Toward a National Eviction Data Collection Strategy Using Natural Language Processing

Tim Thomas
Alex Ramiller
University of California, Berkeley

Cheng Ren
University at Albany, State University of New York

Ott Toomet
University of Washington

During the past decade, eviction research has relied heavily on preprepared (structured) data from third parties and state agencies who have taken the effort to create readable and accessible filing data. However, massive data gaps across the country exist because third parties may not provide a complete count of filings and many states do not have a formalized process to digitize, enumerate, analyze, or release information on evictions. In some states, the bulk of eviction filings are buried in court filings.

To address this issue, the Eviction Research Network developed a natural language processing (NLP) approach to mine court record images to enumerate and map eviction filing counts at the neighborhood level and help researchers identify disparities by location, race, and gender. This approach involved downloading eviction court record images from online county court systems, digitizing the text, isolating and geocoding addresses, and estimating demographics based on names and location.

In a case study for the State of Washington, millions of pages in more than 110,000 eviction filings from 2004 to 2017 were processed to demonstrate this approach. The research shows massive racial and gender disparities, where up to one in five African-American/Black female-headed households were named in eviction filings. Eviction rates peak in areas with the lowest rent and in the most diverse neighborhoods when analyzing neighborhood dynamics related to eviction. This research helped pass several tenant protection policies in the state and informed other strategies on how to address housing precarity. A suggested strategy for collecting eviction data across the country concludes the article.