ENTRUST: Argument Reframing with Language Models and Entailment

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4958–4971 June 6–11, 2021. ©2021 Association for Computational Linguistics 4958 ENTRUST: Argument Reframing with Language Models and Entailment Tuhin Chakrabarty1, Christopher Hidey 3 Smaranda Muresan1,2 Department of Computer Science, Columbia University 1 Data Science Institute, Columbia University 2 Google 3 ∗ {tuhin.chakr, smara}@cs.columbia.edu chris.hidey@gmail.com Abstract Framing involves the positive or negative pre- sentation of an argument or issue depending on the audience and goal of the speaker (Ent- man, 1983). Differences in lexical framing, the focus of our work, can have large ef- fects on peoples’ opinions and beliefs. To make progress towards reframing arguments for positive effects, we create a dataset and method for this task. We use a lexical re- source for connotations to create a parallel cor- pus and propose a method for argument re- framing that combines controllable text gen- eration (positive connotation) with a postde- coding entailment component (same denota- tion). Our results show that our method is ef- fective compared to strong baselines along the dimensions of fluency, meaning, and trustwor- thiness/reduction of fear. 1 Introduction Public opinion has been shown to be significantly influenced by framing effects. Framing refers to the presentation of an issue, where even small changes may have outsized effects on beliefs (Chong and Druckman, 2007). For example, when asked about “welfare,” the American public is largely against increasing spending (with only 20% in favor), but when asked about “assistance to the poor,” 65% believe that the government is not spending enough (Rasinski, 1989). While other research has focused on syntactic framing (Greene and Resnik, 2009) or issue fram- ing (Hartmann, 2019), we focus specifically on lex- ical framing, distinguishing sentences by their con- notative meaning even where they have the same denotative meaning. According to Frege (1892), two sentences with the same truth conditions may refer to the same entities or state of affairs (“refer- ence,” also known as denotation) but be presented ∗The work is not affiliated to Google and was conducted independently outside of the organization Arg1 Alabama’s Supreme Court Chief Justice was suspended... for ordering state probate judges not to grant marriage licenses to gay couples... Rf Arg1 Alabama’s Supreme Court Chief Justice was suspended... for ordering state probate judges not to grant legal marriage equality to gay couples... Arg2 Every nation with territorial claims in the arctic is a member of NATO, except Russia. Rf Arg2 Every nation with sovereign competence in the arctic is a member of NATO, except Russia. Arg3 At this dire moment , we all need to amplify our voices in defense of free speech . Rf Arg3 At this crucial moment , we all need to amplify our voices in support of free speech. Arg4 It is difficult to think of any single act that would do more to restore America’s soft power than the election of Obama to the presidency Rf Arg4 It is difficult to think of any …

Content truncated to 500 words