Predicting and understanding law-making with word vectors and an ensemble model

Area of Science:

• Computational Social Science
• Natural Language Processing
• Political Science

Background:

• Few US bills become law, necessitating predictive models.
• Understanding legislative success factors is crucial for policy analysis.

Purpose of the Study:

• Develop a machine learning model to forecast the probability of US bills becoming law.
• Investigate the predictive power of legislative text versus contextual factors.
• Assess the impact of bill amendments on prediction accuracy.

Main Methods:

• Trained machine learning models on legislative data from the 107th to 113th US Congresses.
• Utilized a language model to embed legislative text into semantic vectors.
• Compared text-only, context-only, and combined models for prediction.
• Analyzed bill text and metadata at introduction versus latest available data.

Main Results:

• Context-only models outperformed text-only models at bill introduction.
• Text-only models surpassed context-only models with the latest bill data.
• Combined text and context models consistently yielded the best prediction performance.
• Identified key variables influencing bill enactment through sensitivity analysis.

Conclusions:

• Machine learning effectively forecasts legislative success.
• Integrating textual content and contextual information optimizes prediction.
• The predictive power of text increases as bills evolve through the legislative process.