RnnTd: An Approach Based on LSTM and Tensor Decomposition for Classification of Crimes in Legal Cases

Abstract

With the rapid development of big data and artificial intelligence technology, wisdom justice and wisdom procuratorial services have become an inevitable trend in the development of contemporary courts and procuratorates. As one of the most basic businesses involved, the classification of crimes in legal cases is one of the hot topics of research. This article proposes a classification model of legal cases based on LSTM and tensor decomposition layer, namely RnnTd. We represent different types of legal cases in terms of tensors, and then apply tensor decomposition layer to decompose the original tensors into core tensors. The core tensor represents the primary tensor elements and tensor structure information of its corresponding original tensor. Further, the core tensors are used to train LSTM to construct a legal case classification model. Compared with the classification models which are based on traditional deep learning algorithms, the tensor decomposition layer based LSTM classification model proposed in this article has weak dependence on vocabulary and grammar information in the original legal case data of legal cases, and does not require heavy manual marking work. It is worth mentioning that our model is more scalable and interpretability. Experiments show that the legal case classification algorithm proposed in this article has higher accuracy and faster convergence than traditional neural networks.