Attention is not Explanation
This work challenges a common assumption in NLP, showing that attention mechanisms are not reliable for model interpretability, which is crucial for researchers and practitioners relying on transparency.
The paper investigates whether attention weights in NLP models provide meaningful explanations for predictions, finding that they largely do not, as attention distributions often correlate poorly with feature importance and can vary widely without affecting predictions.
Attention mechanisms have seen wide adoption in neural NLP models. In addition to improving predictive performance, these are often touted as affording transparency: models equipped with attention provide a distribution over attended-to input units, and this is often presented (at least implicitly) as communicating the relative importance of inputs. However, it is unclear what relationship exists between attention weights and model outputs. In this work, we perform extensive experiments across a variety of NLP tasks that aim to assess the degree to which attention weights provide meaningful `explanations' for predictions. We find that they largely do not. For example, learned attention weights are frequently uncorrelated with gradient-based measures of feature importance, and one can identify very different attention distributions that nonetheless yield equivalent predictions. Our findings show that standard attention modules do not provide meaningful explanations and should not be treated as though they do. Code for all experiments is available at https://github.com/successar/AttentionExplanation.