Research On Legal Judgment Prediction Model Based On Attention Mechanism And Knowledge Fusion

With the advancement of the national legal society and the increasingly sound laws and regulations,the public has taken corresponding legal measures to safeguard their rights,which has become an important means to safeguard their rights and solve life problems,so resulting in a large amount of legal text data.However,how to use information extraction technology to extract valuable parts of legal text data to improve the efficiency and fairness of judicial decisions is one of the main research fields of intelligent justice.Legal Judgment Prediction(LJP),as a part of the intelligent judiciary,has high research value.The factual description of legal cases is an important basis for predicting the judgment results,and LJP can predict multiple sub-tasks of judgment,including legal judgment prediction,charge prediction,and term of penalty prediction,by analyzing the factual description.Legal Judgment Prediction can provide legal professionals with valuable legal advice,alleviate their workload,and provide efficient and convenient legal assistance to those in need,promoting the construction of a country ruled by law.Therefore,how to effectively model the legal judgment process using artificial intelligence technology and improve the performance of judgment is a topic of significant research significance.Most existing studies consider legal judgment prediction as a text classification task and employ machine learning or deep learning methods to classify legal texts.However,due to the diversity and complexity of legal knowledge,traditional machine learning approaches often yield suboptimal prediction results,leaving considerable room for improvement.Deep learning models,on the other hand,possess excellent generalization and data processing capabilities,which have garnered increasing attention in various fields,including the domain of legal judgment prediction.Furthermore,attention mechanisms have been widely applied in deep learning models,achieving significant success.The core idea behind attention mechanisms is to simulate a selective mechanism similar to human cognition,enabling the identification of relevant information closely related to the current task from a large amount of data.This provides a new approach for identifying key information related to a case during the judgment process.This thesis focuses on studying legal judgment prediction based on attention mechanisms and knowledge fusion.The main contributions and research content can be summarized as follows:1)This thesis proposes a Knowledge-Aware Charge Prediction(KACP)model based on attention mechanisms and knowledge awareness.The main focus of this model is charge prediction.Existing methods primarily rely on the factual description of a case to predict the charges,resulting in unreliable prediction outcomes due to the neglect of rich information from legal articles and charges.Therefore,in the actual judgment process,the relevant knowledge of charges and legal articles is crucial for the judgment.The main problem addressed by this model is how to effectively utilize this knowledge to enrich the semantic information of case descriptions,enabling the algorithm to accurately predict the charges.Building upon previous research,this thesis incorporates charges and legal articles into the model and seamlessly integrates them with case facts,enhancing the model’s awareness of legal knowledge and facilitating an understanding of the legal background knowledge related to charges and legal articles.This,in turn,improves the acquisition of key information.To fuse legal articles,a knowledge-aware module with a dual-attention mechanism is designed in the knowledge-awareness layer to enhance the interaction among legal articles,capturing syntactic and semantic features.Specific features of legal articles are introduced from the case facts to obtain enhanced representations of the facts.Then,the charge-aware module is employed to extract fused features of the factual descriptions and charge knowledge from multiple perspectives.The charge-aware module constructs a similarity graph based on charge definition information,which is used to aggregate deep semantic information of charges.It further interacts charge features with the factual vectors to capture the critical components of a case and enhance the representation of the facts.Finally,the learned knowledge representations from legal knowledge and factual representations are input into a classifier to predict the charges.To validate the effectiveness of KACP,extensive comparative experiments are conducted in Chapter 3,comparing it with relevant baseline models using multiple real criminal case datasets.The KACP model outperforms several comparative models.Additionally,Chapter 3 of this paper conducts a large number of ablation experiments to verify the impact of each module in the KACP model on the results.2)This thesis presents a sequential multi-task learning framework with task Dependencies and label Constraints for Legal Judgment Prediction(DCLJP).The framework is mainly used to solve multiple subtasks of legal judgment prediction,such as legal article,charge,and term of penalty prediction.In the actual judgment process,the three subtasks are closely related to each other and affect each other.However,existing methods often treat multiple tasks predicted by legal judgment as independent subtasks,which cannot capture the dependency and constraint relationships between subtasks,resulting in poor judgment prediction results.In order to better utilize the logical relationships between different subtasks,the model can accurately simulate the logic of judges’ decisions in real situations.This thesis formalizes the existence of dependencies between subtasks into a directed acyclic graph,and designs a forward propagation mechanism to capture simple dependencies on the directed acyclic graph.As multi-tasking labels have consistent constraints,we use a calibration function to achieve the goal of the constraint and improve prediction performance.In addition,we further adopt a reasoning method based on first-order predicate logic.It can drive the model to pay more attention to the correct term of penalty corresponding to the seriousness of the circumstance.Experimental results on two real legal datasets demonstrate that the DCLJP model exhibits significantly improved performance compared to various comparative methods.In summary,this thesis conducts research on legal judgment prediction based on attention mechanisms and knowledge fusion.Firstly,we study the task of predicting charges combined with enhanced legal knowledge,and on this basis,we propose a multi-task legal judgment prediction framework that simulates the judicial judgment process,mainly containing three sub-tasks: legal article prediction,charge prediction,and term of penalty prediction.Meanwhile,this thesis conducts multiple sets of comparative experiments,and the experimental results verify the effectiveness and feasibility of the model.