Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Incorporating Attribution Importance for Improving Faithfulness Metrics

Zhixue Zhao, Nikolaos Aletras

TL;DR

This paper tackles the problem of faithfulness evaluation for feature attributions in NLP by showing that hard erasure can misrepresent token importance and degrade robustness to distribution shifts. It introduces Soft Normalized Sufficiency (Soft-NS) and Soft Normalized Comprehensiveness (Soft-NC), which perturb inputs at the embedding level using FA-guided dropout rather than full token deletion. The authors formalize these metrics, provide a comprehensive experimental study across SST, AG, Ev.Inf., and MultiRC using multiple FAs, and demonstrate that Soft-NS and Soft-NC yield higher diagnosticity and more faithful explanations than traditional NS/NC and hard perturbations. The work advances practical faithfulness evaluation and offers code for reproducibility, with implications for more reliable explanation methods in real-world NLP tasks.

Abstract

Feature attribution methods (FAs) are popular approaches for providing insights into the model reasoning process of making predictions. The more faithful a FA is, the more accurately it reflects which parts of the input are more important for the prediction. Widely used faithfulness metrics, such as sufficiency and comprehensiveness use a hard erasure criterion, i.e. entirely removing or retaining the top most important tokens ranked by a given FA and observing the changes in predictive likelihood. However, this hard criterion ignores the importance of each individual token, treating them all equally for computing sufficiency and comprehensiveness. In this paper, we propose a simple yet effective soft erasure criterion. Instead of entirely removing or retaining tokens from the input, we randomly mask parts of the token vector representations proportionately to their FA importance. Extensive experiments across various natural language processing tasks and different FAs show that our soft-sufficiency and soft-comprehensiveness metrics consistently prefer more faithful explanations compared to hard sufficiency and comprehensiveness. Our code: https://github.com/casszhao/SoftFaith

Incorporating Attribution Importance for Improving Faithfulness Metrics

TL;DR

This paper tackles the problem of faithfulness evaluation for feature attributions in NLP by showing that hard erasure can misrepresent token importance and degrade robustness to distribution shifts. It introduces Soft Normalized Sufficiency (Soft-NS) and Soft Normalized Comprehensiveness (Soft-NC), which perturb inputs at the embedding level using FA-guided dropout rather than full token deletion. The authors formalize these metrics, provide a comprehensive experimental study across SST, AG, Ev.Inf., and MultiRC using multiple FAs, and demonstrate that Soft-NS and Soft-NC yield higher diagnosticity and more faithful explanations than traditional NS/NC and hard perturbations. The work advances practical faithfulness evaluation and offers code for reproducibility, with implications for more reliable explanation methods in real-world NLP tasks.

Abstract

Feature attribution methods (FAs) are popular approaches for providing insights into the model reasoning process of making predictions. The more faithful a FA is, the more accurately it reflects which parts of the input are more important for the prediction. Widely used faithfulness metrics, such as sufficiency and comprehensiveness use a hard erasure criterion, i.e. entirely removing or retaining the top most important tokens ranked by a given FA and observing the changes in predictive likelihood. However, this hard criterion ignores the importance of each individual token, treating them all equally for computing sufficiency and comprehensiveness. In this paper, we propose a simple yet effective soft erasure criterion. Instead of entirely removing or retaining tokens from the input, we randomly mask parts of the token vector representations proportionately to their FA importance. Extensive experiments across various natural language processing tasks and different FAs show that our soft-sufficiency and soft-comprehensiveness metrics consistently prefer more faithful explanations compared to hard sufficiency and comprehensiveness. Our code: https://github.com/casszhao/SoftFaith
Paper Structure (35 sections, 7 equations, 3 figures, 6 tables)

This paper contains 35 sections, 7 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Feature Attribution Methods
  4. Measuring Faithfulness
  5. Evaluating Faithfulness Metrics
  6. Faithfulness Evaluation Metrics
  7. Sufficiency and Comprehensiveness
  8. Normalized Sufficiency (NS):
  9. Normalized Comprehensiveness (NC):
  10. Soft Nomralized Sufficiency and Comprehensiveness
  11. Soft Input Perturbation:
  12. Soft Normalized Sufficiency (Soft-NS):
  13. Soft Normalized Comprehensiveness (Soft-NC):
  14. Experimental Setup
  15. Tasks
  16. ...and 20 more sections

Figures (3)

  • Figure 1: Hard and soft erasure criteria for comprehensiveness and sufficiency for two toy feature attribution (FA) methods A and B.
  • Figure 2: The impact of rationale length on normalized comprehensiveness (NC) and sufficiency (NS). Each line represents a FA.
  • Figure 3: The impact of rationale length (shown in ratio) on Diagnosticity scores.