Experimental Evaluation of CNN Parameters for Text Categorization in Legal Document Review

Predictive Coding, also called Text Categorization, has been widely used in legal industry. By leveraging machine learning models such as logistic regression and SVM, the review of documents can be prioritized based on their probability of relevance to the legal case, thus improving review efficiency and cutting cost. In recent years, deep learning models-combined with word embeddings-have shown better performance in predictive coding. However, deep learning models involve many parameters and it is challenging and time-consuming for legal practitioners to select appropriate settings. Based on the experiments on several public legal text datasets, this paper shows the preliminary results about how various key parameter settings impact the performance of Convolutional Neural Networks (CNNs).