Zimbabwe LexLM: Crafting and Validating a Legal NLP Model for Zimbabwe’s Bench and Bar

We introduce Zimbabwe LexLM, a natural language processing (NLP) model designed to assist the legal community in Zimbabwe. Our method involves further pretraining a multilingual transformer model on a specialized collection of local legal texts. This collection includes over 77,000 documents such as case laws, statutes, and legal commentaries from Zimbabwe and nearby jurisdictions. The LexLM architecture enhances the transformer with a citation graph module to understand references between documents and legal context. We tested LexLM on legal classification and retrieval tasks, achieving a macro-F1 score of about 0.865 and an accuracy of around 0.88. These results show that domain-specific training and citation-aware models can be effective in the legal field. Zimbabwe LexLM offers a strong, specialized model and dataset to improve legal research and information retrieval in resource-limited settings. Our work demonstrates how custom language models can assist legal practitioners in accessing information more efficiently, especially in low-resource contexts.