Truly Smart Agents
research directions > learning
learning

Our goal is to make our artificial intelligent agents capable of learning both concepts and facts in a variety of modes:

Learning by being told:
the agent learns by understanding dialog;  this capability is already present to a degree in the latest version of MVP.
Learning by experience:
the agent carries out an action and learns about its effects.
Learning by reading: 
the agent learns concepts and word meanings by reading texts. While this capability has not yet been incorporated in the latest version of MVP, we have implemented an initial version of a system for learning words and concepts from text. We view this capability as essential for overcoming - or at least alleviating - the knowledge bottleneck (the expense of manual knowledge acquisition) and for recovering from unexpected input, that is, dealing with unknown words and phrases in processing language inputs.

gallery (click to enlarge)

An sample of quality of automatically learned lexical senses.
A sample of quality of automatically learned lexical senses.
 Automatic extraction of ontological and lexical resources.
Automatic extraction of ontological and lexical resources.

relevant publications:

[1] English, J. and Nirenburg, S. Striking a balance: Human and computer contributions to learning through semantic analysis. 2010. Proceedings of the Fourth IEEE International Conference on Semantic Computing, Carnegie Mellon University, Pittsburgh, PA, Sept. 22-24.

[2] English, J. and Nirenburg, S. Ontology Learning from Text Using Automatic Ontological-Semantic Text Annotation and the Web as the Corpus. 2007. Proceedings of the AAAI-07 Spring Symposium on Machine Reading.