Name in original language
Ulysses - Consolidação de sumaríos de proposições legislativas
Initiative overview
Legislative consultants, staff in charge of support the laymaking drafting, assist parliamentarians in reporting on legislative bills by producing consolidated executive summaries for each bill, including its justification and other related documents: amendments, attached bills, and approved substitutes.
The production of these summaries can take hours or days, depending on the number of amendments and bills attached to the main bill being reported. This challenge is amplified along with the bills' lifecycle, once the set of documents increases eveyday and the updated version can assume a different meaning along with the time. In some cases, the bill intends to change other bills, resulting in a complex chain of documents.
The AI system uses generative AI tools to assist consultants in producing these summaries, which must be reviewed and supplemented by the consultants themselves according to thematic or contextual needs. These are the main objectives for this AI application: (i) decrease the time required to produce consolidated executive summaries from hours or days to minutes, enabling legislative consultants to work more efficiently; (ii) automatically identify and synthesize key information from all related documents to prevent oversight of critical content; and (iii) provide standardized, well-structured summaries that serve as a robust foundation for consultants to review and refine, ensuring uniformity across legislative reports. This initiative necessitates ongoing investment in ensuring the quality of both metadata and complete textual content of the legislative proposals under analysis and their respective cited legislation.
Results, outcomes and impacts:
A first positive impact achieved was the standardization of AI-generated summaries based on consensus obtained among the consultants who participated in the solution validation. For example, there are rules for initiating summary text, for limiting the number of words generated, for incorporating cited legislation into the summarization context, etc. We have begun collecting feedback-based measurements for ongoing investigation and continuous improvement of the deployed solution
Challenges and lessons learned:
Along the implementation of the AI system, it was necessary to perform extraction of textual and tabular content from documents produced in various different formats (e.g., .doc, .docx, .pdf, .tiff, among others). Some bills and regulations may contain both text and tables simultaneously. Moreover, varying context window constraints and behavioral differences were observed across different generative AI models, which required specific testing and targeted interventions to handle each encountered issue.
