Is Attention Interpretable?

Sofia Serrano, Noah A. Smith

2019 · 115 citations · 29 references

DOIFull text

Open access

Concepts

TL;DR

Attention mechanisms have recently boosted performance on a range of NLP tasks, and because they explicitly weight input components, they are often assumed to reveal which information models find important. We test whether the assumption that attention reveals important information holds by manipulating attention weights in pre‑trained text classification models. We analyze the resulting differences in their predictions. We find that higher attention weights sometimes correlate with greater impact on predictions, but often they do not, and gradient‑based rankings better predict effects than raw magnitudes, concluding that attention is a noisy but not fail‑safe indicator of input importance.

Abstract

Attention mechanisms have recently boosted performance on a range of NLP tasks. Because attention layers explicitly weight input components’ representations, it is also often assumed that attention can be used to identify information that models found important (e.g., specific contextualized word tokens). We test whether that assumption holds by manipulating attention weights in already-trained text classification models and analyzing the resulting differences in their predictions. While we observe some ways in which higher attention weights correlate with greater impact on model predictions, we also find many ways in which this does not hold, i.e., where gradient-based rankings of attention weights better predict their effects than their magnitudes. We conclude that while attention noisily predicts input components’ overall importance to a model, it is by no means a fail-safe indicator.

References

29