Learning to Extract International Relations from Political Context

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1094–1104, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Learning to Extract International Relations from Political Context Brendan O’Connor School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA brenocon@cs.cmu.edu Brandon M. Stewart Department of Government Harvard University Cambridge, MA 02139, USA bstewart@fas.harvard.edu Noah A. Smith School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA nasmith@cs.cmu.edu Abstract We describe a new probabilistic model for extracting events between major polit- ical actors from news corpora. Our un- supervised model brings together famil- iar components in natural language pro- cessing (like parsers and topic models) with contextual political information— temporal and dyad dependence—to in- fer latent event classes. We quantita- tively evaluate the model’s performance on political science benchmarks: recover- ing expert-assigned event class valences, and detecting real-world conflict. We also conduct a small case study based on our model’s inferences. A supplementary appendix, and replica- tion software/data are available online, at: http://brenocon.com/irevents 1 Introduction The digitization of large news corpora has pro- vided an unparalleled opportunity for the system- atic study of international relations. Since the mid- 1960s political scientists have used political events data, records of public micro-level interactions be- tween major political actors of the form “someone does something to someone else” as reported in the open press (Schrodt, 2012), to study the pat- terns of interactions between political actors and how they evolve over time. Scaling this data effort to modern corpora presents an information extrac- tion challenge: can a structured collection of ac- curate, politically relevant events between major political actors be extracted automatically and ef- ficiently? And can they be grouped into meaning- ful event types with a low-dimensional structure useful for further analysis? We present an unsupervised approach to event extraction, in which political structure and linguis- tic evidence are combined. A political context model of the relationship between a pair of polit- ical actors imposes a prior distribution over types of linguistic events. Our probabilistic model in- fers latent frames, each a distribution over textual expressions of a kind of event, as well as a repre- sentation of the relationship between each political actor pair at each point in time. We use syntactic preprocessing and a logistic normal topic model, including latent temporal smoothing on the politi- cal context prior. We apply the model in a series of compar- isons to benchmark datasets in political science. First, we compare the automatically learned verb classes to a pre-existing ontology and hand-crafted verb patterns from TABARI,1 an open-source and widely used rule-based event extraction system for this domain. Second, we demonstrate correlation to a database of real-world international conflict events, the Militarized Interstate Dispute (MID) dataset (Jones et al., 1996). Third, we qualitatively examine a prominent case not included in the MID dataset, Israeli-Palestinian relations, and compare the recovered trends to the historical record. We outline the data used for event discovery (§2), describe our model (§3), inference …

Content truncated to 500 words