Weakly-Supervised Relation Extraction in Legal Knowledge Bases

Abstract

Intelligent legal information systems are becoming popular recently. Relation extraction in massive legal text corpora such as precedence cases is essential for building knowledge bases behind these systems. Recently, most works have applied deep learning to identify relations between entities in text. However, they require a large amount of human labelling, which is labour intensive and expensive in the legal field. This paper proposes a novel method to effectively extract relations from legal precedence cases. In particular, relation feature embeddings are trained in an unsupervised way. With limited labelled data, the proposed method is shown to effective in constructing a legal knowledge base.