Responsible Methods

Deep learning models have revolutionized the field of Natural Language Processing, but their blackbox nature also makes it notoriously difficult to interpret and explain their inner workings. This limitation has become the central obstacle for responsible use of the technology, even as more and more companies and institutions are using it, and the outcomes of these systems are starting to really affect people’s lives. Without interpretability, (i) there is no transparency about the outputs of these systems, (ii) it is difficult to integrate existing expertise and (iii) users cannot easily adapt the systems to their needs. We develop methods that yield insights into existing deep learning models (“posthoc interpretability”) and identify ways to constrain these models such that they are more interpretable without losing their great performance (“explainability by design”).