Challenge
A public institution was inundated with thousands of legislative amendment proposals to the national budget each year. Manually reviewing and analyzing these proposals was time-consuming and resource-intensive. The team faced several challenges:
- Volume of Proposals: With over 2,000 proposals annually, it was nearly impossible to thoroughly evaluate each one in a timely manner.
- Redundancy Detection: Identifying similar or duplicate proposals submitted over the years was difficult, leading to repetitive evaluations.
- Categorization: Organizing proposals into relevant categories for better assessment and delegation was a manual process prone to errors.
- Resource Allocation: Limited staff resources made it challenging to focus on proposals with the most significant impact.
The client needed an efficient solution to automate the analysis process, reduce redundancy, and focus on high-priority proposals.
APPROACH
An automated system leveraging advanced natural language processing (NLP) and machine learning techniques was developed. The approach included:
- Data Collection and Parsing:
- Collected legislative amendment proposals in PDF format from the past three years.
- Developed a parser to extract text from PDFs, handling various formatting issues.
- Text Summarization:
- Utilized OpenAI’s GPT models to generate concise summaries of each proposal.
- Ensured summaries focused on the key legislative changes without unnecessary details.
- Proposal Categorization:
- Defined a set of relevant categories (e.g., Education, Health, Infrastructure).
- Implemented a categorization model to assign each proposal to the most appropriate category or flag it for review if no suitable category existed.
- Embedding and Similarity Analysis:
- Generated embeddings for proposal summaries to capture semantic meaning.
- Calculated cosine similarity between embeddings to identify similar or duplicate proposals.
- Established thresholds to determine equivalence and relevance.
- Matching New Proposals with Past Submissions:
- Compared new proposals against the historical database to detect redundancies.
- Provided recommendations on whether to review, accept, or disregard proposals based on similarity scores.
- Automation and Batch Processing:
- Leveraged batch processing capabilities to handle large volumes of data efficiently.
- Implemented error handling and retries to ensure robustness.
Solution
The solution delivered was an integrated system that automated the analysis of legislative amendment proposals:
- Automated Parsing: Extracted text from PDFs and handled exceptions gracefully.
- AI-Powered Summarization: Generated clear and concise summaries for quick understanding.
- Efficient Categorization: Assigned proposals to predefined categories, aiding in organized reviews.
- Similarity Matching: Identified top matches among past proposals to reduce redundancy.
- User-Friendly Outputs: Produced reports with hyperlinks for easy access to proposal details.
- Scalable Architecture: Designed to handle increasing volumes of data without compromising performance.
Result
The implementation of the automated analysis system led to significant improvements:
- Time Savings: Reduced the manual review time by over 70%, allowing staff to focus on high-impact proposals or proposals that were completely new.
- Enhanced Consistency: Provided standardized summaries and categorizations, improving decision-making consistency.
- Redundancy Reduction: Identified and eliminated duplicate proposals, streamlining the legislative process.
- Resource Optimization: Allocated staff resources more effectively by highlighting priority proposals.
- Data-Driven Insights: Enabled data analytics on proposal trends over the years, informing future policy decisions.
