A Hybrid Text Analysis Method for Contextual Specificity in Political Documents

• In Event: Advances in Machine Learning (II)

Abstract

One of the most problematic phenomena in text analyses of corpora of political documents such as speeches or position papers is that of contextual specificity. For example, both the referent and the valence of terms such as “tightening” or “hardliner” differ not only by domain, say monetary vs. foreign policy, but also by specific issue, speaker, and time period. Even in relatively large corpora, this variation complicates the use of methods such as topic modeling, much less sentiment analysis. To address the issue of contextual specificity, we present a hybrid machine learning method based on training a semantic parser on human-annotated text. Applied to two extensive corpora of policy announcements – one by the Federal Reserve Board on monetary policy, the other by the U.S. State Department and the White House on relations with Russia – made by dozens of speakers/authors, over decades, and on a variety of particular issues, the hybrid method is able to “predict” contextually specific details of newspaper articles about those announcements, such as the nature of the action being announced, the motives seen as underlying it, and how the action itself is viewed.

Authors

• David Sylvan, Graduate Institute of International and Development Studies

• Jean-Louis Arcand, Graduate Institute of International and Development Studies

• Ashley Thornton, Graduate Institute of International and Development Studies