A Just and Comprehensive Strategy for Using NLP to Address Online Abuse

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3658–3666 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3658 A Just and Comprehensive Strategy for Using NLP to Address Online Abuse David Jurgens University of Michigan School of Information jurgens@umich.edu Eshwar Chandrasekharan Georgia Tech School of Interactive Computing eshwar3@gatech.edu Libby Hemphill University of Michigan School of Information libbyh@umich.edu Abstract Online abusive behavior affects millions and the NLP community has attempted to mitigate this problem by developing technologies to de- tect abuse. However, current methods have largely focused on a narrow definition of abuse to detriment of victims who seek both vali- dation and solutions. In this position paper, we argue that the community needs to make three substantive changes: (1) expanding our scope of problems to tackle both more subtle and more serious forms of abuse, (2) develop- ing proactive technologies that counter or in- hibit abuse before it harms, and (3) reframing our effort within a framework of justice to pro- mote healthy communities. 1 Introduction Online platforms have the potential to enable sub- stantial, prolonged, and productive engagement for many people. Yet, the lived reality on social media platforms falls far short of this potential (Papacharissi, 2004). In particular, the promise of social media has been hindered by antisocial, abu- sive behaviors such as harassment, hate speech, trolling, and the like. Recent surveys indicate that abuse happens much more frequently than many people suspect (40% of Internet users report be- ing the subject of online abuse at some point), and members of underrepresented groups are tar- geted even more often (Herring et al., 2002; Drake, 2014; Anti-Defamation League, 2019). The NLP community has responded by de- veloping technologies to identify certain types of abuse and facilitating automatic or computer- assisted content moderation. Current technol- ogy has primarily focused on overt forms of abu- sive language and hate speech, without consid- ering both (i) the success and failure of tech- nology beyond getting the classification correct, and (ii) the myriad forms that abuse can take. Risk of Physical Danger Frequency Hate speech Insults Condescension Microagression Physical Threats Doxxing Promoting Self Harm Figure 1: Abusive behavior online falls along a spec- trum, and current approaches focus only on a narrow range (shown in red text), ignoring nearby problems. Impact comes from both the frequency (on left) and real-world consequences (on right) of behaviors. This figure illustrates the spectrum of online abuse in an hypothetical manner, with its non-exhaustive examples inspired from prior surveys of online experiences (Dug- gan, 2017; Salminen et al., 2018). As Figure 1 shows, a large spectrum of abu- sive behavior exists—some with life-threatening consequences—much of which is currently unad- dressed by language technologies. Explicitly hate- ful speech is just one tool of hate, and related tac- tics such as rape threats, gaslighting, First Amend- ment panic, and veiled insults are effectively em- ployed both off- and online to silence, scare, and exclude participants from what should be …

Content truncated to 500 words