Language technologies for understanding law, politics, and public policy

[摘要] This thesis focuses on the development of machine learning and natural language processing methods and their application to large, text-based open government datasets. We focus on models that uncover patterns and insights by inferring the origins of legal and political texts, with a particular emphasis on identifying text reuse and text similarity in these document collections. First, we present an authorship attribution model on unsigned U.S. Supreme Court opinions, offering insights into the authorship of important cases and the dynamics of Supreme Court decision-making. Second, we apply software engineering metrics to analyze the complexity of the United States Code of Laws, thereby illustrating the structure and evolution of the U.S. Code over the past century. Third, we trace policy trajectories of legislative bills in the United States Congress, enabling us to visualize the contents of four key bills during the Financial Crisis. These applications on diverse open government datasets reveal that text reuse occurs widely in legal and political texts: similar ideas often repeat in the same corpus, different historical versions of documents are usually quite similar, or legitimate reasons for copying or borrowing text may exist. Motivated by this observation, we present a novel statistical text model, Probabilistic Text Reuse (PTR), for finding repeated passages of text in large document collections. We illustrate the utility of PTR by finding template ideas, less-common voices, and insights into document structure in a large collection of public comments on regulations proposed by the U.S. Federal Communications Commission (FCC) on net neutrality. These techniques aim to help citizens better understand political processes and help governments better understand political speech.

[发布机构] Massachusetts Institute of Technology