Automatic legal case summarization crucially condenses a lengthy document into a short structure, while conserving its information content and overall meaning. Manual summarization necessitates a substantial amount of human labour and time. For this reason, automatic text summarization is introduced which saves the legal expert time. One of the key difficulties in legal text summarizations is the study of high-dimensional input results. This issue can be solved by using feature selection. Various heuristic algorithms like particle swarm optimization, genetic algorithm, and ant colony search algorithm provide better accuracy in particular problems, but they cannot be used as universal. Gravitational Search Algorithm is one of the modern optimizations based on heuristics algorithms according to Newton's law. But, the Gravitational Search Algorithm is incapable of recalling information that implies a reduction in memory. Researchers experimented using the Convolutional Neural Network to perform the sentiment analysis of legal text summaries obtained from our summarized legal dataset taken from Brazilian Supreme Court data. In the course of training the model, the dataset is manually annotated by experts. The experiments are performed with various numbers and sizes of filters in different ten convolution neural network models. © 2025 Elsevier B.V., All rights reserved.