Multi-label text classification for risk prediction in contracts

Abstract

To determine valid criteria in detecting risks of contracts is essential for automation of legal tasks such as reviewing contracts. In this paper, we propose a multi-label text classification with a neural network model in order to predict multiple review points in each clause of contracts. On our dataset consisting of over 20k Japanese contracts, in which each clause has 1 ~ 4 label(s) and the labels total 205, our model achieved 31 ~ 64 % accuracy, depending on the number of labels an input text contains, for test data. In addition, we observed probability transition from the first character to the last character of the input texts, character by character, to check the relation between input token and output labels, and we found out that this observation helps us to see where on input texts our model attends to predict labels.