Social Norms-Grounded Machine Ethics in Complex Narrative Situation

Proceedings of the 29th International Conference on Computational Linguistics, pages 1333–1343 October 12–17, 2022. 1333 Social Norms-Grounded Machine Ethics in Complex Narrative Situation Tao Shen, Xiubo Geng, Daxin Jiang∗ Microsoft Corporation {shentao,xigeng,djiang}@microsoft.com Abstract Ethical judgment aims to determine if a per- son in a narrative situation acts under people’s social norms under a culture, so it is crucial to understand actions in narratives and achieve machine ethics. Recent works depend on data- driven methods to directly judge the ethics of complex real-world narratives but face two ma- jor challenges. First, they cannot well handle dilemma situations due to a lack of basic knowl- edge about social norms. Second, they focus merely on sparse situation-level judgment re- gardless of the social norms involved during the judgment, leading to a black box. In this work, inspired by previous knowledge-grounded and -augmented paradigms, we propose to comple- ment a complex situation with grounded social norms. Besides a norm-grounding knowledge model, we present a novel norm-supported eth- ical judgment model in line with neural mod- ule networks to alleviate dilemma situations and improve norm-level explainability. Em- pirically, our model improves state-of-the-art performance on two narrative judgment bench- marks. 1 Introduction In natural language processing (NLP) literature, ethical judgment aims to determine if a person (e.g., narrator or someone else) in a narrative sit- uation is morally wrong or correct (Lourie et al., 2021). For example, in a narrative situation “I helped him but got taunted”, the narrator is morally good while the other is bad. It attracts more in- terest from academia and industry as it plays an indispensable role in human-centric applications and benefits a wide range of downstream tasks, e.g., dialogue systems and storytelling. Recently, Forbes et al. (2020) propose a gener- ative model, NORM TRANSFORMER, which can extract actions1 from a narrative situation and then ∗Corresponding author. 1In this paper, “action” denotes a verb-centric “event” Figure 1: Two situations and the involved social norms. The first one is adapted from Forbes et al. (2020) while the second one is grounded by our model, where the norm in red is context-irrelevant to the situation. judge the ethics towards the actions. However, it can only generate action-level ethical judgment for simple narrative situations (i.e., sentences with limited events), so it usually fails to perform in complex narrative situations of many real-world ap- plications. Here, “complex” is usually reflected in over-long narrative contexts (multiple paragraphs) and/or dilemma situations. Take Situation 1 in Fig- ure 1 as a dilemma example: although we keep in mind that “it’s bad to punch others”, we cannot conclude the narrator is morally bad due to “it’s bad to steal”. On the contrary, Lourie et al. (2021) propose a data-driven method to directly judge complex narra- tive situations (e.g., real-life anecdotes from the In- ternet). Empowered by pre-trained language mod- els (e.g., BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019)), the proposed method achieves satisfactory performance to boost its real-world ap- plications but encounters two major challenges …

Content truncated to 500 words