People's actions are motivated by goals -- partial descriptions of desired or avoided world states. They can vary in specificity from the very abstract (e.g. "eat some food") to extremely concrete (e.g., "eat a fried Walter's hotdog with relish mustard in Rye, NY"). Goals are subjective and an individual's goals change over time. They are acquired through many kinds of learning: they can be transmitted and evolve through social mediums, but are ultimately validated by biological reinforcement signals.

Although nobody yet knows how goals are represented in human brains, we can represent their interrelationships using two kinds of links. Analogous to how a object visually is a collection of parts arranged in space, a goal is composed of a series of sub-goals in time, possibly with ordering constraints. For example, in order to "eat at a restaurant" one must first "go to the restaurant" "order food" and then "eat the food". Disjoint sub-goals (eat at home versus eat at restaurant) that implement a common parent goal (eat dinner) are related by the 'is-a' inheritance relationship. We represent relations between goals by composition (part of) relationships and taxonomic (is-a) relationships. A key problem is developing methods to distinguish these two types of relationships.

This project aims to construct a goal network that relates desired actions by composition (part-of) and taxonomic (is-a) relationships. Goals are also annotated with properties relating their desirability, common locations, how long they take place, and other information. We aim to have a complete corpus of common human goals.

The results of the text mining and goal construction can be viewed by running the flex application, GoalViz, in the visualization directory along with the backend server and goal information retrieval system: server/server.py

If you are interested in working on this project, please contact Dustin Smith.