Utilizing VGG19 and ResNet50 Neural Network Models for Comprehensive Intelligent Learning and Analysis of Legal Precedents and Texts

Analysis of legal text presents its unique challenges compared to textual data due to diverse and largely unstructured content and the sophistication of legal language. In the present study, we investigate the suitability of deep learning for this challenging domain by predicting legal precedent based on helical text and image data with the help of several state-of-the-art deep learning models such as Convolutional Neural Network (CNN), VGG16, VGG19, and ResNet50. The dataset consists of a wide range of cases across various domains: civil, criminal, administrative, constitutional, and appellate domain for a total of 350 cases. The dataset is divided into 70% for training and 30% for testing. Prior to training the models, data preprocessing includes textual cleaning, tokenization, and then converting text to a numeric format using the Bag-of-Words model. Image data is resized, normalized, and augmented to increase size by 10%. The performance of the models is assessed by the standard evaluation metric including accuracy, precision, recall, and f1 score. Results indicate that the CNN model is the best performer with 96.78% accuracy followed by VGG16 with 94.5% accuracy, VGG19 with 93.2% accuracy, and ResNet50 with 90.25% accuracy. In the discussion section, we provide insights into the performance on the strengths and weaknesses of CNN along with other models. Their analysis clearly shows CNN as the most suitable model for small samples and outperforming the other models for precision, recall and F1 score. Potential use cases of this research can involve legal practice as it could greatly help lawyers carrying out precedent analysis.