Recent advances in NLU focus on representation learning, where numerical distributions representing words or sentences are learned using a machine learning model. However, these representations do not have an explicit structure (internal schematic representation) and are hence not interpretable, which impedes human understanding and theoretical advancement of the field. This leads to a central question in AI of whether these high-performing models are able to capture and reason over the implications of events in the real world, or whether they simply memorize all the training examples and perform a small amount of generalization.
In this thesis, we address the problem of creating humanly-interpretable representations of knowledge, and of exploiting representation structure to retain only the parts of the encoded information that are relevant to a task at hand. We propose methods to create more explainable representations that distinguish aspects of meaning based on the specific task in question. Our proposal focuses on representations that express causal inferences of events and how different aspects of meaning direct the impact of an event on entities. We study representations at different levels of semantics (lexical/conceptual, sentence, and discourse); representations for words, sentences, and large-scale complex events.
Eduard Hovy (Chair)
Alan Ritter (Georgia Institute of Technology)
Zoom Participation. See announcement.