Exploring Bill Similarity with Attention Mechanism for Enhanced Legislative Prediction

Common legislative prediction methods often emphasize bill content or social relationships. This paper, motivated by the insight that similar policy texts reflect comparable political ideologies and can lead to similar voting outcomes, proposes a deep learning method that exploits attention mechanisms to incorporate semantic similarity between bills into legislative prediction models. Our approach uses attention scores to identify bills that are most similar to the one being predicted, and combines the encoded features of these similar bills as additional auxiliary information. By integrating these related features, the model goes beyond the semantic information of individual bills, leading to a more comprehensive use of roll-call data. Empirical results show that utilizing bill similarity along with traditional social relationships, voter characteristics, and bill content significantly improves performance in terms of accuracy, recall, precision, and F1 score compared to models that ignore bill similarity. The results also confirm that legislators tend to maintain consistent views or voting patterns on bills that are similar in nature. In addition, we demonstrate that the attention mechanism is more effective than conventional similarity measures, such as cosine similarity and Euclidean distance, in capturing the similarities between bills.