Double Granularity Graph Network For Chinese Legal Question Answering

Legal question answering is a critical task in artificial intelligence. Since most legal data are presented in text, using natural language processing (NLP) to solve legal question answering is a current research direction. Compared with traditional question answering tasks, legal question answering often contains some potential information, such as legal events, crime process, litigants, and victims. This potential information suggests the legal question answering model reasoning's theme and can help the model improve its reasoning ability. In addition, the legal question answering task must answer based on relevant legal clauses, and the number of relevant legal clauses is usually a lot. Hence, the model needs to eliminate the influence of redundant and noisy clauses. Therefore, we propose a double-granularity-based graph neural network that can reason through potential legal events. Based on this research, we design an attention mechanism based on text interaction and calculate the attention by different window sizes score to decrease the influence of noise graph nodes. Finally, we evaluate the proposed model on the JEC-QA benchmark dataset to demonstrate our method's effectiveness. Experimental results show that the model performs well on the Chinese legal examination data and outperforms classical baselines. Out-performs the best baseline model by 7.08 in overall performance and outperforms the best baseline model in single-choice and multiple-choice questions; they are 7.52 and 6.41, respectively.