligence system. I introduce a data type called a chain which clusters together similar experiences. I use graphs to store readily available historical and causal relations for experiences. The resulting memory system can handle three types of commonsense reasoning tasks. It can generalize, going from two specific examples to the knowledge that all birds can fly. It can predict, hypothesizing that since a dog likes to bark at people, it will bark when a burglar appears. And it can answer questions, providing a response when asked about the location of my car. This memory system is encoded in approximately 2,000 lines of Java.