TalkDown: A Corpus for Condescension Detection in Context

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3711–3719, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 3711 TalkDown: A Corpus for Condescension Detection in Context Zijian Wang Symbolic Systems Program Stanford University zijwang@stanford.edu Christopher Potts Department of Linguistics Stanford University cgpotts@stanford.edu Abstract Condescending language use is caustic; it can bring dialogues to an end and bifurcate com- munities. Thus, systems for condescension de- tection could have a large positive impact. A challenge here is that condescension is often impossible to detect from isolated utterances, as it depends on the discourse and social con- text. To address this, we present TALKDOWN, a new labeled dataset of condescending lin- guistic acts in context. We show that extending a language-only model with representations of the discourse improves performance, and we motivate techniques for dealing with the low rates of condescension overall. We also use our model to estimate condescension rates in various online communities and relate these differences to differing community norms. 1 Introduction Condescending language use can derail conversa- tions and, over time, disrupt healthy communi- ties. The caustic nature of this language traces in part to the ways that it keys into differing so- cial roles and levels of power (Fournier et al., 2002). It is common for people to be condescend- ing without realizing it (Wong et al., 2014), but a lack of intent only partly mitigates the dam- age it can cause. Thus, condescension detection is a potentially high-impact NLP task that could open the door to many applications and future re- search directions, including, for example, support- ing productive interventions in online communi- ties (Spertus, 1997), educating people who use condescending language in writing, helping lin- guists to understand the implicit linguistic acts as- sociated with condescension, and helping social scientists to study the relationship between con- descension and other variables like gender or so- cioeconomic status. Progress on this task is currently limited by a >Are you struggling with this whole English language thing? Stop being so condescending and engage in a real discussion. Comment: You might just admit that you misunderstood. Are you struggling with this whole English language thing? Reply: Context Quoted Accusation of Condescension Figure 1: In this example, the REPLY quotes from part of the COMMENT and says that this QUOTED text is condescending. lack of high-quality labeled data. A deeper chal- lenge is that condescension is often impossible to detect from isolated utterances. First, a charac- teristic of condescending language is that it is not overtly negative or critical – it might even include (insincere) praise (Huckin, 2002). Second, con- descension tends to rest on a pair of conflicting pragmatic presuppositions: a speaker presumption that the speaker has higher social status than the listener, and a listener presumption that this is in- correct. For example, an utterance that is entirely friendly if said by one friend to another might be perceived as highly …

Content truncated to 500 words