Design Process
The project leverages data from Congress.gov API, primarily utilizing the /bills and /members endpoints to gather information on legislative measures and lawmakers. By combining these datasets, we conducted text analysis using large language models (LLMs) to assess the environmental impact of various bills.
Text Analysis Using LLMs
We tested the efficiency of multiple LLMs, manually evaluating over 100 bills—both positive and negative for the environment—against models fine-tuned locally and OpenAI’s hosted service. Our analysis found that OpenAI's GPT-4o provided the most accurate assessments from an environmentally conscious perspective, independent of cost considerations. To refine results, we adjusted parameters such as tokens, temperature, and Top-P. GPT-4o was tasked with assigning a score from 0 to 10, where 10 represents strong environmental benefits and 0 indicates harm. Using these scores, we developed an Eco-Policy Score for each legislator, calculated as:
Eco-Policy Score = (sum of scores for sponsored bills) + 0.1 (sum of scores for co- sponsored bills).
This metric provides a quantified perspective on lawmakers’ environmental policy stances.
