Identifying Legal Core Elements Based on Extra-Long Network and Bidirectional Long Short-Term Memory

The accurate identification of core elements in legal documents plays a crucial role in the outcome of a case and therefore receives widespread attention. However, current methods often fail to simultaneously extract the long-distance dependency and temporal relationship of the case description information. Therefore, this paper proposes a legal core element recognition method based on Extra-Long Network (XLNet) and Bidirectional Long Short-Term Memory (BiLSTM), referred to as XLNet-BiLSTM. Specifically, this method utilizes the XLNet network of Transformer to process long-distance dependency relationships, and applies BiLSTM to capture more detailed temporal sequence information, thus effectively identifying the contextual information and element labels of case descriptions. The proposed method is trained and predicted by case classification in the CAIL2019 element recognition task data-set. In terms of the comprehensive performance indicator F1, divorce, labor, and loan achieve 78.63%, 76.93%, and 77.20% respectively, all being the best, and 1.5, 9.4, and 3.5 percentage points higher than the second place respectively. The experimental results indicate that the XLNet-BiLSTM model performs well in the multi-label binary classification task of legal core elements.