From Polarity to Intensity: Mining Morality from Semantic Space

Proceedings of the 29th International Conference on Computational Linguistics, pages 1250–1262 October 12–17, 2022. 1250 From Polarity to Intensity: Mining Morality from Semantic Space Chunxu Zhao, Pengyuan Liu, Dong Yu� School of Information Science, Beijing Language and Culture University, China chunxu1212@gmail.com, liupengyuan@pku.edu.cn, yudong_blcu@126.com Abstract Most works on computational morality focus on moral polarity recognition, i.e., distinguish- ing right from wrong. However, a discrete polarity label is not informative enough to re- flect morality as it does not contain any de- gree or intensity information. Existing ap- proaches to compute moral intensity are limited to word-level measurement and heavily rely on human labelling. In this paper, we propose MORALSCORE, a weakly-supervised frame- work1 that can automatically measure moral intensity from text. It only needs moral polar- ity labels, which are more robust and easier to acquire. Besides, the framework can capture la- tent moral information not only from words but also from sentence-level semantics which can provide a more comprehensive measurement. To evaluate the performance of our method, we introduce a set of evaluation metrics and con- duct extensive experiments. Results show that our method achieves good performance on both automatic and human evaluations. 1 Introduction Moral intensity is a degree of feeling that a person has about a behaviour (Barnett, 2001). As shown in Figure 1, although speeding on streets and killing a child are both immoral, the latter is more se- vere in most people’s perception. Understanding the above difference is an ability that humans have gradually developed in everyday life. It affects indi- viduals’ ethical judgments and reflects the ideology of our society (Jones, 1991). As AI gets ever more involved in people’s lives, it has become increas- ingly important for machine to acquire this ability and behave ethically. Researchers have studied the problem from early rule-based methods to today’s deep learning-based paradigms (Yu et al., 2018; Hendrycks et al., 2021). It remains a fundamental but unsolved problem in computational morality (Moor, 2006). 1https://github.com/blcunlp/MoralScore Bob donates 1M$ to charity. Ivy puts the rubbish in the dustbin. Bob donates 1M$ to charity. Joe is speeding on city streets. Tom kills a child. Polarity Intensity Polarity Intensity 0 0 1 1 0.1 0.4 0.7 0.9 Text Figure 1: Moral Intensity Example. From the numerical measurement of morality, both moral polarity and its degree can be reflected. Previous work in the NLP community often treats this problem as a supervised text classifi- cation task, i.e., judging the moral polarity for a text (Xie et al., 2020; Nahian et al., 2021). This way of modelling morality is inadequate because it oversimplifies morality into a Bernoulli distribu- tion, i.e., being only moral or immoral. We model morality into a continuous distribution by introduc- ing moral intensity to include degree information. Computing moral intensity is challenging in two aspects: 1) In supervised settings, unlike labelling moral polarity, building a large corpus with precise intensity values is time consuming and prone to subjectivity. 2) In unsupervised settings, there is no direct link between text and …

Content truncated to 500 words