RePAIR: Recommend political actors in real-time from news websites

Extracting a structured representation of political events from news reports is at the intersection of the computational and social sciences. A traditional approach is to use dictionary-based pattern lookups to identify actors and actions involved in potential events. A key complication of this approach is updating the dictionaries with new actors (e.g., when a new president takes office). Currently, the dictionaries are curated by humans, updated infrequently, and at a high cost. This means that tools dependent on the actor dictionaries (e.g., PETRARCH) overlook events when actors are missing in the dictionary. Since these tools use only the syntactic structure of the sentence (e.g., parse tree, etc.) for their event coding, missing actors will generate events which fail to capture actual political interaction. To overcome these issues, we propose a framework RePAIR to recommend new political actors in real-time from the political news articles with RSS feeds related to national/international politics across the globe. The framework identifies semantic structure of a sentence using an Automatic Content Extraction (ACE) method and uses a frequency based actor ranking algorithm to recommend the most frequent new political actors over multiple time windows. We also suggest the associated role of recommended new actors from the role of co-occurred political actors in the existing CAMEO actor dictionary. Further we integrate an external knowledge base (e.g., Wikipedia) into our framework to capture the evolving roles of existing actors over time and recommend new roles for them. Furthermore, we consider PETRARCH and BBN ACCENT event coders for actor recommendation, and a graph-based actor role recommendation using weighted label propagation as baselines and compare them with our framework. Experimental results show our approaches outperform them significantly.