Let the CAT out of the bag: Contrastive Attributed explanations for Text

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7190 - 7206 December 7-11, 2022 ©2022 Association for Computational Linguistics Let the CAT out of the bag: Contrastive Attributed explanations for Text Saneem A. Chemmengath1∗ Amar Prakash Azad2 Ronny Luss2 Amit Dhurandhar2 1Microsoft 2IBM Research schemmengath@microsoft.com, amarazad@in.ibm.com, {rluss, adhuran}@us.ibm.com Abstract Contrastive explanations for understanding the behavior of black box models has gained a lot of attention recently as they provide po- tential for recourse. In this paper, we propose a method Contrastive Attributed explanations for Text (CAT) which provides contrastive ex- planations for natural language text data with a novel twist as we build and exploit attribute classifiers leading to more semantically mean- ingful explanations. To ensure that our con- trastive generated text has the fewest possible edits with respect to the original text, while also being fluent and close to a human gener- ated contrastive, we resort to a minimal pertur- bation approach regularized using a BERT lan- guage model and attribute classifiers trained on available attributes. We show through qual- itative examples and a user study that our method not only conveys more insight because of these attributes, but also leads to better qual- ity (contrastive) text. Quantitatively, we show that our method outperforms other state-of- the-art methods across four data sets on four benchmark metrics. 1 Introduction Explainable AI (XAI) has seen an explosion of in- terest over the last five years, not just in research (Molnar, 2019; Arya et al., 2019), but also in the real world where governments (Yannella and Ka- gan, 2018; Gunning, 2017) and industry have made sizeable investments. The primary driver for this level of interest has been the inculcation of deep learning technologies (Goodfellow et al., 2016), which are inherently black box, into decision mak- ing systems that affect billions of people. Trust thus has become a central theme in relying on these black box systems, and one way to achieve it is seemingly through obtaining explanations. Although many feature-based (Ribeiro et al., 2016; Lundberg and Lee, 2017; Simonyan et al., ∗1Work done while at IBM Research. 2013) and exemplar-based methods (Gurumoor- thy et al., 2019; Koh and Liang, 2017; Kim et al., 2016) have been proposed to explain local instance level decisions of black box models, contrastive/- counterfactual explanations have seen a surge of interest recently (Wachter et al., 2017; Dhurand- har et al., 2018; Madaan et al., 2021; Luss et al., 2021; Ross et al., 2021). One reason for this is that contrastive explanations are viewed as one of the main tools to achieve recourse (Karimi et al., 2021). For example, companies commonly use chatbots to communicate with customers, which starts by pass- ing customer text through a classifier that decides which support department should handle the cus- tomer. A common problem is locating bias in these classifier models since “the chatbot will continue to show the behavior” due to biased knowledge bases(Brown, 2021); recourse in terms of remov- ing bias could be achieved by …

Content truncated to 500 words