“Nice Try, Kiddo”: Investigating Ad Hominems in Dialogue Responses

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 750–767 June 6–11, 2021. ©2021 Association for Computational Linguistics 750 “Nice Try, Kiddo”: Investigating Ad Hominems in Dialogue Responses Emily Sheng1, Kai-Wei Chang2, Premkumar Natarajan1, Nanyun Peng1,2 1 Information Sciences Institute, University of Southern California 2 Computer Science Department, University of California, Los Angeles {ewsheng,pnataraj}@isi.edu, {kwchang,violetpeng}@cs.ucla.edu Abstract Ad hominem attacks are those that target some feature of a person’s character instead of the position the person is maintaining. These at- tacks are harmful because they propagate im- plicit biases and diminish a person’s credi- bility. Since dialogue systems respond di- rectly to user input, it is important to study ad hominems in dialogue responses. To this end, we propose categories of ad hominems, compose an annotated dataset, and build a classifier to analyze human and dialogue sys- tem responses to English Twitter posts. We specifically compare responses to Twitter top- ics about marginalized communities (#Black- LivesMatter, #MeToo) versus other topics (#Vegan, #WFH), because the abusive lan- guage of ad hominems could further amplify the skew of power away from marginalized populations. Furthermore, we propose a con- strained decoding technique that uses salient n-gram similarity as a soft constraint for top-k sampling to reduce the amount of ad hominems generated. Our results indicate that 1) responses from both humans and DialoGPT contain more ad hominems for discussions around marginalized communities, 2) different quantities of ad hominems in the training data can influence the likelihood of generating ad hominems, and 3) we can use constrained de- coding techniques to reduce ad hominems in generated dialogue responses. 1 Introduction Ad hominems attack an opponent’s character or identity instead of the points the opponent is mak- ing, and can exist in any conversational setting between two or more entities. From an argumen- tation perspective, ad hominems are fallacies, and fallacies rely on faulty reasoning to advance a point (Hansen, 2020). These ad hominem fallacies are related to abusive language, toxicity, and microag- gressions, and can be expressed with both subtle and explicitly offensive language. Table 1 presents Post: Many are trying to co-opt and mischaracterize the #blacklivesmatter movement. We won’t allow it! Resp: I hate how much of a victim complex you guys have. Post: You’re the reason we need the #MeToo movement. Resp: Nice try, kiddo. Post: Stop eating them if you don’t want them to go ex- tinct! #govegan Resp: I don’t like your username Table 1: Ad hominem responses to Twitter posts. examples of ad hominem responses to Twitter posts. Undesirable in any response, ad hominems are un- productive in furthering a meaningful discussion and can reinforce falsehoods. However, these at- tacks appeal to emotions and implicit biases to ar- gue a point, and are thus often effectively harmful regardless of whether the attacks are true, recog- nized, or retracted (Yap, 2013). Our work is motivated by this fallacy’s potential to amplify the spread of harmful societal biases. For communities that …

Content truncated to 500 words