Abstract:
Access to legal knowledge in Bangladesh remains highly unequal, with women and children being among the most vulnerable groups who struggle to obtain affordable, understandable, and timely legal support. Complex legal language, financial barriers, and limited professional guidance further widen this gap. This thesis proposes NyaAI, an AI-powered legal assistance system designed to provide reliable and accessible legal information for women and children in Bangladesh. The system was implemented using both traditional retrieval methods (TF-IDF, BM25) and semantic transformer-based approaches (MiniLM embeddings) on a dataset of 3,131 legal queries and documents. After preprocessing and splitting into training, validation, and testing sets, the models were evaluated using Precision@5, Recall@5, F1@5, and Success Rate. Results showed that BM25 achieved the best overall performance with Precision@5 = 0.8133, Recall@5 = 0.9667, F1@5 = 0.8834, and Success Rate = 96.7%. TF-IDF followed closely with an F1@5 of 0.8674 and the same success rate of 96.7%. In contrast, the neural models performed significantly worse, with F1@5 scores of 0.6302 (MiniLM-L6) and 0.5853 (MiniLM-L12), and success rates dropping to 83.3% and 76.7% respectively. The research contributes both academically and practically. It demonstrates that traditional IR models can outperform transformer-based methods in domain-specific, resource-constrained environments. Furthermore, the selected BM25 model was integrated into a web-based application, enabling users to submit queries and receive relevant legal information quickly and efficiently. These findings highlight the feasibility of building a practical, AI-driven legal assistant that supports marginalized groups in understanding and exercising their legal rights in Bangladesh.
