Using graphs extracted from legal corpora to infer legal rules

Abstract

The study explores the use of graphs for linking various entities from legal corpora to identify outcomes for replying to search queries (user input). We explore various NLP techniques to automate the generation of such a graph by parsing legal corpora directly, including open information extraction. Additionally, the use of Label Propagation Algorithm (LPA) and Latent Dirichlet Allocation (LDA) are evaluated and tested for further improving the results on top of the graph data. From a practical perspective, the idea behind this project is to parse the full legal corpora of a jurisdiction (complete set of laws of a country), and automatically infer a graph which can behave similarly to a rule-engine, which, in turn, a person can subsequently easily ask for feedback. An example final solution for the end user would be the user asking a direct question "Can I produce wine?". The system would then be able to answer with information that is relevant to that question, as well as referencing the source texts and excerpts of that information for more completeness. Hence, the resulting experience to the end user can be compared to having a hybrid solution similar to a combination of a search engine and a knowledge base but focused specifically within the legal domain. This aims at providing richer value to the end user by offering more direct and specific answers rather than just references to text. We identified the main benefit of using Latent Dirichlet Allocation (LDA) to improve considerably the results when combined with the graph. Such a solution is aimed to the general public, as well as persons working in the legal domain alike. A primary object of this solution is to help users look up and find relevant answers, texts and statements within legal corpora with ease