Incorporating Attribution Importance for Improving Faithfulness Metrics

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4732–4745 July 9-14, 2023 ©2023 Association for Computational Linguistics Incorporating Attribution Importance for Improving Faithfulness Metrics Zhixue Zhao Nikolaos Aletras Department of Computer Science, University of Sheffield United Kingdom {zhixue.zhao, n.aletras}@sheffield.ac.uk Abstract Feature attribution methods (FAs) are popu- lar approaches for providing insights into the model reasoning process of making predic- tions. 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 cri- terion, 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 ig- nores the importance of each individual token, treating them all equally for computing suf- ficiency and comprehensiveness. In this pa- per, we propose a simple yet effective soft erasure criterion. Instead of entirely remov- ing or retaining tokens from the input, we ran- domly mask parts of the token vector repre- sentations proportionately to their FA impor- tance. Extensive experiments across various natural language processing tasks and differ- ent FAs show that our soft-sufficiency and soft- comprehensiveness metrics consistently prefer more faithful explanations compared to hard sufficiency and comprehensiveness.1 1 Introduction Feature attribution methods (FAs) are popular post- hoc explanation methods that are applied after model training to assign an importance score to each token in the input (Kindermans et al., 2016; Sundararajan et al., 2017). These scores indicate how much each token contributes to the model pre- diction. Typically, the top-k ranked tokens are then selected to form an explanation, i.e. rationale (DeY- oung et al., 2020). However, it is an important challenge to choose a FA for a natural language processing (NLP) task at hand (Chalkidis et al., 2021; Fomicheva et al., 2022) since there is no sin- 1Our code: https://github.com/casszhao/So ftFaith Figure 1: Hard and soft erasure criteria for comprehen- siveness and sufficiency for two toy feature attribution (FA) methods A and B. gle FA that is consistently more faithful (Atanasova et al., 2020). To assess whether a rationale extracted with a given FA is faithful, i.e. actually reflects the true model reasoning (Jacovi and Goldberg, 2020), vari- ous faithfulness metrics have been proposed (Arras et al., 2017; Serrano and Smith, 2019; Jain and Wallace, 2019; DeYoung et al., 2020). Sufficiency and comprehensiveness (DeYoung et al., 2020), also referred to as fidelity metrics (Carton et al., 2020), are two widely used metrics which have been found to be effective in capturing rationale faithfulness (Chrysostomou and Aletras, 2021a; Chan et al., 2022). Both metrics use a hard era- sure criterion for perturbing the input by entirely removing (i.e. comprehensiveness) or retaining (i.e. sufficiency) the rationale to observe changes in predictive likelihood. However, the hard erasure criterion ignores the different importance of each individual token, treat- ing them all equally for computing sufficiency and comprehensiveness. Moreover, …

Content truncated to 500 words