What does a Text Classifier Learn about Morality? An Explainable Method for Cross-Domain Comparison of Moral Rhetoric

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 14113–14132 July 9-14, 2023 ©2023 Association for Computational Linguistics What does a Text Classifier Learn about Morality? An Explainable Method for Cross-Domain Comparison of Moral Rhetoric Enrico Liscio1, Oscar Araque2, Lorenzo Gatti3, Ionut Constantinescu4, Catholijn M. Jonker1,6, Kyriaki Kalimeri5, and Pradeep K. Murukannaiah1 1TU Delft, Delft, the Netherlands 2Universidad Politécnica de Madrid, Madrid, Spain 3University of Twente, Enschede, the Netherlands 4ETH Zürich, Zürich, Switzerland 5ISI Foundation, Turin, Italy 6Leiden University, Leiden, the Netherlands {e.liscio,c.m.jonker,p.k.murukannaiah}@tudelft.nl o.araque@upm.es l.gatti@utwente.nl iconstantinescu100@gmail.com kyriaki.kalimeri@isi.it Abstract Moral rhetoric influences our judgement. Al- though social scientists recognize moral ex- pression as domain specific, there are no sys- tematic methods for analyzing whether a text classifier learns the domain-specific expres- sion of moral language or not. We propose Tomea, a method to compare a supervised clas- sifier’s representation of moral rhetoric across domains. Tomea enables quantitative and qual- itative comparisons of moral rhetoric via an in- terpretable exploration of similarities and dif- ferences across moral concepts and domains. We apply Tomea on moral narratives in thirty- five thousand tweets from seven domains. We extensively evaluate the method via a crowd study, a series of cross-domain moral classifi- cation comparisons, and a qualitative analysis of cross-domain moral expression. 1 Introduction Moral narratives play a fundamental role in stance taken on controversial social issues (Fulgoni et al., 2016). Recognizing moral narratives helps under- stand the argumentation around important topics such as vaccine hesitancy (Kalimeri et al., 2019b), violent protests (Mooijman et al., 2018), and cli- mate change (Dickinson et al., 2016). Language reveals deep psychological constructs, including moral values (Graham et al., 2013). Thus, language is an important avenue for analyzing moral expression. In particular, supervised text classification models have been showing promising results on morality prediction (Lourie et al., 2021; Hendrycks et al., 2021; Alshomary et al., 2022). These models leverage the wisdom of crowds (via annotations of moral expression) to attain a de- scriptive understanding of morality. However, the supervised learning paradigm can lead to black-box models (Danilevsky et al., 2020). Understanding what these models learn is crucial, especially for the morality classification task, which is likely to be used in sensitive applications like healthcare (Wen et al., 2019; Carriere et al., 2021). Moral expression is context dependent (Hill and Lapsley, 2009; Brännmark, 2015; Kola et al., 2022), where context refers to factors such as actors, ac- tions, judges, and values (Schein, 2020). For a text classifier, the domain from which the training data is sourced represents the context. For example, in the context of recent Iranian protests, tweets tagged #mahsaamini can form the training domain. We expect this domain to have a different moral expres- sion than the training domain of #prolife tweets, representing a different context. Recent works (Liscio et al., 2022a; Huang et al., 2022) analyze the out-of-domain performance of morality classifiers. However, what leads classi- fiers to perform differently across domains has not been systematically explored. Such an …

Content truncated to 500 words