Query-focused Scenario Construction

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2712–2722, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 2712 Query-Focused Scenario Construction Su Wang1,2 Greg Durrett3 Katrin Erk1 1Department of Linguistics 2Department of Statistics and Data Science 3Department of Computer Science The University of Texas at Austin shrekwang@utexas.edu gdurrett@cs.utexas.edu katrin.erk@mail.utexas.edu Abstract The news coverage of events often contains not one but multiple incompatible accounts of what happened. We develop a query-based system that extracts compatible sets of events (scenarios) from such data, formulated as one- class clustering. Our system incrementally evaluates each event’s compatibility with al- ready selected events, taking order into ac- count. We use synthetic data consisting of ar- ticle mixtures for scalable training and evalu- ate our model on a new human-curated dataset of scenarios about real-world news topics. Stronger neural network models and harder synthetic training settings are both important to achieve high performance, and our final sce- nario construction system substantially outper- forms baselines based on prior work. 1 Introduction While a situation is developing, news reports often contain multiple contradictory stories (scenarios) of what happened, and it is hard to piece together the individual scenarios. For example, surround- ing the disappearance of the Saudi journalist Jamal Khashoggi, there were initially multiple conflict- ing accounts of what happened. One states that he was the victim of a murder scheme; an alternative suggests that he walked out of the consulate alive. The task of identifying these individual scenarios is also being considered in the Active Interpreta- tion of Disparate Alternatives (AIDA) program,1 and in a recent Text Analysis Conference (TAC).2 We frame the task as query-based scenario dis- covery: given a topic (e.g., the disappearance of Jamal Khashoggi) and a query (e.g. Jamal Khashoggi was murdered), we want to retrieve 1https://www.darpa.mil/program/active- interpretation-of-disparate-alternatives 2https://tac.nist.gov/2018/SM-KBP/ index.html Query: Jamal Khashoggi was murdered. Jamal Khashoggi entered the consulate of Saudi Arabia consulate in Istambul. He exited the Saudi consulate after a few minutes. The team wanted to arrest Khashoggi but botched it. He never exited the Saudi consulate but died there. Khashoggi, according to the reporter, was seen on a flight leaving Turkey for Estonia. A team flew from Saudi Arabia to Turkey prior to Khashoggi’s appointment at the consulate specifically to intercept him. The team was sent by the Saudi crown prince with the order to murder Khashoggi. Jamal A. Khashoggi works for The Washington Post, and is the editor-in-chief of Al-Arab News. Figure 1: An example for query-based scenario con- struction. Given the query, we want to select event- denoting sentences from a document mixture to build a target scenario with a sequence of compatible events. The mixture also contains sentences which may be ir- relevant or part of an alternative scenario. a scenario, a set of compatible events, from the given reports. We formulate query-based scenario discovery as one-class clustering (Bekkerman and Crammer, 2008). We specifically focus on dis- covering …

Content truncated to 500 words