Extraction of important features for risk prediction in contracts

Abstract

Contract review requires legal knowledge, which makes it difficult for non-experts while easy for experts in the legal department. To overcome such legal disparity, we need to automate the review process, especially risk decision in contracts. In this paper, we formulate risk decision in contracts as binary text classification, train classifiers using machine learning models and evaluate them. To identify a text span to be revised in a contract, we apply LIME, a method for estimating important features for prediction, to BERT classifier and extract important tokens from text. It is observed that the extracted tokens match most of the gold ones annotated by experts. Furthermore, we present revision examples that result in the inverted risk prediction and contribute to the prediction. We show that LIME can help to identify a text span to be revised in review work and present correction examples with high transparency.