A Judicial Information Retrieval Method and System Based on Graph Neural Networks

Abstract

This invention relates to the field of data processing technology, and in particular to a judicial information retrieval method and system based on graph neural networks. The method includes: extracting key information from judicial case text data using named entity recognition and relation extraction tools to obtain legal fact case subgraphs and legal issue case subgraphs; enhancing the structural information of the legal fact case subgraphs and legal issue case subgraphs through topology enhancement; encoding case text information and structural information using a pre-trained language model; guiding model training based on a multi-level loss function system, integrating cross-subgraph contrast loss and structural consistency loss; and using a graph neural network model to introduce a global attention mechanism for feature extraction and aggregation of the legal fact case subgraphs and legal issue case subgraphs. This model effectively encodes the complex relational structures in legal documents using graph learning algorithms, language modeling techniques, and contrastive learning methods, thereby achieving accurate matching and retrieval of relevant cases.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a judicial information retrieval method and system based on graph neural networks. Background Technology

[0002] Existing legal case retrieval methods can be broadly categorized into two types: statistical models and language models. Statistical models focus on measuring term frequency as a measure of case similarity, combining keyword frequency with its distribution across the entire document set to determine the relevance between a given case and a query case. While this method is simple and direct, it often overlooks the semantic connections and contextual relationships between words in legal language. Language models, on the other hand, represent case texts as vectors using case text representation strategies, performing nearest neighbor searches in a high-dimensional space to determine case similarity. Common case text representation strategies include sentence, paragraph, and full-text levels. Although language models achieve higher accuracy than traditional statistical models, they neglect the structural information of legal cases. Furthermore, the length of legal case texts often exceeds most input limitations, requiring text segmentation methods for processing, leading to the loss of legal information from a global perspective. Currently, methods that comprehensively consider the information structure of legal case texts and accurately identify legal case semantics are rare.

[0003] Furthermore, most existing legal case retrieval methods rely on traditional statistical and language models, which typically input the raw text of legal documents directly into the system for analysis. However, research has found that raw legal documents often contain a large number of complex legal terms, lengthy sentence structures, and intricate paragraph relationships. This makes it difficult for retrieval systems to accurately capture the deep semantic connections between legal cases, thus affecting the accuracy and relevance of retrieval results. Legal case retrieval methods need to mine the structural information and deep semantic connections of case texts. Secondly, analysis of current language models reveals that traditional language models usually rely on vector representations of text and nearest neighbor search to measure case similarity. While this method improves semantic matching capabilities to some extent, its effectiveness is still limited by text representation strategies. Especially when dealing with long and complex legal texts, the model consumes huge computational resources, resulting in slow retrieval speeds and limitations in multi-level semantic understanding. In industrial applications facing massive amounts of legal data and real-time retrieval requirements, the efficiency and scalability issues of traditional methods become particularly prominent. Legal case retrieval methods need to solve the problem of efficient retrieval of complex legal texts. Especially when dealing with complex legal issues involving multiple fields, there are currently few legal case retrieval methods that can simultaneously achieve both retrieval accuracy and efficiency. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a judicial information retrieval method and system based on graph neural networks.

[0005] In a first aspect, the present invention provides a judicial information retrieval method based on graph neural networks, which adopts the following technical solution:

[0006] A judicial information retrieval method based on graph neural networks includes:

[0007] Obtain judicial case text data;

[0008] Key information is extracted from judicial case text data using named entity recognition and relation extraction tools to obtain legal fact case subgraphs and legal issue case subgraphs. Among them, the structural information of the legal fact case subgraphs and legal issue case subgraphs is enhanced by a topology augmentation method. The training of the model is guided by a combination of cross-subgraph comparison loss and structural consistency loss based on a multi-level loss function system.

[0009] Using graph neural networks, a global attention mechanism is introduced to extract and aggregate features from legal fact case subgraphs and legal issue case subgraphs to obtain graph embedding vectors.

[0010] The graph embedding vectors are summarized using a readout function to obtain the final embedding vector;

[0011] By calculating the cosine similarity between the embedding vector of the query case and the embedding vector of the sample case, the similarity of all cases is measured and ranked.

[0012] Furthermore, the step of using named entity recognition and relation extraction tools to extract key information from judicial case text data to obtain legal fact case subgraphs and legal issue case subgraphs includes using the named entity recognition tool jieba and the relation extraction tool OpenNRE to extract key information from legal case texts, and creating nodes and edges of the graph based on the key information to obtain legal fact case subgraphs and legal issue case subgraphs.

[0013] Furthermore, the method of enhancing the structural information of the legal fact case subgraph and the legal issue case subgraph through topology enhancement includes converting the text in the nodes and edges into tokens that the model can recognize, and then inputting the tokens into the DeBERTa model for encoding to obtain node and edge features; performing standardization and normalization on the graph-level features and inputting them into the MLP to obtain the embedding of global information.

[0014] Furthermore, the model training is guided by a multi-level loss function system that integrates cross-subgraph comparison loss and structural consistency loss. This includes setting a cross-subgraph comparison loss function to identify the implicit relationships between legal facts and legal issues, and setting a structural consistency loss function to dynamically adjust node connection weights to match the content of similar cases and maintain convergence in the structural form of the legal relationship network. This is expressed as follows:

[0015] ,

[0016] in, and Represents dynamic weights. Represents the Jensen-Shannon divergence, which measures the difference in the distribution of the current case factual structure from that of typical similar structures. and represents the adjacency matrices of the case fact subgraph and the problem subgraph of the training samples, respectively. and These represent the structural prototype adjacency matrices of the case fact subgraph and the problem subgraph, respectively.

[0017] Furthermore, the method of introducing a global attention mechanism based on graph neural networks to extract and aggregate features from the legal fact case subgraph and the legal issue case subgraph to obtain graph embedding vectors includes aggregating important features globally, the features representing the nodes themselves, and integrating information about global nodes and their connecting edges through the global attention mechanism. At the same time, graph-level information is incorporated into the aggregation process to obtain node features. The feature update formula for each node is as follows:

[0018] ,

[0019] in, These are the updated node features. These are the initial features of node v. Represents graph-level features, Characteristics representing connecting edges, N (v) is the set of all nodes. It is the global attention weight. W s , W g , W e and W vu These are the weight matrices for node self-update, graph-level feature, edge feature update, and node association update, respectively, and ReLU represents the activation function.

[0020] Furthermore, the introduction of a global attention mechanism for feature extraction and aggregation of the legal fact case subgraph and the legal issue case subgraph to obtain graph embedding vectors also includes performing average pooling on node features based on global attention weights to obtain global node features. Combining the node features and global node features, complete graph representations, i.e., graph embedding vectors, are generated for the fact subgraph and the issue subgraph respectively; the local node features are represented as follows:

[0021]

[0022] in, and These are the sets of nodes in the case subgraph and the problem subgraph, respectively. and It is the global representation of the graph after pooling.

[0023] Furthermore, the process of measuring and ranking all cases by calculating the cosine similarity between the query case embedding vector and the sample case embedding vector includes calculating the legal fact subgraph embedding vector and the question subgraph embedding vector for each case, comparing the similarity between the query case embedding vector and the sample case embedding vector using cosine similarity, ranking all cases by similarity, and returning the case with the highest similarity as follows:

[0024] ,

[0025] in, and These are the graph-level features of the legal fact diagram and the legal issue diagram, respectively. and It is a complete graphical representation of legal fact diagrams and legal issue diagrams.

[0026] Secondly, a judicial information retrieval system based on graph neural networks includes:

[0027] The data acquisition module is configured to acquire judicial case text data;

[0028] The graph feature module is configured to extract key information from judicial case text data using named entity recognition and relation extraction to obtain legal fact case subgraphs and legal issue case subgraphs; among them, the structural information of the legal fact case subgraphs and legal issue case subgraphs is enhanced by a topology enhancement method; and the model training is guided by a multi-level loss function system that integrates cross-subgraph comparison loss and structural consistency loss.

[0029] The embedding module is configured to extract and aggregate features from the legal fact case subgraph and the legal issue case subgraph based on a global attention mechanism using a graph neural network to obtain graph embedding vectors; the graph embedding vectors are then aggregated through a readout function to obtain the final embedding vector.

[0030] The retrieval module is configured to measure and rank all cases by calculating the cosine similarity between the embedding vectors of query cases and the embedding vectors of sample cases.

[0031] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned judicial information retrieval method based on a graph neural network.

[0032] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a judicial information retrieval method based on a graph neural network.

[0033] In summary, the present invention has the following beneficial technical effects:

[0034] (1) In this invention, an innovative model for legal case retrieval is proposed by combining artificial intelligence and graph neural network technology. This model uses graph learning algorithms, language model technology and contrastive learning methods to effectively encode the complex relational structure in legal documents, thereby achieving accurate matching and retrieval of relevant cases.

[0035] (2) The legal case retrieval model proposed in this invention effectively addresses the shortcomings of traditional legal retrieval methods in processing long texts by introducing a graph attention network. This model enables more efficient and accurate case matching and reasoning even with limited judicial resources, thereby better helping legal professionals quickly find relevant cases.

[0036] (3) It solves the problem of insufficient understanding of the semantic structure of cases in existing legal retrieval systems. By constructing a structure based on legal semantic graphs, the model can capture the logical relationships between various elements in legal documents more deeply, and through knowledge distillation technology, it can efficiently transmit these semantic information to each layer of the model, thereby greatly improving the accuracy and efficiency of legal case retrieval. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of a judicial information retrieval method based on graph neural networks according to Embodiment 1 of the present invention. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to the accompanying drawings.

[0039] Example 1

[0040] Reference Figure 1 This embodiment of a judicial information retrieval method based on graph neural networks includes:

[0041] Obtain judicial case text data;

[0042] Key information is extracted from judicial case text data to obtain legal fact case subgraphs and legal issue case subgraphs;

[0043] The structural information of the legal fact case subgraph and the legal issue case subgraph is enhanced by a topology-based approach.

[0044] The case text and structural information are encoded using a pre-trained language model, and global feature information is introduced;

[0045] Design a multi-level loss function system that integrates cross-subgraph comparison loss and structural consistency loss to guide model training;

[0046] Using a graph neural network model, a global attention mechanism is introduced to extract and aggregate features from the legal fact case subgraph and the legal issue case subgraph to obtain graph embedding vectors.

[0047] The graph embedding vectors are summarized using a readout function to obtain the final embedding vector;

[0048] By calculating the cosine similarity between the embedding vector of the query case and the embedding vector of the sample case, the similarity of all cases is measured and ranked.

[0049] Specifically, this includes the following operations:

[0050] S1. Extract triples from case texts using entity recognition and relation extraction tools, construct structural information of cases from two aspects: legal facts and legal issues, and convert each legal case into a legal fact case subgraph and a legal issue case subgraph.

[0051] S2. A topology enhancement method is introduced, which optimizes the fact subgraph and the problem subgraph based on two core relationships in legal knowledge: the hierarchical structure of legal provisions and the relationship between case and legal provision citation.

[0052] S3. Encode the extracted case text information and structural information using a pre-trained language model to generate feature vectors for nodes and edges. Introduce graph-level global features in each subgraph to capture global information of the graph and improve the expressive power of node representation. Set global nodes based on global features.

[0053] S4. Design a multi-level loss function system that integrates cross-subgraph comparison loss and structural consistency loss to guide the model to effectively distinguish between cases similar to and dissimilar to the query case.

[0054] S5. An edge-graph attention layer is used to process the two subgraphs of the legal case. By aggregating the feature vectors of nodes and edges, the representation of each node is updated, and a more expressive graph embedding vector is learned through a global attention mechanism. A readout function is used to summarize the graph-level global features and the average of the feature vectors of all nodes, which is used as the final graph embedding vector.

[0055] Step S1 specifically includes:

[0056] S1.1: Key information is extracted from legal case texts using the Named Entity Recognition tool (jieba) and the Relationship Extraction tool (OpenNRE). This information includes entities in the case (such as parties, events) and the relationships between entities. The extraction results are represented in the form of triples (h, r, t), where h and t represent entities, and r is the relationship between them.

[0057] S1.2: Create nodes for the graph based on the entities in the triples, and create edges between nodes based on the relations in the triples. For each case, construct two subgraphs to describe the legal fact graph G of the case. lf And the legal issues diagram G describing the legal issues of the case li .

[0058] S1.3: Legal Fact Subgraph G lf The node types include parties involved, specific actions, and physical evidence. Edge types include party-specific action, party-physical evidence, and specific action-physical evidence relationships. Legal issue subgraph G li The node types include points of contention, legal provisions, and legal concepts. The edge types include dispute-legal provision associations.

[0059] Step S2 specifically includes:

[0060] S2.1: In the fact subgraph, based on the key actions described in the case (such as "intentional injury" or "property damage"), relevant legal provisions are matched from the legal provisions library. Legal provision nodes are added and connected via "case-legal provision" edges to form a hybrid graph structure.

[0061] S2.2: In the problem subgraph, add legal provision hierarchy nodes to encode the hierarchical relationship of legal provisions in the legal provisions library (such as "General Provisions-Specific Provisions-Specific Clauses") as directed edges to form a legal provision logical link.

[0062] S2.3: Calculate the number of connected components, clustering degree, graph diameter, average path length, and graph density of the legal fact subgraph and the legal problem subgraph respectively as global features F of the two subgraphs. glf and F gli .

[0063] Step S3 specifically includes:

[0064] S3.1: First, the text in the nodes and edges is converted into a token form that the model can recognize. Then, the tokens are input into the DeBERTa model for encoding to obtain the node and edge features. For node u, node v, and edge e... uv The text attribute encoding process is as follows:

[0065]

[0066]

[0067] in, , , They are node u, node v, and edge e, respectively. uv The text information is represented by Tokenizer, which converts the text information into tokens. DeBERTa is a pre-trained language model called DeBERTa.

[0068] S3.2: Graph-level features Fg lf and Fg li Standardization and normalization are performed, and the embeddings are input into the MLP to obtain global information. A global node is set in both the legal fact graph and the legal problem graph, and the embeddings are used as features of the global node. The calculation formula is as follows:

[0069]

[0070] in, and Let F represent the mean and variance of feature F, respectively. Representing the characteristics of standardization and normalization, h Feature embedding representing global nodes.

[0071] Step S4 specifically includes:

[0072] The design incorporates a multi-level loss function system, integrating cross-subgraph comparison loss and structural consistency loss, enabling the model to both grasp case feature comparisons and maintain the legal structural paradigm, thus forming an embedding space that conforms to the laws of judicial cognition.

[0073] S4.1 Design a cross-subgraph contrastive loss function to identify the implicit relationship between legal facts and legal issues.

[0074] ,

[0075] in, and To embed the fact subgraphs and problem subgraphs of the training sample cases, The temperature coefficient controls the distinguishability of similar cases. Representing the j Embedding of a problem subgraph from a training sample, N This represents the total number of sample cases in the training batch.

[0076] S4.2 Design a structural consistency loss function, which dynamically adjusts the node connection weights to ensure that similar cases not only match in content, but also maintain convergence in the structural form of the legal relationship network.

[0077] ,

[0078] in, and Represents dynamic weights. Represents the Jensen-Shannon divergence, which measures the difference in the distribution of the current case factual structure from that of typical similar structures. and represents the adjacency matrices of the case fact subgraph and the problem subgraph of the training samples, respectively. and These represent the structural prototype adjacency matrices of the case fact subgraph and the problem subgraph, respectively.

[0079] S4.3 Combines the cross-subgraph comparison loss and the structural consistency loss to form a multi-level loss function.

[0080]

[0081] Step S5 specifically includes:

[0082] The edge graph attention layer is the core component of the method. It is used to effectively aggregate information about nodes, their neighboring nodes, and edges in the graph neural network. It introduces the global attention mechanism of the Transformer to effectively capture information about non-neighboring nodes.

[0083] S5.1: A global attention mechanism is used to aggregate important features globally. The node representation not only considers its own features but also integrates information from global nodes and their connecting edges, incorporating graph-level information into the aggregation process. The feature update formula for each node is as follows:

[0084] ,

[0085] in, These are the updated node features. These are the initial features of node v. Represents graph-level features, Characteristics representing connecting edges, N (v) is the set of all nodes. It is the global attention weight. W s , W g , W e and W vu These are the weight matrices for node self-update, graph-level feature, edge feature update, and node association update, respectively, and ReLU represents the activation function.

[0086] S5.2: This is the global attention weight, calculated through the query-key dot product. The formula for calculating the global attention weight is as follows:

[0087] ,

[0088] Softmax converts the weight values ​​into probability values ​​between 0 and 1. Representative node i In the l The query vector of the layer. node j In the l Layer key vector, and Represents the learnable weight matrix. d Represents feature dimension, Represents edge encoding.

[0089] ,

[0090] in, It is the edge feature on the path. W edge It is a projection matrix. L It is the path length.

[0091] S5.3 For each node v in graph G i Its eigenvector is Average pooling is used to average the feature vectors of all nodes:

[0092]

[0093] in, and These are the sets of nodes in the case subgraph and the problem subgraph, respectively. and It is the global representation of the graph after pooling.

[0094] S5.4 Combining the representations of node features and global node features, generate a complete graph representation for both the fact subgraph and the problem subgraph.

[0095] ,

[0096] in, and These are the graph-level features of the legal fact diagram and the legal issue diagram, respectively. and It is a complete graphical representation of legal fact diagrams and legal issue diagrams.

[0097] S5.5 Calculate the legal fact subgraph embedding vector and the question subgraph embedding vector for each case, and use cosine similarity to compare the similarity between the query case embedding vector and the sample case embedding vector.

[0098] ,

[0099] Sort all cases by similarity and return the case with the highest similarity.

[0100] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned judicial information retrieval method based on a graph neural network.

[0101] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor as described in the graph neural network-based judicial information retrieval method.

[0102] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A judicial information retrieval method based on graph neural networks, characterized in that, include: Obtain judicial case text data; Key information is extracted from judicial case text data using named entity recognition and relation extraction tools to obtain legal fact case subgraphs and legal issue case subgraphs. The structural information of these subgraphs is enhanced using a topology augmentation method. A multi-level loss function system is used to guide model training, integrating cross-subgraph comparison loss and structural consistency loss. Specifically, a cross-subgraph comparison loss function is designed to identify the implicit relationships between legal facts and legal issues, expressed as follows: ， in, and To embed the fact subgraphs and problem subgraphs of the training sample cases, The temperature coefficient controls the distinguishability of similar cases. Representing the j Embedding of a problem subgraph from a training sample, N This represents the total number of sample cases in the training batch; A global attention mechanism is introduced to extract and aggregate features from the legal fact case subgraph and the legal issue case subgraph to obtain graph embedding vectors; The graph embedding vectors are summarized using a readout function to obtain the final embedding vector; By calculating the cosine similarity between the embedding vectors of query cases and the embedding vectors of sample cases, the similarity of all cases is measured and ranked. The multi-level loss function system, which integrates cross-subgraph comparison loss and structural consistency loss to guide model training, includes setting a cross-subgraph comparison loss function to identify the implicit relationship between legal facts and legal issues, and setting a structural consistency loss function to dynamically adjust node connection weights to match the content of similar cases and maintain convergence in the structural form of the legal relationship network. This is expressed as follows: ， in, and Represents dynamic weights. The Jensen-Shannon divergence represents the difference in the distribution of the current case's factual structure compared to similar typical structures. and represents the adjacency matrices of the case fact subgraph and the problem subgraph of the training samples, respectively. and The structural prototype adjacency matrices represent the factual subgraph and the problem subgraph, respectively; combined with cross-subgraph comparative loss. With structural consistency loss This forms a multi-level loss function. : .

2. The judicial information retrieval method based on graph neural networks according to claim 1, characterized in that, The method of extracting key information from judicial case text data using named entity recognition and relation extraction tools to obtain legal fact case subgraphs and legal issue case subgraphs includes using the named entity recognition tool jieba and the relation extraction tool OpenNRE to extract key information from legal case texts, and creating nodes and edges of the graph based on the key information to obtain legal fact case subgraphs and legal issue case subgraphs.

3. The judicial information retrieval method based on graph neural networks according to claim 2, characterized in that, The method described above enhances the structural information of the legal fact case subgraph and the legal issue case subgraph by means of topology enhancement. This includes converting the text in the nodes and edges into tokens that the model can recognize, and then inputting the tokens into the DeBERTa model for encoding to obtain node and edge features. The graph-level features are then standardized and normalized and input into the MLP to obtain the embedding of global information.

4. The judicial information retrieval method based on graph neural networks according to claim 3, characterized in that, The introduction of a global attention mechanism involves feature extraction and aggregation of the legal fact case subgraph and the legal issue case subgraph to obtain graph embedding vectors. This includes aggregating important features globally, the node's own representation features, and integrating information about global nodes and their connecting edges. Graph-level information is also incorporated into the aggregation process. The feature update formula for each node is as follows: ， in, These are the updated node features. These are the initial features of node v. Represents graph-level features, Characteristics representing connecting edges, N (v) is the set of all nodes. It is the global attention weight. W s , W g , W e and W vu These are the weight matrices for node self-update, graph-level feature, edge feature update, and node association update, respectively, and ReLU represents the activation function.

5. The judicial information retrieval method based on graph neural networks according to claim 4, characterized in that, The introduction of a global attention mechanism for feature extraction and aggregation of the legal fact case subgraph and the legal issue case subgraph to obtain graph embedding vectors includes performing average pooling on node features based on global attention weights to obtain global node features. Combining the node features and global node features, complete graph representations, i.e., graph embedding vectors, are generated for both the fact subgraph and the issue subgraph. The local node features are represented as follows: ， in, and These are the sets of nodes in the case subgraph and the problem subgraph, respectively. and It is the global representation of the graph after pooling.

6. The judicial information retrieval method based on graph neural networks according to claim 5, characterized in that, The process involves calculating the cosine similarity between the query case embedding vector and the sample case embedding vector to measure and rank the similarity of all cases. This includes calculating the legal fact subgraph embedding vector and the question subgraph embedding vector for each case, comparing the similarity between the query case embedding vector and the sample case embedding vector using cosine similarity, ranking all cases based on similarity, and returning the case with the highest similarity. ， in, and These are the graph-level features of the legal fact diagram and the legal issue diagram, respectively. and It is a complete graphical representation of legal fact diagrams and legal issue diagrams.

7. A judicial information retrieval system based on graph neural networks, executing the judicial information retrieval method based on graph neural networks as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire judicial case text data; The graph feature module is configured to extract key information from judicial case text data using named entity recognition and relation extraction tools to obtain legal fact case subgraphs and legal issue case subgraphs; among them, the structural information of the legal fact case subgraphs and legal issue case subgraphs is enhanced by a topology enhancement method; and the model training is guided by a multi-level loss function system that integrates cross-subgraph comparison loss and structural consistency loss. The embedding module is configured to introduce a global attention mechanism to extract and aggregate features from the legal fact case subgraph and the legal issue case subgraph to obtain a graph embedding vector; the graph embedding vector is then summarized by a readout function to obtain the final embedding vector. The retrieval module is configured to measure and rank all cases by calculating the cosine similarity between the embedding vectors of query cases and the embedding vectors of sample cases.

8. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded by the processor of the terminal device and executed as described in claim 1.

9. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is configured to implement instructions; and the computer-readable storage medium is configured to store multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as described in claim 1.

Images