Reference Resolution in Japanese Legal Texts at Passage Levels

Abstract

Sentences in the domain of legal texts are usually long and complicated. At the discourse level, they contains lots of reference phenomena which make the understanding of laws become more difficult. This paper investigates the task of reference resolution in the legal domain. The aim is to create a system which can automatically extracts referents for references in a real time. This is a new interesting task in the research of Legal Engineering. It does not only help readers in comprehending the law, support law makers in developing and amending laws, but also support in building an information system which works based on laws, etc. The main issues are to detect references and then resolve them to their referents. To detect references, we use a powerful machine learning technique rather than rule-based approaches as used in previous works. In resolving them, we design regular expressions to catch up the position of referents. We also build a corpus using Japanese National Pension Law to train and test our model. Our final system achieved 91.6% in the F1 score in detecting references, 96.18% accuracy in resolving them, and 88.5% in the F1 score in the end-to-end system.