Scaling requirements extraction to the crowd: Experiments with privacy policies

Abstract

Natural language text sources have increasingly been used to develop new methods and tools for extracting and analyzing requirements. To validate these new approaches, researchers rely on a small number of trained experts to perform a labor-intensive manual analysis of the text. The time and resources needed to conduct manual extraction, however, has limited the size of case studies and thus the generalizability of results. To begin to address this issue, we conducted three experiments to evaluate crowdsourcing a manual requirements extraction task to a larger number of untrained workers. In these experiments, we carefully balance worker payment and overall cost, as well as worker training and data quality to study the feasibility of distributing requirements extraction to the crowd. The task consists of extracting descriptions of data collection, sharing and usage requirements from privacy policies. We present results from two pilot studies and a third experiment to justify applying a task decomposition approach to requirements extraction. Our contributions include the task decomposition workflow and three metrics for measuring worker performance. The final evaluation shows a 60% reduction in the cost of manual extraction with a 16% increase in extraction coverage.