Who Blames or Endorses Whom? Entity-to-Entity Directed Sentiment Extraction in News Text

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4091–4102 August 1–6, 2021. ©2021 Association for Computational Linguistics 4091 Who Blames or Endorses Whom? Entity-to-Entity Directed Sentiment Extraction in News Text Kunwoo Park Soongsil University∗ kunwoo.park@ssu.ac.kr Zhufeng Pan UCLA panzhufeng@cs.ucla.edu Jungseock Joo UCLA jjoo@comm.ucla.edu Abstract Understanding who blames or supports whom in news text is a critical research question in computational social science. Traditional methods and datasets for sentiment analysis are, however, not suitable for the domain of political text as they do not consider the direc- tion of sentiments expressed between entities. In this paper, we propose a novel NLP task of identifying directed sentiment relationship between political entities from a given news document, which we call directed sentiment extraction. From a million-scale news cor- pus, we construct a dataset of news sentences where sentiment relations of political entities are manually annotated. We present a simple but effective approach for utilizing a pretrained transformer, which infers the target class by predicting multiple question-answering tasks and combining the outcomes. We demonstrate the utility of our proposed method for social science research questions by analyzing pos- itive and negative opinions between political entities in two major events: 2016 U.S. pres- idential election and COVID-19. The newly proposed problem, data, and method will fa- cilitate future studies on interdisciplinary NLP methods and applications.1 1 Introduction Sentiment analysis is a useful technique for opinion mining from text data. Most existing work either focuses on sentence-level classification (Van Hee et al., 2018; Zampieri et al., 2019), or aims to detect the sentiment polarity towards specific targets (Pon- tiki et al., 2016; Cortis et al., 2017). These ap- proaches typically do not distinguish sources and targets of the sentiment. They mainly use user- generated content (UGC), such as tweets or restau- rant reviews from Yelp, which assumes each user ∗This work was done while the first author was a postdoc- toral researcher at UCLA. 1https://github.com/bywords/directed_ sentiment_analysis Figure 1: Overview of the directed sentiment extraction (account holder) is the source of the sentiment, and that the target is also clearly defined or easily iden- tifiable (e.g., the restaurant). This assumption does not hold for political news analysis where a large number of political actors blame or support each other in news articles. The key interest for political sentiment analysis is to identify “who” blame or support “whom” (Balahur and Steinberger, 2009) rather than simply assigning a global sentiment polarity to a given document or sentence. For example, from a sentence like “X supported Y for criticizing Z,” we can infer X is positive toward Y and both X and Y are negative toward Z. However, existing sentiment analysis methods are not suitable to detect such sentiment relationships between entities. This paper proposes a new NLP task of identi- fying directed sentiment relationships from a po- litical entity to another in news articles: directed sentiment extraction. For this task, we introduce a newly annotated corpus from a million-scale news 4092 dataset. As demonstrated in Figure 1, …

Content truncated to 500 words