A Hybrid Stacking Model for Legal Document Classification Using TF-IDF Features

In Legal Natural Language Processing (Legal NLP), legal document classification is a fundamental task that can help legal professionals to automatically analyze case documents and determine the treatment of previous judicial decisions in new cases. Traditional machine learning models are good at efficiency but don't have a semantic understanding while deep learning models are more flexible in extracting contextual information, but are prone to over-fitting and have high computational costs. In this paper, a hybrid stacking model is proposed, which integrates the classical machine learning, ensemble learning, and neural learning, so that it can achieve a good balance between accuracy, robustness and efficiency. The proposed architecture is divided into two texts and then uses the TF-IDF feature extraction for the legal document. The extracted features are fed to the base classifiers as the linear support vector machine, extra trees classifier and multi-layer perceptron and the outputs of the base classifiers are passed to the logistic regression meta-classifier for stacking. The model was tested on a synthetic legal corpus that was created using a set of 660 documents, split 80:20 train:test and five-fold crossvalidation. The hybrid model performed best with a test accuracy of 94.7%, a cross-validation accuracy of 92.1% and a macro F1-score of 94.7%, which is similar to the best individual classifiers, and has better robustness due to stacking.