Multi-label legal text classification with BiLSTM and attention

Abstract

Like many other knowledge fields, the legal area has experienced an information-overloaded scenario. However, to extract data from legal documents is a challenge due to the complexity of legal concepts and terms. This work aims to address Bidirectional Long Short-Term Memory (BiLSTM) to perform Portuguese legal text classification to solve such challenges. The proposed model is a shallow network with one BiLSTM layer and one Attention layer trained over two small data sets extracted from two Brazilian courts: the Superior Labour Court (TST) and 1st Region Labour Court. The experimental results show that combining the BiLSTM layer and the Attention layer for long judicial texts helps capture the past and future contexts and extract multiple tags. As the main contribution of this research, the proposed model can quickly process multi-label and multi-class data sets and adapt to new contexts in different languages.