Abstract by Nathan Tibbetts
Analyzing and Improving Embedding Spaces
Word embedding spaces, like those produced by the Word2vec algorithm, have been shown to be an effective method of representing common knowledge for autonomous agents, allowing them to reason about the world and make decisions - but frequent faults in its reasoning have led us to seek methods of improving Word2vec. Subsequent efforts to develop new analogy algorithms revealed gaps in our current understanding of embedding spaces, motivating the need for an effective method of quantitatively characterizing them, as well as comparing their relative performances as we ultimately develop new training algorithms. Thus, my current research has been in building a toolset to analyze, quantify, and compare both the innate mathematical properties of embedding spaces and their performances on analogies.