SBIR Phase I: Machine Assisted Comparative Policy Analysis in Public Health

点击翻译按钮获取中文摘要

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project will be to enhance the capabilities of a research tool for companies, legal experts and researchers undertaking nationwide comparative policy analysis in the public health domain. The tool will assist experts in identifying relevant policy documents by determining and scoring the significance of statutory provisions in context of specific legal questions. The quantitative approach can enable novel policy tracking. Rather than experts setting up alerts for updates to a specific set of documents, this tool learns from the legal text used to answer legal questions to allow for real time tracking and discovery of updates and other relevant documents. This approach to policy tracking can present experts with timely information on updates, along with revealing new documents as they are introduced. Timely analysis can inform policy-makers, facilitating the crafting of optimized evidence-based public health legislation. Reducing the cost and effort of the most time-consuming aspects of legal research can make precise scientific policy analysis affordable and accessible commercially and in real time. This Small Business Innovation Research (SBIR) Phase I project will decrease the time required to produce timely analysis of public health policy across 50 states. This research will apply machine learning, natural language processing and graph theory techniques to extract logical legal ontologies by computing similarities of public health provisions in statutory text. In domain specific problems, large sets of examples of annotated text are required. In the legal domain there is little available expert-labeled legal corpora and purposefully curating this kind of dataset is prohibitively expensive. To address this challenge, the proposed solution integrates transparently into legal experts' workflow while generating ontology that mirrors the approach of a domain expert. The second challenge is that searching for patterns in the relations of a very large network of documents can be very expensive computationally. The proposed solution addresses this by extracting clues from the expert workflow to identify shortcuts that simplify and constrain the larger problem. These clues, combined with sparse expert labeled data can produce a more accurate baseline for optimization of scoring and similarity comparison of larger sets. By being integrated into more workflows, the transparent annotation process and algorithm could be applied to other policy domains.