Source-linked AI summary
Legal Judgment Prediction via Multi-Perspective Bi-Feedback Network
Wenmian Yang, Weijia Jia, XIaojie Zhou, Yutao Luo
TL;DR
LJP must predict several ordered legal outcomes, but existing methods underuse dependencies among task results and word-collocation information for similar cases with different penalties. The paper introduces MPBFN with WCA, combining forward prediction, backward verification, and attention over collocations and number semantics; it reports significant improvements over baselines on all judgment prediction tasks.
Problem
Existing LJP methods inefficiently use dependencies among multiple subtasks and often ignore word-collocation and number semantics needed to distinguish similar cases with different penalties.
Method
MPBFN uses multi-perspective forward prediction and backward verification, while WCA integrates word-collocation features and number semantics through attention.
Results
MPBFN-WCA achieves significant improvements over baselines on all judgment prediction tasks.
Takeaways & Limitations
The model’s result-based bi-feedback structure and collocation-aware attention address multi-task dependencies and penalty distinctions within LJP.
Takeaways & Limitations
The study assumes each task has a single-label prediction and filters out multi-label samples; future work targets multi-label and multi-defendant prediction.
Abstract
from arXiv · showhide
The Legal Judgment Prediction (LJP) is to determine judgment results based on the fact descriptions of the cases. LJP usually consists of multiple subtasks, such as applicable law articles prediction, charges prediction, and the term of the penalty prediction. These multiple subtasks have topological dependencies, the results of which affect and verify each other. However, existing methods use dependencies of results among multiple subtasks inefficiently. Moreover, for cases with similar descriptions but different penalties, current methods cannot predict accurately because the word collocation information is ignored. In this paper, we propose a Multi-Perspective Bi-Feedback Network with the Word Collocation Attention mechanism based on the topology structure among subtasks. Specifically, we design a multi-perspective forward prediction and backward verification framework to utilize result dependencies among multiple subtasks effectively. To distinguish cases with similar descriptions but different penalties, we integrate word collocations features of fact descriptions into the network via an attention mechanism. The experimental results show our model achieves significant improvements over baselines on all prediction tasks.
1 Introduction
LJP predicts legal judgments from case facts, but existing methods underuse dependencies among subtasks and word-collocation information for distinguishing similar cases with different penalties. The paper proposes MPBFN with WCA and reports significant improvements over baselines on all subtasks.
- LJP predicts applicable law articles, charges, and penalty terms from case fact descriptions.
- Existing methods often ignore interactions among multi-task prediction results despite the strict real-world order of legal subtasks.
- Similar fact descriptions can correspond to different penalties because word collocations and number semantics are difficult to distinguish.
- MPBFN uses multi-perspective forward prediction and backward verification to incorporate dependencies among subtasks.
- WCA extracts number semantics and word-collocation information to distinguish cases with similar descriptions but different penalties.
- MPBFN-WCA significantly outperforms all baselines on all subtasks using two real-world datasets.
2 Related Work
Related work applies machine learning and neural methods to LJP, including approaches for charge, article, and multi-label prediction. Prior work also formalizes ordered legal subtasks, but does not model interactions among their results.
- LJP has evolved from traditional text classification and machine learning toward neural network methods with improved reported results.
- Prior neural approaches jointly model charge prediction and relevant article extraction, generate court views and charges, or infer charge attributes.
- Zhong et al. formalize multiple legal judgment subtasks and their strict order through a topological multi-task learning framework.
- The prior topological framework ignores interactions among task results, leaving a distinction from the proposed result-based approach.
- General multi-task learning commonly shares representations, encoding layers, or hard and soft parameters among related tasks.
3 Method
The method section presents MPBFN and WCA for exploiting topological dependencies among legal subtasks and emphasizing word-collocation information. It introduces the formulation, architecture, encoder, attention mechanism, and training process.
- The method section introduces MPBFN and WCA as the paper’s two central components.
- It first presents problem formulation and an overview before describing the fact-description encoder.
- The section then details MPBFN and WCA, followed by the model’s training process.
3.1 Problem Formulation
Each case is represented by a fact-description word sequence and extracted word-collocation sequence. The model jointly predicts applicable law articles, charges, and penalty terms as a multi-task classification problem.
- A case fact description is represented as a word sequence A = {a1, ..., al}, where l is the number of words.
- The method extracts n word collocations from the fact description with Stanford CoreNLP and forms COL = {col1, ..., coln}.
- Given A and COL, the model predicts applicable law articles, charges, and penalty terms in a multi-task classification setting.
3.2 Overview
MPBFN-WCA combines a CNN fact encoder, topology-aware bi-feedback prediction, and word-collocation attention. It uses subtask dependencies for forward prediction and backward verification, while adding number and collocation features for penalty prediction.
- MPBFN-WCA consists of a CNN encoder, the MPBFN decoder, and the WCA mechanism.
- The network defines law article, charge, and penalty prediction as tasks 1, 2, and 3 in a strict topological order.
- Bi-feedback uses each task to infer follow-up tasks and verify pre-order tasks through forward prediction and backward verification.For example, law article predictions guide charge prediction, while charge predictions verify applicable law articles.
- WCA extracts number and collocation features and combines them with MPBFN to predict penalties for similar fact descriptions more accurately.The general framework represents original fact semantics, comprehensive forward results, and final task predictions.
3.3 Neural Encoder for Fact Descriptions
The fact-description encoder converts words into embeddings, applies sliding-window convolution, and uses max-pooling to produce a fact semantic vector.
- The CNN encoder converts each word in the fact sequence into a word embedding sequence.
- Convolution applies a learned matrix to concatenated embeddings within each sliding window, producing window-level feature vectors.
- Max-pooling over the convolution outputs produces the fact semantic vector used by later components.
3.4 Multi-Perspective based Bi-Feedback Network
The MPBFN maps task-result distributions into semantic representations, uses them to predict later tasks and verify earlier tasks, and merges the resulting perspectives into final predictions.
- Each subtask result category is represented with a latent state matrix, whose weighted combination forms a result-based latent state vector.
- The latent state vector is mapped into task-specific semantic space and merged with fact semantics through an element-wise product.
- Forward Prediction: Forward prediction uses each task's result-based semantics to predict its follow-up tasks.
- Backward Verification: Backward verification maps a task's result to gate vectors for pre-order tasks and uses sigmoid values to determine available categories.
- Multi-Perspective Integration: Comprehensive forward results are summarized in topological order, intersected across perspectives, and merged with comprehensive backward results to obtain final task predictions.
3.5 Word Collocation Attention mechanism
WCA combines number-aware embeddings with task-specific attention over word collocations, then merges collocation semantics with fact semantics for penalty prediction.
- WCA is designed to improve penalty prediction when fact descriptions are similar by modeling numbers and word collocations.
- Number Embedding: Numbers are converted into fixed-length bit vectors, embedded digit by digit, and merged with the corresponding unit's word embedding.
- Number Embedding: Number embeddings are used in collocation embeddings rather than directly in the fact-description encoder.
- Collocation Encoding: Each two-word collocation is encoded with an LSTM, replacing a number-with-unit word embedding with its number-aware vector.
- Collocation Attention: Task-specific attention weights collocations according to the task representation to form a result-based collocation semantic vector.
- Penalty Prediction: The collocation-based semantic vector is merged with fact semantics and used for task 3, the term of penalty, including its forward predictions.
3.6 Training
The model assumes single-label prediction for each task and optimizes the subtasks jointly with summed cross-entropy loss. Training is end-to-end with Adam and dropout applied to reduce overfitting.
- 3.6 Training: Each prediction task is assumed to have a single label, and cross-entropy losses are summed into one overall loss.
- 3.6 Training: The model is trained end-to-end with Adam optimization, using dropout after the fact semantic vector layer to prevent overfitting.
4 Experiments
Experiments use filtered CAIL-small and CAIL-big datasets, compare MPBFN-WCA with established baselines using standard classification metrics, and evaluate individual modules. MPBFN-WCA achieves the strongest reported performance across tasks and datasets, while WCA particularly improves penalty prediction.
- 4.1 Dataset Construction and Experimental Setup: Experiments use CAIL-small and CAIL-big, filtering multi-label samples and infrequent charges or law articles while converting penalty terms into nonoverlapping intervals.
- 4.2 Results: The comparison includes FLA, HARNN, CNN, and TOPJUDGE, with accuracy, macro-precision, macro-recall, and macro-F1 used for evaluation.
- 4.2 Results: MPBFN-WCA achieves the best performance on all metrics, improving F1score over TOPJUDGE by 2.75%, 1.54%, and 13.3% across the three CAIL-small prediction tasks.
- 4.2 Results: On CAIL-big, MPBFN-WCA improves F1score over TOPJUDGE by 2.24%, 1.64%, and 8.92% for law article, charge, and penalty prediction, respectively.
- 4.2 Results: MPFP remains better than baselines without WCA and backward verification, while MPBFN remains better than TOPJUDGE on both datasets.
- 4.2 Results: WCA improves term of penalty prediction by 4.81% on F1score in CAIL-small and 3.11% in CAIL-big compared with MPBFN.
5 Conclusion
The paper addresses result-based dependencies among LJP subtasks with multi-perspective forward prediction and backward verification, while using collocation attention to reduce penalty-prediction errors. Experiments report significant improvements over baselines across all judgment-prediction tasks, with future work extending beyond the current setting.
- 5 Conclusion: MPBFN uses multi-perspective forward prediction and backward verification to exploit topological dependencies among multiple LJP subtasks.
- 5 Conclusion: Word collocations and attention are combined to reduce misjudgment in term of penalty prediction.
- 5 Conclusion: The model achieves significant improvements over baselines on all judgment prediction tasks.
- 5 Conclusion: Future work will consider multi-label and multidefendant legal prediction and apply the network to other multi-task text classification problems.