Towards Faithfully Interpretable NLP Systems: How should we define and evaluate faithfulness?

With the growing popularity of deep-learning based NLP models, comes a need\nfor interpretable systems. But what is interpretability, and what constitutes a\nhigh-quality interpretation? In this opinion piece we reflect on the current\nstate of interpretability evaluation research. We call for more clearly\ndifferentiating between different desired criteria an interpretation should\nsatisfy, and focus on the faithfulness criteria. We survey the literature with\nrespect to faithfulness evaluation, and arrange the current approaches around\nthree assumptions, providing an explicit form to how faithfulness is "defined"\nby the community. We provide concrete guidelines on how evaluation of\ninterpretation methods should and should not be conducted. Finally, we claim\nthat the current binary definition for faithfulness sets a potentially\nunrealistic bar for being considered faithful. We call for discarding the\nbinary notion of faithfulness in favor of a more graded one, which we believe\nwill be of greater practical utility.\n