Identification of Decision Rules from Legislative Documents Using Machine Learning and Natural Language Processing

Abstract

Decision logic extraction from natural language texts can be a tedious, labor-intensive task. This is especially true for legislative texts, since they do not always follow usual speech and writing patterns. This paper explores the possibility of using machine learning and natural language processing approaches to identify decision rules within legislative documents, and ultimately provides the possibility of building an extraction algorithm on top of the solution to extract and visualize decision logic automatically. Such a novel method for decision rules identification bears the potential to reduce human labor, minimize mistakes, and lessen context dependency. To accomplish this, we use pre-trained word vectorization in conjunction with a complex multi-layer convolutional neural network (CNN). The relevant data used in this project was generated from the Austrian income tax code and labeled by hand. A quantitative evaluation shows that our approach can be trained on as little as a single code of law and still obtain significant accuracy.