Toward a Framework for the Large Scale Textual and Contextual Analysis of Government Information Declassification Patterns

The US government protects a massive amount of secret data as part of its Security Classification System. This information is expensive to protect and maintain. In order to keep citizens informed, as well as to keep costs down, the government is constantly releasing newly declassified documents to the public. According to OpenTheGovernment.org’s annual Secrecy Report Card, human readers manually declassified almost 29 million pages of information in 2009 alone (McDermott and Bennett, 2010). Scholars interested in learning about government transparency history and policy face a daunting task in examining even a small portion of these documents. In order to make the process of learning about the content of these documents easier, we investigate the documents available through Gale’s Declassified Documents Reference System, an electronic repository of once-classified documents created throughout the 20th century, along two dimensions. First, we perform survival analysis to consider questions relating to time, such as when documents were created and how long they tend to remain classified. Then, we examine the contents of the documents via the use of statistical topic models. Finally, we combine temporal and content information, both by using the output of a topic model to inform a proportional hazards survival model and by considering time information during inference in a topic model. In this paper, we present a range of results that arise from a combined analysis of the temporal features and textual content of declassified documents. Since the results of statistical topic models tend to be easily-interpretable by humans, these new ways of looking at data relating to declassified government documents will likely be useful to experts in fields relating to government policy on secrecy.