Learning meaning in Genesis

Abstract

GENESIS is an existing software system for understanding and reasoning about language and vision. One of the modules in GENESIS uses about 1000 lines of Java code representing 31 rules to turn English sentences into a variety of more meaningful semantic representations. I reproduced the functionality of these rules by training the existing rule-learning program UNDERSTAND with 43 human-readable examples of English sentences and corresponding semantic representations, resulting in 18 rules. These new rules and the framework for training and using them provides GENESIS with a more robust and extensible semantic parser. This research also led me to make several improvements to UNDERSTAND, making it both more powerful and easier to train.