Augmenting Human-Centered Racial Covenant Detection and Georeferencing with Plug-and-Play NLP Pipelines

Abstract

Though no longer legally enforceable, racial covenants in twentieth-century property deeds continue to shape spatial and socioeconomic inequalities. Understanding this legacy requires identifying racially restrictive language and geolocating affected properties. The Mapping Prejudice project addresses this by engaging volunteers on the Zooniverse crowdsourcing platform to transcribe covenants from scanned deeds and link them to modern parcel maps using transcribed legal descriptions. While the project has explored automation, it values crowdsourcing for its social impact and technical advantages. Historically, Mapping Prejudice relied on lexicon-based searching and, more recently, fuzzy matching to flag suspected covenants. However, fuzzy matching has increased false positives, burdening volunteers and raising scalability concerns. Additionally, while many properties can be mapped automatically, others still require time-intensive manual geolocation. We present a human-centered computing approach with two plug-and-play NLP pipelines: (1) a context-aware text labeling model that flags racially restrictive language with high precision and (2) a georeferencing module that extracts geographic descriptions from deeds and resolves them to real-world locations. Evaluated on historical deed documents from six counties in Minnesota and Wisconsin, our system reduces false positives in racial term detection by 25.96% while maintaining 91.73% recall and achieves 85.58% georeferencing accuracy within 1x1 square-mile ranges. These tools enhance document filtering and enrich spatial annotations, accelerating volunteer participation and reducing manual cleanup while strengthening public engagement.