Mfa-Ljp: Improved Legal Judgment Prediction Using Bert Embeddings-Based Multifractal Analysis

Legal Judgment Prediction (LJP) is the task of automatically predicting judgment results based on case fact descriptions. However, these fact descriptions can be regarded as lengthy and complex chains of events, and traditional text feature extraction methods struggle to capture the complex multi-scale features from them. To address this, we propose a legal judgment prediction model based on multifractal analysis and BERT embeddings (MFA-LJP). In this work, we treat the [CLS] embeddings from BERT as one-dimensional time series and perform multifractal analysis to reveal the multifractal properties that these time series exhibit. Furthermore, for each time series, we extract multifractal parameters and concatenate them with the corresponding output from the BiGRU to create a new feature vector. Finally, this feature vector is used as the input to a fully connected layer for the prediction tasks of applicable law articles, charges, and terms of penalty, respectively. Experiments on CAIL datasets demonstrate the effectiveness of MFA. With only 11 multifractal parameters added, MFA-LJP can achieve considerable performance improvement in three sub-tasks. Compared to the current best baseline model HD-LJP, on the CAIL-small dataset, MFA-LJP improves the F1 scores for applicable law articles, charges, and terms of penalty by 5.63%, 2.54%, and 2.95%, respectively. Similarly, on the CAIL-big dataset, MFA-LJP also shows notable achievements, with F1 scores for applicable law articles, charges, and terms of penalty improving by 3.08%, 4.33%, and 8.91%, respectively.