The availability and generation of digitized newspaper collections have provided researchers in several domains with a powerful tool to advance their research. More specifically, digitized historical newspapers give us a magnifying glass into the past. In this paper, we propose a scalable and customizable big data analysis system that enables researchers to study complex questions about our society as depicted in news media for the past few centuries by applying cutting-edge text analysis tools to large historical newspaper collections. We discuss our experience with building a preliminary version of such a system, including how we have addressed the following challenges: processing millions of digitized newspaper pages from various publications worldwide, which amount to hundreds of terabytes of data; applying article segmentation and Optical Character Recognition (OCR) to historical newspapers, which vary between and within publications over time; retrieving relevant information to answer research questions from such data collections by applying human-in-the-loop machine learning; and enabling users to analyze topic evolution and semantic dynamics with multiple compatible analysis operators. We also present some preliminary results of using the proposed system to study the social construction of juvenile delinquency in the United States and discuss important remaining challenges to be tackled in the future.