PhD Proposal: Rethinking Evaluation and Interpretability in NLP

Recognizing issues in existing approaches, such as the discovery of adversarial examples, has been crucial to the development of better problem formulations and more robust models. Despite the recent interest in adversarial evaluation and interpretation for NLP models, there is still large room for improvement. For example, it's not clear how we can formulate the NLP counterpart of certifiable robustness from vision. In this talk, I'll first take a critical look at how we currently do interpretation and evaluation in NLP. Specifically, we look at how methods that are supposed to help us understand the models can mislead us. Then I'll present some preliminary results on how we might move towards fixing these issues.Examining Committee:

Chair: Dr. Jordan Boyd-Graber Dept rep: Dr. John Dickerson Members: Dr. Hal Daumé III Dr. Marine Carpuat Dr. Alexander Rush