Machine learning dataset for criminal law suggestions using case studies within the context of Bangladesh

Abstract

With the growth of Language Modeling and the upcoming natural language processing assisted tools which aim for text generation, can someday render bureaucracy mean ingless while also decreasing human workload. To bring about that day, we need datasets which are able to train those models. Especially in the case of Bangladesh where there are very few datasets based on Bangladesh’s legal case studies. In this paper, we have created a Multi-Label and Binary classification dataset through data augmentation using verified case studies from the Manupatra database with the criminal law subject within the context of Bangladesh and have tested them with a few models such as DistilBERT, BERT, GPT-2, GPT-3, XLNet and Mamba to classify the acts involved and court in a case study for a 2000 data split and 3000 data split. So far, we have collected 3000 criminal case studies for our augmented dataset. Experimental results showed that out of all the models, Mamba performed the best while GPT-2 came up with the worst results. DistilBERT showed almost similar results to BERT and XLNet despite their computational differences during the benchmarking process of our augmented dataset.