Design And Implement Of Interpretable Legal Judgment Prediction Model Based On ILJP

Based on the current situation of poor interpretability of legal intelligence system,limit research on element identification,unsatisfactory prediction effect of few-shot and confusing legal causes,an Interpretable Legal Judgment Prediction Model ILJP was proposed,which combines the factual elements identification task with the cause prediction task in the same framework,and improves the prediction performance of few-shot and confusing causes through the factual elements and the introduced the external hierarchical dependence between legal causes.Moreover,the identified factual elements and the predicted cause path can provide some interpretable support for the model.The main work content of this thesis is as follows:（1）Recognition of factual elements based on LEP.In view of the current situation of limit research on factual element identification and the poor interpretability of the model,this thesis designs a LEP model to identify the phrases that play an important role in the judgment structure.The LEP model encodes the fact description based on XLNet-RCNN-S network,and obtains the semantic representation of specific elements based on FE-MHA mechanism.The experimental results show that compared with XLNet model,the Macro-F1 and micro-F1 values of LEP model have increased by 2.66% and 4.04% respectively in CAIL dataset,and the highest increase of 4.05% and 5.42% in CIVIL dataset,indicating that the model can capture rich label features and is effective for the identification of factual elements.（2）Legal Cause prediction based on HLCP.Aiming at the current situation of unsatisfactory prediction effect of few-shot and confusing legal causes,this thesis designs a HLCP model for cause prediction.The model introduces the hierarchical structure of cause,transforms cause prediction into sequence generation task,and predicts the cause path through the improved Seq2Seq-attention model.The experimental results show that compared with XLNet model,the ACC and Macro-F1 values of LEP model are increased by 2.93% and 4.55%respectively on CAIL dataset and 2.29% and 3.41% on CIVIL dataset,indicating that the model can use the hierarchical structures of causes make up the imbalance of data distribution.（3）An Interpretable Legal Judgment Prediction Model based on ILJP.On the premise of ensuring the interpretability of the model,this thesis designs a legal judgment prediction model based on LEP and HLCP models,which combines the factual elements identification task with the cause prediction task in the same framework.More specifically,ILJP model proposes an element-attentive mechanism,which integrates the identified factual elements into the cause prediction task to improve the prediction performance.The experimental results show that compared with XLNet model,the ACC and Macro-F1 values of ILJP model are increased by5.09% and 6.88% respectively on CAIL dataset,and 0.88% and 3.22% on FSC dataset,indicating that ILJP model can alleviate the imbalance of data distribution and effectively improve the performance of factual element prediction,few-shot and confusing cause prediction of the model,which can also provide some interpretable support for the cause prediction.