Case evidence logic deduction method based on graph neural network

Abstract

The invention discloses a case evidence logic deduction method based on a graph neural network. The method comprises the following steps: S1, forming a standardized case evidence data set; S2, forming a multi-level case evidence graph; s3, performing vectorization representation on case evidence data nodes and edges of the multi-level case evidence graph, and calculating weights of the case evidence data nodes; s4, inputting the multi-level case evidence graph into a graph neural network, and calculating a logic relation weight between case evidence data nodes; s5, identifying case evidence data nodes with low credibility; and S6, based on the logical relationship weight of the case evidence data nodes and a causal reasoning result, calculating an optimal case evidence deduction path by adopting a path search algorithm, identifying a case core fact chain, and generating a case deduction conclusion. According to the invention, different types of case evidence data can be reasoned in respective hierarchies, information mixing is effectively avoided, and the accuracy and interpretability of case fact deduction are improved.

Description

Technical Field

[0001] The present invention relates to the field of case evidence technology, and in particular to a case evidence logic deduction method based on graph neural network. Background Art

[0002] With the development of artificial intelligence technology, the judicial field has gradually introduced intelligent analysis tools to assist case trials and evidence reasoning. The logical deduction of case evidence is an important part of judicial reasoning. However, in the traditional case analysis and evidence reasoning process, the construction and deduction of case facts still mainly rely on the experience and judgment of judges and lawyers, and lack of systematic and intelligent evidence chain construction methods. At present, the evidence analysis of most cases relies on manual retrieval, manual induction and manual reasoning. That is, judges and lawyers need to read the case materials one by one, extract key information from a large amount of text, image, audio and time series evidence, and infer the case fact chain based on personal experience. Although this method can ensure that legal professionals have an in-depth understanding of the case, the case analysis process is time-consuming and inefficient due to the diverse types of case evidence and the complex interrelationships.

[0003] In recent years, some legal technology systems have begun to be applied to case analysis and reasoning, mainly relying on legal knowledge graphs and keyword-based case retrieval technology to assist case analysis. Such technologies have improved the efficiency of case information retrieval to a certain extent, but there are still many problems in case reasoning and fact chain construction: on the one hand, the existing legal information retrieval system is mainly based on keyword matching, which cannot accurately deduce case facts and is difficult to identify the logical relationship of key evidence in complex cases; on the other hand, the existing case analysis system usually lacks the ability to model the causal relationship of evidence and is difficult to reason with different types of evidence, resulting in low accuracy and credibility of case deduction. In addition, existing technologies generally fail to effectively handle the hierarchical associations between case evidence and cannot distinguish the logical relationship between basic evidence, derived evidence and legal rules in the deduction process, which affects the scientificity and explainability of case reasoning.

[0004] In summary, existing technologies have obvious deficiencies in the systematization of evidence chain construction, automation of reasoning process, quantitative assessment of evidence credibility, and verification of the rationality of legal rules, and are unable to meet the intelligent needs of modern judicial case analysis. Summary of the invention

[0005] One purpose of the present invention is to propose a case evidence logic deduction method based on graph neural network. The present invention enables different types of case evidence data to be reasoned at their respective levels, effectively avoiding information confusion and improving the accuracy and interpretability of case fact deduction.

[0006] A case evidence logic deduction method based on a graph neural network according to an embodiment of the present invention includes the following steps: S1. Collect case-related information and pre-process different types of case evidence data to form a standardized case evidence data set; S2. Construct a multi-level case evidence graph based on the standardized case evidence data set, set multi-level case evidence nodes, and form a multi-level case evidence graph; S3, vectorizing the case evidence data nodes and edges of the multi-level case evidence graph, and calculating the weights of the case evidence data nodes; S4, inputting the multi-level case evidence graph into the graph neural network to calculate the logical relationship weights between the case evidence data nodes; S5. Construct a causal reasoning model based on the logical relationship weights of case evidence data nodes, calculate the causal strength between case evidence data nodes, and identify case evidence data nodes with low credibility; S6. Based on the logical relationship weights and causal reasoning results of the case evidence data nodes, a path search algorithm is used to calculate the optimal case evidence deduction path, identify the core fact chain of the case, generate the case deduction conclusion, and verify the rationality of the case deduction conclusion in combination with the constraints of the legal rule nodes.

[0007] Optionally, S1 includes the following steps: S11. Obtain case-related information and construct a case evidence data set, which includes text case evidence data, image case evidence data, voice case evidence data, and time series case evidence data. The text case evidence data includes witness testimony, contract agreements, and judgment documents, the voice case evidence data includes telephone recordings and conversation records, and the time series case evidence data includes transaction records and location information. S12. Perform word segmentation, entity recognition and semantic analysis on text case evidence data, extract case-related entity sets, perform target detection and feature extraction on image case evidence data, obtain case key target sets, perform speech recognition on voice case evidence data, transcribe it into text form, generate speech-text mapping sets, perform time alignment and trend analysis on time series case evidence data, and construct a time series feature matrix; S13. Integrate the formatting results of case evidence data to form a standardized case evidence data set : ; in, is the case text entity set, is the case image target set, is a speech-to-text mapping collection, is the time series feature matrix.

[0008] Optionally, S2 includes the following steps: S21. For each standardized case evidence data in the standardized case evidence data set , using multimodal feature fusion function to generate evidence node representation : ; Among them, LayerNorm represents the layer normalization operation, r represents the modal type of evidence data, which is taken from the standardized case evidence dataset. is the feature extraction function for modality r, mapping the evidence data d into a preliminary feature vector, is the characteristic transformation matrix of mode r, is the corresponding bias vector, The modal weight calculated by the self-attention mechanism; Through the mapping function The node set that constitutes the preliminary case evidence graph is: ; And the corresponding preliminary multi-level case evidence map: ; S22, the initial node set According to their characteristic distribution, they are divided into different levels: ; in, is any evidence node, l represents the hierarchical type of the case evidence graph, which is taken from the base case evidence layer, the inferred case evidence layer, and the legal rule layer, It is a predefined hierarchical prototype vector, reflecting the feature center that each layer should have; Then the node sets of each level are defined as: ; S23. Based on the relationship between case evidence data nodes, and deriving the case evidence layer Establish an edge between any basic node With the inference node , set its edge weight to: ; in, represents the cosine similarity between two nodes, is the Gaussian kernel scale parameter, which is used to control the similarity attenuation. is the preset similarity threshold, is an indicator function, which takes the value 1 when the condition is met, otherwise it takes the value 0; This forms the edge set: ; S24. Based on the correlation between the case evidence data nodes and the legal rules, establish edges between the basic and derivation layer nodes and the legal rule layer nodes. Legal rule nodes , set its edge weight to: ; in, represents the Sigmoid function, which is used to map the similarity to [0,1], represents the inner product of two vectors; This forms the edge set: ; in, Thresholds associated with legal rules; S25, merging the edge set between the base and the derivation layer constructed in step S23 with the edge set between the case evidence and the legal rules constructed in step S24 to form a complete edge set of the multi-level case evidence graph: ; The fully constructed multi-layered case evidence diagram is as follows: .

[0009] Optionally, S3 includes the following steps: S31. Based on the multi-level case evidence graph, feature embedding is performed on each case evidence data node to generate an initial feature vector : ; in, Indicates the type embedding of case evidence data nodes, corresponding to basic case evidence, derived case evidence or legal rules, Represents the content embedding of case evidence data nodes, based on the feature extraction results of case evidence data text, image, speech or time series, Relationship embedding representing case evidence data nodes, which is calculated based on the correlation between case evidence data and other case evidence data; S32. Vectorize each edge based on the edge set and calculate the edge feature representation : ; in, are the initial feature vectors of case evidence data nodes u and v respectively, Calculate the function for edge features, using weighted concatenation or nonlinear transformation; S33. Based on the embedding vectors of case evidence data nodes and the vectorized representation of edges, a structured case evidence representation matrix is ​​constructed: ; in, is the feature matrix of the case evidence data node, each row corresponds to the initial feature vector of a case evidence data node, N is the total number of case evidence data nodes, and d is the dimension of the feature vector; Construct an adjacency matrix based on edge weight information: ; in, is the adjacency matrix of the multi-level case evidence graph, which represents the association strength between the case evidence data nodes; S34. Feature matrix based on case evidence data nodes and the adjacency matrix , calculate the global importance score of the case evidence data node: ; in, is the set of neighbor nodes of node v, is the attention weight, which indicates the influence of node u on node v: ; Among them, a is the trainable parameter vector, is the activation function; Calculate the normalized weight of the case evidence data node: .

[0010] Optionally, S4 includes the following steps: S41. Based on the adjacency matrix, the information aggregation between case evidence data nodes is calculated through the graph convolution mechanism: ; in, is the updated feature representation of the case evidence data node v in the l+1th layer of the graph neural network, is the edge weight between case evidence data nodes u and v, is the normalization factor of the case evidence data node v, is the trainable parameter matrix of layer l, is the bias vector; S42. Calculate the attention weights between case evidence data nodes based on the attention mechanism ; S43. Define the hierarchical relationship of the case evidence data nodes according to the hierarchical division, perform feature propagation, and calculate and derive the update features of the case evidence data nodes: ; in, To derive case evidence data nodes The updated feature vector at layer l+1 is, and are the feature vectors of the basic case evidence data node and the legal rule node at the lth layer, and They are the attention coefficients between the basic case evidence data node and the derived case evidence data node, and between the legal rule node and the derived case evidence data node; S44. Calculate the logical relationship weights between the case evidence data nodes based on the final feature representation of the case evidence data nodes: ; in, is the logical relationship weight between case evidence data nodes u and v, reflecting the strength of association between the two nodes. and is the feature representation of case evidence data nodes u and v at the final layer L, is the logical relationship mapping matrix, is the bias term.

[0011] Optionally, S5 includes the following steps: S51. Define the structural equation of the causal reasoning model based on the logical relationship weight matrix between case evidence data nodes: ; in, is the causal effect representation of the case evidence data node v, is the causal coefficient in the causal inference model, is the feature representation of the case evidence data node u at the final layer L, is the noise term, which indicates the random influence of the case evidence data node v; S52. Calculate the causal strength of case evidence data node u to node v based on the causal reasoning model: ; in, is the causal strength matrix of case evidence data node u to node v, reflecting the strength of the causal relationship between the two. is the adjustment coefficient of causal strength, which is used to control the nonlinear mapping of causal strength; S53. Calculate the credibility score of case evidence data nodes based on causal strength ; S54. Update the weight of the case evidence data node based on the credibility score: ; in, is the updated weight of the case evidence data node v, To adjust the coefficient, control the influence of the credibility score on the weight; S55. Setting credibility threshold , identify the set of case evidence data nodes with low credibility: ; in, It is a set of case evidence data nodes with low credibility, indicating that there are case evidence data nodes with incomplete or erroneous information.

[0012] Optionally, the S6 includes the following steps: S61. Define the path search weight matrix between case evidence data nodes based on the logical relationship weight matrix and the causal strength matrix: ; in, represents the path weight between case evidence data nodes u and v, is the logical relationship weight between case evidence data nodes u and v, As a regulatory factor, it controls the balance between the weight of the logical relationship and the causal strength; S62. Based on the path weight matrix P between the case evidence data nodes, the case initial evidence data node s and the case conclusion node t are set, and the optimal case evidence deduction path is solved by the shortest path search algorithm: ; in, The optimal path for case evidence deduction, A set of deduction paths for all possible case evidence, Indicates the path The sum of the path weights of all edges above; S63. Based on the optimal case evidence deduction path, calculate the key nodes on the path and define the core fact chain of the case: ; in, It is a collection of key case evidence data nodes on the core fact chain of the case. is the threshold of causal strength, which is used to screen case evidence data nodes with high causal influence; S64, based on the core facts of the case Defining Case Conclusions ; S65. Calculate the legal rule consistency score based on the legal rule nodes and case deduction conclusions: ; in, Score the consistency between the case deduction conclusion and the legal rules, is the weight of the logical relationship between the case evidence data node v and the legal rule node k on the case core fact chain; Based on the legal rule consistency score Setting reasonableness criteria ,like , then the case deduction conclusion is judged to be reasonable, otherwise the case evidence deduction path will be readjusted and iterative optimization will be performed.

[0013] The beneficial effects of the present invention are: (1) The present invention divides case evidence data into a basic case evidence layer, a derived case evidence layer and a legal rule layer by constructing a multi-level case evidence graph, and adopts an adaptive feature embedding method to generate a structured case evidence representation matrix based on the type, content and relevance of the evidence data, so that the case evidence data can be stored and modeled in a unified graph structure. The graph structure is used to explicitly represent the causal relationship between the case evidence data, making the reasoning path clearer, which can effectively solve the problems of incomplete evidence chain and unclear logical relationship in traditional case analysis. At the same time, different types of case evidence data can be reasoned at their respective levels, effectively avoiding information confusion and improving the accuracy and interpretability of case fact deduction.

[0014] (2) The present invention adopts graph neural network to perform logical reasoning of case evidence, and combines the attention mechanism to optimize the influence of the core evidence nodes of the case, so that the reasoning model can adaptively learn the high-order relationship between the case evidence data. In the process of information propagation, the feature propagation method based on hierarchical attention is adopted to enable the derived case evidence data nodes to obtain key information from the basic case evidence data nodes and legal rule nodes, thereby enhancing the rationality and reliability of the derived evidence in the case reasoning process. The graph neural network reasoning method can automatically learn the intrinsic correlation of case evidence data, improve the degree of automation of evidence logic analysis, and reduce the influence of human subjective factors on the case reasoning results.

[0015] (3) The present invention constructs a causal reasoning model based on the logical relationship weights and causal strength of case evidence data, calculates the credibility of case evidence data nodes, and adopts a dynamic weight adjustment method to optimize the case deduction path. In order to verify the rationality of case deduction, a legal rule consistency scoring mechanism is introduced to ensure that the case reasoning results comply with relevant legal norms. Compared with the existing case reasoning method based on statistical analysis, the causal reasoning method of the present invention can effectively identify the causal chain of case evidence data and find the optimal deduction path through a path search algorithm, thereby improving the reliability of the core fact chain of the case. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flow chart of a case evidence logic deduction method based on graph neural network proposed by the present invention. DETAILED DESCRIPTION

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0018] refer to Figure 1 , a case evidence logic deduction method based on graph neural network, including the following steps: S1. Collect case-related information and pre-process different types of case evidence data to form a standardized case evidence data set; S2. Construct a multi-level case evidence graph based on the standardized case evidence data set, set multi-level case evidence nodes, and form a multi-level case evidence graph; S3, vectorizing the case evidence data nodes and edges of the multi-level case evidence graph, and calculating the weights of the case evidence data nodes; S4, inputting the multi-level case evidence graph into the graph neural network to calculate the logical relationship weights between the case evidence data nodes; S5. Construct a causal reasoning model based on the logical relationship weights of the case evidence data nodes, calculate the causal strength between the case evidence data nodes, and identify the case evidence data nodes with low credibility; S6. Based on the logical relationship weights and causal reasoning results of the case evidence data nodes, a path search algorithm is used to calculate the optimal case evidence deduction path, identify the core fact chain of the case, generate the case deduction conclusion, and verify the rationality of the case deduction conclusion in combination with the constraints of the legal rule nodes.

[0019] In this implementation, S1 includes the following steps: S11. Obtain case-related information and construct a case evidence data set. The case evidence data set includes text case evidence data, image case evidence data, voice case evidence data, and time series case evidence data. Text case evidence data includes witness testimony, contract agreements, and judgment documents. Voice case evidence data includes telephone recordings and conversation records. Time series case evidence data includes transaction records and location information. S12. Perform word segmentation, entity recognition and semantic analysis on text case evidence data, extract case-related entity sets, perform target detection and feature extraction on image case evidence data, obtain case key target sets, perform speech recognition on voice case evidence data, transcribe it into text form, generate speech-text mapping sets, perform time alignment and trend analysis on time series case evidence data, and construct a time series feature matrix; S13. Integrate the formatting results of case evidence data to form a standardized case evidence data set : ; in, is the case text entity set, is the case image target set, is a speech-to-text mapping collection, is the time series feature matrix.

[0020] In this implementation, S2 includes the following steps: S21. For each standardized case evidence data in the standardized case evidence data set , using multimodal feature fusion function to generate evidence node representation : ; Among them, LayerNorm represents the layer normalization operation, r represents the modal type of evidence data, which is taken from the standardized case evidence dataset. is the feature extraction function for modality r, mapping the evidence data d into a preliminary feature vector, is the characteristic transformation matrix of mode r, is the corresponding bias vector, The modal weight calculated by the self-attention mechanism; Through the mapping function The node set that constitutes the preliminary case evidence graph is: ; And the corresponding preliminary multi-level case evidence map: ; S22, the initial node set According to their characteristic distribution, they are divided into different levels: ; in, is any evidence node, l represents the hierarchical type of the case evidence graph, which is taken from the base case evidence layer, the inferred case evidence layer, and the legal rule layer, It is a predefined hierarchical prototype vector, reflecting the feature center that each layer should have; Then the node sets of each level are defined as: ; S23. Based on the relationship between case evidence data nodes, and deriving the case evidence layer Establish an edge between any basic node With the inference node , set its edge weight to: ; in, represents the cosine similarity between two nodes, is the Gaussian kernel scale parameter, which is used to control the similarity attenuation. is the preset similarity threshold, is an indicator function, which takes the value 1 when the condition is met, otherwise it takes the value 0; This forms the edge set: ; S24. Based on the correlation between the case evidence data nodes and the legal rules, establish edges between the basic and derivation layer nodes and the legal rule layer nodes. Legal rule nodes , set its edge weight to: ; in, represents the Sigmoid function, which is used to map the similarity to [0,1], represents the inner product of two vectors; This forms the edge set: ; in, Thresholds associated with legal rules; S25, merging the edge set between the base and the derivation layer constructed in step S23 with the edge set between the case evidence and the legal rules constructed in step S24 to form a complete edge set of the multi-level case evidence graph: ; The fully constructed multi-layered case evidence diagram is as follows: .

[0021] In this implementation, S3 includes the following steps: S31. Based on the multi-level case evidence graph, feature embedding is performed on each case evidence data node to generate an initial feature vector : ; in, Indicates the type embedding of case evidence data nodes, corresponding to basic case evidence, derived case evidence or legal rules, Represents the content embedding of case evidence data nodes, based on the feature extraction results of case evidence data text, image, speech or time series, Relationship embedding representing case evidence data nodes, which is calculated based on the correlation between case evidence data and other case evidence data; S32. Vectorize each edge based on the edge set and calculate the edge feature representation : ; in, are the initial feature vectors of case evidence data nodes u and v respectively, Calculate the function for edge features, using weighted concatenation or nonlinear transformation; S33. Based on the embedding vectors of case evidence data nodes and the vectorized representation of edges, a structured case evidence representation matrix is ​​constructed: ; in, is the feature matrix of the case evidence data node, each row corresponds to the initial feature vector of a case evidence data node, N is the total number of case evidence data nodes, and d is the dimension of the feature vector; Construct an adjacency matrix based on edge weight information: ; in, is the adjacency matrix of the multi-level case evidence graph, which represents the association strength between the case evidence data nodes; S34. Feature matrix based on case evidence data nodes and the adjacency matrix , calculate the global importance score of the case evidence data node: ; in, is the set of neighbor nodes of node v, is the attention weight, which indicates the influence of node u on node v: ; Among them, a is the trainable parameter vector, is the activation function; Calculate the normalized weight of the case evidence data node: .

[0022] In this implementation, S4 includes the following steps: S41. Based on the adjacency matrix, the information aggregation between case evidence data nodes is calculated through the graph convolution mechanism: ; in, is the updated feature representation of the case evidence data node v in the l+1th layer of the graph neural network, is the edge weight between case evidence data nodes u and v, is the normalization factor of the case evidence data node v, is the trainable parameter matrix of layer l, is the bias vector; S42. Calculate the attention weights between case evidence data nodes based on the attention mechanism ; S43. Define the hierarchical relationship of the case evidence data nodes according to the hierarchical division, perform feature propagation, and calculate and derive the update features of the case evidence data nodes: ; in, To derive case evidence data nodes The updated feature vector at layer l+1 is, and are the feature vectors of the basic case evidence data node and the legal rule node at the lth layer, and They are the attention coefficients between the basic case evidence data node and the derived case evidence data node, and between the legal rule node and the derived case evidence data node; S44. Calculate the logical relationship weights between the case evidence data nodes based on the final feature representation of the case evidence data nodes: ; in, is the logical relationship weight between case evidence data nodes u and v, reflecting the strength of association between the two nodes. and is the feature representation of case evidence data nodes u and v at the final layer L, is the logical relationship mapping matrix, is the bias term.

[0023] In this implementation, S5 includes the following steps: S51. Define the structural equation of the causal reasoning model based on the logical relationship weight matrix between case evidence data nodes: ; in, is the causal effect representation of the case evidence data node v, is the causal coefficient in the causal inference model, is the feature representation of the case evidence data node u at the final layer L, is the noise term, which indicates the random influence of the case evidence data node v; S52. Calculate the causal strength of case evidence data node u to node v based on the causal reasoning model: ; in, is the causal strength matrix of case evidence data node u to node v, reflecting the strength of the causal relationship between the two. is the adjustment coefficient of causal strength, which is used to control the nonlinear mapping of causal strength; S53. Calculate the credibility score of case evidence data nodes based on causal strength ; S54. Update the weight of the case evidence data node based on the credibility score: ; in, is the updated weight of the case evidence data node v, To adjust the coefficient, control the influence of the credibility score on the weight; S55. Setting credibility threshold , identify the set of case evidence data nodes with low credibility: ; in, It is a set of case evidence data nodes with low credibility, indicating that there are case evidence data nodes with incomplete or erroneous information.

[0024] In this implementation, S6 includes the following steps: S61. Define the path search weight matrix between case evidence data nodes based on the logical relationship weight matrix and the causal strength matrix: ; in, represents the path weight between case evidence data nodes u and v, is the logical relationship weight between case evidence data nodes u and v, As a regulatory factor, it controls the balance between the weight of the logical relationship and the causal strength; S62. Based on the path weight matrix P between the case evidence data nodes, the case initial evidence data node s and the case conclusion node t are set, and the optimal case evidence deduction path is solved by the shortest path search algorithm: ; in, The optimal path for case evidence deduction, A set of deduction paths for all possible case evidence, Indicates the path The sum of the path weights of all edges above; S63. Based on the optimal case evidence deduction path, calculate the key nodes on the path and define the core fact chain of the case: ; in, It is a collection of key case evidence data nodes on the core fact chain of the case. is the threshold of causal strength, which is used to screen case evidence data nodes with high causal influence; S64, based on the core facts of the case Defining Case Conclusions ; S65. Calculate the legal rule consistency score based on the legal rule nodes and case deduction conclusions: ; in, Score the consistency between the case deduction conclusion and the legal rules, is the weight of the logical relationship between the case evidence data node v and the legal rule node k on the case core fact chain; Based on the legal rule consistency score Setting reasonableness criteria ,like , then the case deduction conclusion is judged to be reasonable, otherwise the case evidence deduction path will be readjusted and iterative optimization will be performed.

[0025] The present invention divides case evidence data into a basic case evidence layer, a derived case evidence layer and a legal rule layer by constructing a multi-level case evidence graph, and adopts an adaptive feature embedding method to generate a structured case evidence representation matrix based on the type, content and relevance of the evidence data, so that the case evidence data can be stored and modeled in a unified graph structure. The graph structure is used to explicitly represent the causal relationship between the case evidence data to make the reasoning path clearer, which can effectively solve the problems of incomplete evidence chain and unclear logical relationship in traditional case analysis. At the same time, different types of case evidence data can be reasoned at their respective levels, effectively avoiding information mixing and improving the accuracy and interpretability of case fact deduction.

[0026] The present invention adopts graph neural network to perform logical reasoning of case evidence, and combines the attention mechanism to optimize the influence of the core evidence nodes of the case, so that the reasoning model can adaptively learn the high-order relationship between the case evidence data. In the process of information propagation, the feature propagation method based on hierarchical attention is adopted to enable the derived case evidence data nodes to obtain key information from the basic case evidence data nodes and legal rule nodes, thereby enhancing the rationality and reliability of the derived evidence in the case reasoning process. The graph neural network reasoning method can automatically learn the intrinsic correlation of case evidence data, improve the degree of automation of evidence logic analysis, and reduce the influence of human subjective factors on the case reasoning results.

[0027] The present invention constructs a causal reasoning model based on the logical relationship weights and causal strength of case evidence data, calculates the credibility of case evidence data nodes, and adopts a dynamic weight adjustment method to optimize the case deduction path. For the rationality verification of case deduction, a legal rule consistency scoring mechanism is introduced to ensure that the case reasoning results comply with relevant legal norms. Compared with the existing case reasoning method based on statistical analysis, the causal reasoning method of the present invention can effectively identify the causal chain of case evidence data and find the optimal deduction path through a path search algorithm, thereby improving the reliability of the core fact chain of the case.

[0028] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A case evidence logic deduction method based on graph neural network, characterized in that: The steps include: S1. Collect case-related information and pre-process different types of case evidence data to form a standardized case evidence data set; S2. Construct a multi-level case evidence graph based on the standardized case evidence data set, set multi-level case evidence nodes, and form a multi-level case evidence graph; S3, vectorizing the case evidence data nodes and edges of the multi-level case evidence graph, and calculating the weights of the case evidence data nodes; S4, inputting the multi-level case evidence graph into the graph neural network to calculate the logical relationship weights between the case evidence data nodes; S5. Construct a causal reasoning model based on the logical relationship weights of case evidence data nodes, calculate the causal strength between case evidence data nodes, and identify case evidence data nodes with low credibility; S6. Based on the logical relationship weights and causal reasoning results of the case evidence data nodes, a path search algorithm is used to calculate the optimal case evidence deduction path, identify the core fact chain of the case, generate the case deduction conclusion, and verify the rationality of the case deduction conclusion in combination with the constraints of the legal rule nodes.

2. According to claim 1, a case evidence logic deduction method based on graph neural network is characterized in that: The S1 comprises the following steps: S11. Obtain case-related information and construct a case evidence data set, which includes text case evidence data, image case evidence data, voice case evidence data, and time series case evidence data. The text case evidence data includes witness testimony, contract agreements, and judgment documents, the voice case evidence data includes telephone recordings and conversation records, and the time series case evidence data includes transaction records and location information. S12. Perform word segmentation, entity recognition and semantic analysis on text case evidence data, extract case-related entity sets, perform target detection and feature extraction on image case evidence data, obtain case key target sets, perform speech recognition on voice case evidence data, transcribe it into text form, generate speech-text mapping sets, perform time alignment and trend analysis on time series case evidence data, and construct a time series feature matrix; S13. Integrate the formatting results of case evidence data to form a standardized case evidence data set : ； in, is the case text entity set, is the case image target set, is a speech-to-text mapping collection, is the time series feature matrix.

3. According to claim 2, a case evidence logic deduction method based on graph neural network is characterized in that: The S2 comprises the following steps: S21. For each standardized case evidence data in the standardized case evidence data set , using multimodal feature fusion function to generate evidence node representation : ； Among them, LayerNorm represents the layer normalization operation, r represents the modal type of evidence data, which is taken from the standardized case evidence dataset. is the feature extraction function for modality r, mapping the evidence data d into a preliminary feature vector, is the characteristic transformation matrix of mode r, is the corresponding bias vector, The modal weight calculated by the self-attention mechanism; Through the mapping function The node set that constitutes the preliminary case evidence graph is: ； And the corresponding preliminary multi-level case evidence map: ； S22, the initial node set According to their characteristic distribution, they are divided into different levels: ； in, is any evidence node, l represents the hierarchical type of the case evidence graph, which is taken from the base case evidence layer, the inferred case evidence layer, and the legal rule layer, It is a predefined hierarchical prototype vector, reflecting the feature center that each layer should have; Then the node sets of each level are defined as: ； S23. Based on the relationship between case evidence data nodes, and deriving the case evidence layer Establish an edge between any basic node With the inference node , set its edge weight to: ； in, represents the cosine similarity between two nodes, is the Gaussian kernel scale parameter, which is used to control the similarity attenuation. is the preset similarity threshold, is an indicator function, which takes the value 1 when the condition is met, otherwise it takes the value 0; This forms the edge set: ； S24. Based on the correlation between the case evidence data nodes and the legal rules, establish edges between the basic and derivation layer nodes and the legal rule layer nodes. Legal rule nodes , set its edge weight to: ； in, represents the Sigmoid function, which is used to map the similarity to [0,1], represents the inner product of two vectors; This forms the edge set: ； in, Thresholds associated with legal rules; S25, merging the edge set between the base and the derivation layer constructed in step S23 with the edge set between the case evidence and the legal rules constructed in step S24 to form a complete edge set of the multi-level case evidence graph: ； The fully constructed multi-layered case evidence diagram is as follows: 。 4. According to claim 1, a case evidence logic deduction method based on graph neural network is characterized in that: The S3 comprises the following steps: S31. Based on the multi-level case evidence graph, feature embedding is performed on each case evidence data node to generate an initial feature vector : ； in, Indicates the type embedding of case evidence data nodes, corresponding to basic case evidence, derived case evidence or legal rules, Represents the content embedding of case evidence data nodes, based on the feature extraction results of case evidence data text, image, speech or time series, Relationship embedding representing case evidence data nodes, which is calculated based on the correlation between case evidence data and other case evidence data; S32. Vectorize each edge based on the edge set and calculate the edge feature representation : ； in, are the initial feature vectors of case evidence data nodes u and v respectively, Calculate the function for edge features, using weighted concatenation or nonlinear transformation; S33. Based on the embedding vectors of case evidence data nodes and the vectorized representation of edges, a structured case evidence representation matrix is ​​constructed: ； in, is the feature matrix of the case evidence data node, each row corresponds to the initial feature vector of a case evidence data node, N is the total number of case evidence data nodes, and d is the dimension of the feature vector; Construct an adjacency matrix based on edge weight information: ； in, is the adjacency matrix of the multi-level case evidence graph, which represents the association strength between the case evidence data nodes; S34. Feature matrix based on case evidence data nodes and the adjacency matrix , calculate the global importance score of the case evidence data node: ； in, is the set of neighbor nodes of node v, is the attention weight, which indicates the influence of node u on node v: ； Among them, a is the trainable parameter vector, is the activation function; Calculate the normalized weight of the case evidence data node: 。 5. According to claim 1, a case evidence logic deduction method based on graph neural network is characterized in that: The S4 comprises the following steps: S41. Based on the adjacency matrix, the information aggregation between case evidence data nodes is calculated through the graph convolution mechanism: ； in, is the updated feature representation of the case evidence data node v in the l+1th layer of the graph neural network, is the edge weight between case evidence data nodes u and v, is the normalization factor of the case evidence data node v, is the trainable parameter matrix of layer l, is the bias vector; S42. Calculate the attention weights between case evidence data nodes based on the attention mechanism ; S43. Define the hierarchical relationship of the case evidence data nodes according to the hierarchical division, perform feature propagation, and calculate and derive the update features of the case evidence data nodes: ； in, To derive case evidence data nodes The updated feature vector at layer l+1 is, and are the feature vectors of the basic case evidence data node and the legal rule node at the lth layer, and They are the attention coefficients between the basic case evidence data node and the derived case evidence data node, and between the legal rule node and the derived case evidence data node; S44. Calculate the logical relationship weights between the case evidence data nodes based on the final feature representation of the case evidence data nodes: ； in, is the logical relationship weight between case evidence data nodes u and v, reflecting the strength of association between the two nodes. and is the feature representation of case evidence data nodes u and v at the final layer L, is the logical relationship mapping matrix, is the bias term.

6. The case evidence logic deduction method based on graph neural network according to claim 1 is characterized in that: The S5 comprises the following steps: S51. Define the structural equation of the causal reasoning model based on the logical relationship weight matrix between case evidence data nodes: ； in, is the causal effect representation of the case evidence data node v, is the causal coefficient in the causal inference model, is the feature representation of the case evidence data node u at the final layer L, is the noise term, which indicates the random influence of the case evidence data node v; S52. Calculate the causal strength of case evidence data node u to node v based on the causal reasoning model: ； in, is the causal strength matrix of case evidence data node u to node v, reflecting the strength of the causal relationship between the two. is the adjustment coefficient of causal strength, which is used to control the nonlinear mapping of causal strength; S53. Calculate the credibility score of case evidence data nodes based on causal strength ; S54. Update the weight of the case evidence data node based on the credibility score: ； in, is the updated weight of the case evidence data node v, To adjust the coefficient, control the influence of the credibility score on the weight; S55. Setting credibility threshold , identify the set of case evidence data nodes with low credibility: ； in, It is a set of case evidence data nodes with low credibility, indicating that there are case evidence data nodes with incomplete or erroneous information.

7. According to claim 1, a case evidence logic deduction method based on graph neural network is characterized in that: The S6 comprises the following steps: S61. Define the path search weight matrix between case evidence data nodes based on the logical relationship weight matrix and the causal strength matrix: ； in, represents the path weight between case evidence data nodes u and v, is the logical relationship weight between case evidence data nodes u and v, As a regulatory factor, it controls the balance between the weight of the logical relationship and the causal strength; S62. Based on the path weight matrix P between the case evidence data nodes, the case initial evidence data node s and the case conclusion node t are set, and the optimal case evidence deduction path is solved by the shortest path search algorithm: ； in, The optimal path for case evidence deduction, A set of deduction paths for all possible case evidence, Indicates the path The sum of the path weights of all edges above; S63. Based on the optimal case evidence deduction path, calculate the key nodes on the path and define the core fact chain of the case: ； in, It is a collection of key case evidence data nodes on the core fact chain of the case. is the threshold of causal strength, which is used to screen case evidence data nodes with high causal influence; S64, based on the core facts of the case Defining Case Conclusions ; S65. Calculate the legal rule consistency score based on the legal rule nodes and case deduction conclusions: ； in, Score the consistency between the case deduction conclusion and the legal rules, is the weight of the logical relationship between the case evidence data node v and the legal rule node k on the case core fact chain; Based on the legal rule consistency score Setting reasonableness criteria ,like , then the case deduction conclusion is judged to be reasonable, otherwise the case evidence deduction path will be readjusted and iterative optimization will be performed.

Images