Abstract

Natural language processing in the legal domain is in its infancy. Still, it is a much-needed solution to repetitive and time-consuming tasks such as contract review. There are several ways in which a contract can be reviewed using natural language processing including the classification of norm sentences into the categories of permission, obligation and/or prohibition. Such a process requires a thorough annotation scheme comprised of clear guidelines and adequate resources (data, time, expertise) followed by an appropriate word embedding method.

This project outlines the methodology for an annotation scheme on a small dataset, highlights the significance of domain-specific word embedding, and further demonstrates the efficiency of convolutional neural network classifiers on multilabel classification tasks. The best result of 0.02 loss and 98% precision was achieved, a remarkable performance for a developing field.

Natural Language Processing for Legal Document Review: Categorising Deontic Modalities in Contracts