Flexible Instance-Specific Rationalization of NLP Models

Recent research on model interpretability in natural language processing extensively uses feature scoring methods for identifying which parts of the input are most important a to make prediction (i.e. explanation or rationale). However, previous has shown that there is no clear best method across various text classification tasks while practitioners typically have several other ad-hoc choices regarding length and type rationale (e.g. short long, contiguous not). Inspired by this, we propose simple yet effective flexible allows selecting optimally each data instance: (1) method; (2) length; (3) rationale. Our inspired erasure approaches assume faithful should be one with highest difference between model's output distribution using full after removing as respectively. Evaluation four standard datasets shows our proposed provides more faithful, comprehensive highly sufficient explanations compared fixed method, type. More importantly, demonstrate practitioner not required any order extract rationales approach.