Three Essays on Natural Language Processing and Information Extraction with Applications to Political Violence and International Security

Abstract

This dissertation consists of three essays that develop new word order aware text analysis methods to extract information from text and apply these techniques to questions in security studies. Most existing text analysis tools in political science classify or cluster documents, discarding the order of words. In contrast, the techniques I introduce take word order into account in order to extract pieces of information within documents. Each paper introduces new text analysis techniques and applies them to measuring key variables in conflict and security studies. One paper introduces a new technique for extracting descriptions of political events, using both the syntax, or grammar, of a sentence and the semantic meaning of words. I apply the new event extraction technique to two domains, extracting descriptions of human rights abuses from State Department reports and descriptions of political violence from Indian newspapers, finding that the standards of the human rights reports have changed over time, and that existing techniques greatly undercount both political violence and police responses to it in India. Another dissertation paper uses new text analysis techniques, including for Arabic text, to create a new dataset of over 100,000 civilian casualties in the Syrian civil war to test theories of violence against civilians. I find that theories emphasizing strategic violence have much greater explanatory power than theories emphasizing territorial control or threats against the regime. A third dissertation paper extends this geolocation work, providing a neural network-based method for linking the events described in text with the locations where they occurred, even when multiple events and locations occur in the same sentence.