Mitigating Toxic Degeneration with Empathetic Data: Exploring the Relationship Between Toxicity and Empathy

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4926 - 4938 July 10-15, 2022 ©2022 Association for Computational Linguistics Mitigating Toxic Degeneration with Empathetic Data: Exploring the Relationship Between Toxicity and Empathy Allison Lahnala† and Charles Welch† and Béla Neuendorf and Lucie Flek Conversational AI and Social Analytics (CAISA) Lab Department of Mathematics and Computer Science, University of Marburg http://caisa-lab.github.io {allison.lahnala,welchc,neuendob,lucie.flek}@uni-marburg.de Abstract Content Warning: This paper includes exam- ples of religious-based discriminatory language that may be offensive and upsetting. Large pre-trained neural language models have supported the effectiveness of many NLP tasks, yet are still prone to generating toxic language hindering the safety of their use. Using empa- thetic data, we improve over recent work on controllable text generation that aims to reduce the toxicity of generated text. We find we are able to dramatically reduce the size of fine- tuning data to 7.5-30k samples while at the same time making significant improvements over state-of-the-art toxicity mitigation of up to 3.4% absolute reduction (26% relative) from the original work on 2.3m samples, by strategi- cally sampling data based on empathy scores. We observe that the degree of improvement is subject to specific communication components of empathy. In particular, the cognitive compo- nents of empathy significantly beat the original dataset in almost all experiments, while emo- tional empathy was tied to less improvement and even underperforming random samples of the original data. This is a particularly implica- tive insight for NLP work concerning empathy as until recently the research and resources built for it have exclusively considered empathy as an emotional concept. 1 Introduction Pre-trained neural language models are prone to generating toxic language, hindering the ability to use them safely (Gehman et al., 2020). Re- cent work on controllable text generation has shown promise in successfully altering such text attributes (Liu et al., 2021). However, partly due to the subjective nature of this task (Jurgens et al., 2019), the selection of negative, non-toxic exam- ples for modeling has been somewhat arbitrary. †Authors contributed equally. Meanwhile, there is a growing body of research in natural language processing around the concept of empathetic communication - a number of data resources and approaches have been proposed for training empathy recognition and generation mod- els (Sharma et al., 2020; Rashkin et al., 2019). Though the definitions of toxicity and empathy vary across literature, we observe an opposition between the concepts in terms of response appro- priateness and intent toward others, which is the basis of the research question driving this work: is there an opposing relationship between toxic and empathetic language that can be leveraged to better model these phenomena? Toxic language is often described as harassing or offensive language that decreases the likelihood of participation in discussions or other cooperative efforts (Wulczyn et al., 2017). In NLP literature, empathetic language is usually conveyed as lan- guage that shows an understanding and acknowl- edgement of the interlocutor’s emotions (Rashkin et al., 2019; …

Content truncated to 500 words