No Permanent Friends or Enemies: Tracking Relationships between Nations from News

Proceedings of NAACL-HLT 2019, pages 1660–1676 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1660 No Permanent Friends or Enemies: Tracking Relationships between Nations from News Xiaochuang Han College of Computing Georgia Institute of Technology Atlanta, GA xc@gatech.edu Eunsol Choi Paul G. Allen School of CSE University of Washington Seattle, WA eunsol@cs.washington.edu Chenhao Tan Dept. of Computer Science Unversity of Colorado Boulder Boulder, CO chenhao@chenhaot.com Abstract Understanding the dynamics of international politics is important yet challenging for civil- ians. In this work, we explore unsupervised neural models to infer relations between na- tions from news articles. We extend exist- ing models by incorporating shallow linguis- tics information and propose a new automatic evaluation metric that aligns relationship dy- namics with manually annotated key events. As understanding international relations re- quires carefully analyzing complex relation- ships, we conduct in-person human evalua- tions with three groups of participants. Over- all, humans prefer the outputs of our model and give insightful feedback that suggests fu- ture directions for human-centered models. Furthermore, our model reveals interesting re- gional differences in news coverage. For in- stance, with respect to US-China relations, Singaporean media focus more on “strength- ening” and “purchasing”, while US media fo- cus more on “criticizing” and “denouncing”. 1 Introduction In the context of growing globalization (Baylis et al., 2017), understanding complex international relations is increasingly relevant to our daily life. Yet this is a challenging task due to the inher- ently dynamic nature of international relations. As Kissinger famously said, “America has no perma- nent friends or enemies, only interests.” Staying informed becomes even harder in the continuous streams of information from news outlets and so- cial media. This very availability of such information, how- ever, opens up exciting opportunities for natu- ral language processing to support individuals in understanding international relations. Supervised extraction has been incredibly useful at identi- fying pre-defined relations and events (Dodding- ton et al., 2004; Mintz et al., 2009) but fails to capture emerging or complex information needs. Topic models and neural models have been pro- posed to explore relations between entities with- out supervision (O’Connor et al., 2013; Chaney et al., 2016; Iyyer et al., 2016). In particular, Iyyer et al. (2016) introduces an unsupervised neu- ral model for tracking relations between fictional characters, and this approach outperforms base- lines from topic models and hidden Markov mod- els. In this work, we incorporate linguistic insights into this model to track relation dynamics between nations from news articles. Our model reconstructs textual information in the embedding space using relation embeddings, as proposed in Iyyer et al. (2016). We inte- grate simple yet effective linguistic insights: ver- bal predicates often describe the relationship be- tween entities,1 while nouns and proper nouns pro- vide the context of this relationship. For exam- ple, in “U.S. denounces Russia for its interference in the 2016 election”, denounce describes the re- lation, and election and interference provide the context. We show that this intuition leads the model to …

Content truncated to 500 words