Generative Entity-to-Entity Stance Detection with Knowledge Graph Augmentation

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9950 - 9969 December 7-11, 2022 ©2022 Association for Computational Linguistics Generative Entity-to-Entity Stance Detection with Knowledge Graph Augmentation Xinliang Frederick Zhang1, Nick Beauchamp2, and Lu Wang1 1Computer Science and Engineering, University of Michigan, Ann Arbor, MI 2Department of Political Science, Northeastern University, Boston, MA 1{xlfzhang,wangluxy}@umich.edu 2n.beauchamp@northeastern.edu Abstract Stance detection is typically framed as predict- ing the sentiment in a given text towards a target entity. However, this setup overlooks the importance of the source entity, i.e., who is expressing the opinion. In this paper, we emphasize the need for studying interactions among entities when inferring stances. We first introduce a new task, entity-to-entity (E2E) stance detection, which primes models to iden- tify entities in their canonical names and dis- cern stances jointly. To support this study, we curate a new dataset with 10,619 annotations labeled at the sentence-level from news articles of different ideological leanings. We present a novel generative framework to allow the gen- eration of canonical names for entities as well as stances among them. We further enhance the model with a graph encoder to summarize entity activities and external knowledge sur- rounding the entities. Experiments show that our model outperforms strong comparisons by large margins. Further analyses demonstrate the usefulness of E2E stance detection for un- derstanding media quotation and stance land- scape, as well as inferring entity ideology. 1 Introduction News media often employ ideological language to sway their readers, including criticizing entities they disagree with, and praising those conforming to their values (Baum and Groeling, 2008; Leven- dusky, 2013). However, in many cases, the sources do not directly express their sentiments, in part due to the norm that “objective” news media should restrict their role to narrating events and quoting others. In the realm of political news, many re- ported events consist of individuals or groups who themselves are engaged in praise or blame. The seemingly neutral act of choosing who to quote, and about what, as illustrated in Fig. 1, may be shaped by ideology and have significant effects on readers (Gentzkow and Shapiro, 2006; Gentzkow et al., 2015). Trump’s rhetoric, including calling Central Americans trying to enter the United States “an invasion,” and his hard-line immigration policies have exposed him to con- demnation since the El Paso shooting. “How far is it from Trump’s saying this ‘is an invasion’ to the shooter in El Paso declaring ‘his attack is a response to the Hispanic invasion of Texas?’ Not far at all,” Biden was due to say, according to an advance copy of his speech. “In both clear language and in code, this president has fanned the flames of white supremacy in this nation.” [0] Joe Biden NEG Donald Trump [1] Joe Biden NEG white supremacy [2] Donald Trump POS white supremacy Figure 1: Sample stance triplet annotations for a target sentence. Entities in SEESAW can be Person or Topic, and are annotated in canonical forms. Multiple stances are expressed, …

Content truncated to 500 words