Authors

Katie Zhang, Yihao Xu, Chengzhu Jia

Research Question

Can entailment detection, retrieval and value-frame classification identify policy conflicts that domain experts recognize as real? The project develops and pilots a computational approach that combines qualitative policy analysis with NLP to find where policies governing urban Nature-based Solutions (NbS) reinforce or contradict one another.

Background

Nature-based Solutions protect, restore and manage ecosystems to address societal challenges. They cut across housing, water, energy and land-use systems that are governed through separate mandates and competing priorities; housing policy, for example, may encourage development in areas environmental policy seeks to protect. These conflicts limit the effectiveness, equity and scalability of NbS.

Most computational policy analysis surfaces topical overlap or keyword co-occurrence, not the competing values and institutional priorities that determine whether policies work together. In New York City, where NbS commitments are spread across many agencies and plans, a working tool could help the Mayor's Office of Climate & Environmental Justice and partner agencies see where policies undermine one another before conflicts surface in implementation.

Methodology

Working from a corpus of NbS-relevant policy documents compiled with the faculty lead, the team will build and evaluate a pipeline that:

• Segments and structures heterogeneous policy text into comparable units.
• Applies textual entailment to classify statement pairs as supporting, contradicting or neutral.
• Uses embedding-based similarity to surface candidate conflict pairs from a large document space.
• Classifies values of nature (instrumental, intrinsic, relational) and tests whether value framing predicts where conflicts cluster.
• Serves results through a Retrieval-Augmented Generation (RAG) interface that returns source passages alongside flagged tensions.

The team will measure precision and recall against expert interpretation, run error analysis, and document where general-domain models transfer poorly to regulatory prose.

Deliverables

• Machine learning model
• Academic paper
• Literature review
• RAG interface

Data Sources

• Corpus of NbS-relevant policy documents compiled and interpretively framed with the faculty lead