I'm working on understanding both representation learning systems (blackbox systems) and vector representations. For example, matrix factorization systems, word embeddings, and natural language inference systems.
More concretely, I worked on:
a) Extracting knowledge from a matrix factorization model, so one could understand why this complex model made a particular prediction. This knowledge is in the form of an interpretable model such as logic rules, decision trees and Bayesian networks. These interpretable models allow us to explain how the matrix factorization model arrived to a particular output; thus, we can explain how the matrix factorization model made incorrect predictions. (AAAI spring symposium and Cognitive Computation workshop papers)
b) Understanding to what extent pretrained word embeddings encode a hypernymy relation; i.e., given the word embeddings of two concepts, can we know if the isa relationship holds between these two concepts only by inspecting the vector representations? For example, can we know that a cat is an animal just by inspecting their respective word embeddings? Well, it turns out that into some extent we can. (EACL short paper 2017)
c) Understanding what factors affect the robustness of natural language inference systems: Are NLI systems as robust as their test set accuracy shows? Not really. Three neural systems analyzed (one of them a stateoftheart system) are shown to be insensitive to a simple transformation to input data; also, these models pick a bias from the training set and they seem to struggle with unseen antonym word pairs; though, not very surprisingly (as discussed in the EACL 2017 paper) these models seem to learn hypernymy. (NAACL long paper 2018)
News
