Legal document element recognition method based on big data

Abstract

The present invention relates to the field of data processing technology, and in particular to a method for identifying elements of legal documents based on big data. The content includes: obtaining legal document data to obtain original data, performing intelligent enhancement processing on the original data after preprocessing to obtain legal document data after intelligent enhancement, performing feature classification processing on the legal document data after intelligent enhancement to obtain element classification data; performing multi-level association recognition processing on the element classification data to generate a multi-level element association graph, performing structured processing based on the multi-level element association graph to obtain the identification result of the legal document elements. The method solves the problem of not being able to accurately identify key elements and the deep-level association relationship between elements in the identification of legal document elements.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for identifying legal document elements based on big data. Background Art

[0002] In the current legal document processing, accurately identifying and extracting the core elements of legal documents, especially the complex relationships between the elements, is one of the key challenges for intelligence and automation in the legal field. Legal documents usually contain highly complex syntactic structures, strict logical levels, and a large number of professional terms. They not only contain various legal clauses, factual descriptions, and party information, but also include hierarchical associations and logical references between different clauses and events. The processing of this information requires extremely high accuracy and efficiency.

[0003] Traditional text analysis methods, such as keyword extraction, rule-based matching, or simple statistical models, are unable to fully cope with the particularity of legal documents. On the one hand, these text analysis methods are often unable to cope with complex grammar and long-distance dependencies, and it is difficult to accurately extract deep-level element relationships; on the other hand, traditional text analysis methods perform poorly in dealing with association relationships, and it is difficult to automatically identify causal relationships, reference relationships, or contextual logic between clauses from unstructured texts. Therefore, how to improve the accuracy of element recognition and automatically analyze the relationship between elements is a key problem in current legal document information extraction technology.

[0004] In summary, in the identification of elements of legal documents, there is a problem of not being able to accurately identify key elements and the deep-seated correlations between elements. Summary of the invention

[0005] The present invention provides a legal document element recognition method based on big data to solve the problem that key elements and the deep-level correlation between elements cannot be accurately identified in legal document element recognition.

[0006] The method for identifying elements of legal documents based on big data includes the following steps:

[0007] S1. Obtaining legal document data to obtain original data, pre-processing the original data and then performing intelligent enhancement processing to obtain intelligently enhanced legal document data, performing feature classification processing on the intelligently enhanced legal document data to obtain element classification data;

[0008] S2. Perform multi-level association recognition processing on the element classification data to generate a multi-level element association graph, perform structured processing based on the multi-level element association graph, and obtain the recognition results of the legal document elements.

[0009] Preferably, the S1 specifically includes:

[0010] The original data is preprocessed to obtain preprocessed legal document data; a legal document intelligent enhancement processing algorithm is introduced to enhance the preprocessed legal document data to obtain intelligently enhanced legal document data.

[0011] Preferably, the S1 specifically includes:

[0012] In the process of implementing the intelligent enhancement processing algorithm for legal documents, each word in the preprocessed legal document data is semantically encoded through multi-level non-linear coding to obtain an embedding vector after non-linear coding; based on the embedding vector after non-linear coding, a nested structured transformation is used to generate a sentence vector; the sentence vector is hierarchically transformed and mapped to generate a nested structured sentence vector; the nested structured sentence vector is nonlinearly dynamically mapped and feature amplified to obtain the intelligently enhanced legal document data.

[0013] Preferably, the S2 specifically includes:

[0014] Based on the feature classification data, a multi-level recursive association deduction and adaptive graph reconstruction algorithm are introduced to perform multi-level association recognition processing and generate a multi-level feature association graph.

[0015] Preferably, the S2 specifically includes:

[0016] In the process of implementing the multi-level recursive association deduction and adaptive graph reconstruction algorithm, the initial graph structure is constructed, each feature category label in the feature classification data is used as a feature node, and the features of each feature node are initialized to generate the initial feature vector and edge weight of each feature node.

[0017] Preferably, the S2 specifically includes:

[0018] In the process of implementing the multi-level recursive association deduction and adaptive graph reconstruction algorithm, multi-level recursive deduction is performed on the initial graph structure, the feature vectors and edge weights of the feature nodes are updated through each level of recursion, and the association strength is recalculated and adjusted in each level of recursion.

[0019] Preferably, the S2 specifically includes:

[0020] In the implementation process of multi-level recursive derivation, the feature recursive update formula of the element node is as follows:

[0021] ,

[0022] in, It is Feature nodes after layer update The eigenvector of It is Edge weights for layer iterations; It is Feature nodes after layer update The eigenvector of Represents a feature node The set of neighbor nodes of is the weight coefficient of the initial eigenvector; Is a feature node The initial eigenvector of is the normalization term; It is Feature nodes after layer update The eigenvector of Used to prevent eigenvalues ​​from being too large; is a nonlinear activation function; is the adaptive inverse incidence matrix.

[0023] Preferably, the S2 specifically includes:

[0024] In the implementation process of multi-level recursive derivation, the edge weight update formula is as follows:

[0025] ,

[0026] in, For the Feature nodes after layer update and its neighbor nodes The edge weights between ; Tuning parameter for edge weight sensitivity; is the squared Euclidean distance of the feature node feature vector; It is Feature nodes after layer update The eigenvector of It is Feature nodes after layer update The eigenvector of represents the inner product of the adaptive inverse incidence matrix; Represents a feature node Adaptive inverse correlation matrix of ; Represents a feature node The adaptive inverse incidence matrix of .

[0027] Preferably, the S2 specifically includes:

[0028] In the process of implementing the multi-level recursive association derivation and adaptive graph reconstruction algorithm, edge reconstruction is performed after each level of recursion to generate element association graphs at each level; the element association graphs at each level are hierarchically weighted fused to generate a multi-level element association graph.

[0029] The beneficial effects of the technical solution of the present invention are:

[0030] 1. In the process of enhanced processing, through nonlinear dynamic mapping, sentence expansion and structural transformation, feature representations of different dimensions are generated at multiple levels, making the overall feature expression of the sentence more diversified, enhancing the expressiveness of sentence features, and making the legal document intelligent enhanced processing algorithm more flexible in processing complex grammatical structures, long sentences and ambiguous semantics, thereby improving the ability to recognize elements of legal documents.

[0031] 2. In the multi-level recursive association derivation and adaptive graph reconstruction, the adaptive association edge weight adjustment based on position, time, and semantic features is adopted to make the association relationship between adjacent, similar or similar element nodes clearer. In particular, the dynamically updated edge weight and edge reconstruction enable the element association graph to identify and reflect the hierarchical association relationship between elements, and realize the deep semantic connection between elements.

[0032] 3. The introduction of the adaptive reverse correlation matrix dynamically generates edge weights according to the changes in the characteristics of the element nodes, thereby enhancing the ability of the multi-level recursive correlation derivation and adaptive graph reconstruction algorithm to adapt to complex correlation relationships in the recursive process. Through multi-level recursive correlation derivation and feature recursive updating, the semantic reversal and long-distance dependency between elements can be captured, further improving the understanding and recognition of the structural level of legal documents. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the method for identifying elements of legal documents based on big data described in the present invention. DETAILED DESCRIPTION

[0034] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0036] The specific scheme of the method for identifying elements of legal documents based on big data provided by the present invention is described in detail below with reference to the accompanying drawings.

[0037] See attached Figure 1, which shows a flow chart of a method for identifying legal document elements based on big data provided by an embodiment of the present invention, the method comprising the following steps:

[0038] S1. Obtain legal document data to obtain original data, perform intelligent enhancement processing on the original data after preprocessing to obtain intelligently enhanced legal document data, perform feature classification processing on the intelligently enhanced legal document data to obtain element classification data;

[0039] Legal document data is obtained from multiple data sources such as legal databases and locally stored documents to obtain raw data, and the raw data is preprocessed such as data cleaning, format standardization, text segmentation and annotation to obtain preprocessed legal document data, wherein the technology used for preprocessing is the existing technology well known in the art and will not be elaborated here.

[0040] Furthermore, the legal document intelligent enhancement processing algorithm is introduced to enhance the pre-processed legal document data to obtain the intelligently enhanced legal document data; the legal document intelligent enhancement processing algorithm processes the pre-processed legal document data layer by layer, comprehensively considering the deep encoding of features, the construction of nested structures and the application of nonlinear dynamic mapping, ensuring that the ultimately generated intelligently enhanced legal document data has high-dimensional, deep-level and diversified feature expressions. The specific implementation process is as follows:

[0041] Firstly, multi-layer semantic encoding is performed on each word in the preprocessed legal document data to obtain a deep embedding vector, which provides a semantic basis for the subsequent nested structured transformation; multi-level feature encoding is used to enrich the semantic expression and obtain the basic feature vector.

[0042] Any sentence data in the preprocessed legal document data is represented as For each element , initially generate its word vector representation, each word vector It can be obtained by the following formula:

[0043] ,

[0044] in, Yes word The initial embedding vector of represents the basic semantic features of the word; is the word vector matrix, , is the embedding dimension of the word vector, The size of the vocabulary is determined according to the specific scenario; Indicates transpose.

[0045] In order to embed the word Deep feature encoding is performed at different semantic levels, and multi-level nonlinear encoding is introduced. Through multiple levels of encoding, the semantic features of each word have multi-level expression capabilities, thus adapting to complex syntactic structures. The multi-level nonlinear encoding formula is as follows:

[0046] ,

[0047] in, It is After layer nonlinear coding The embedding vector of Represents the initial word vector; It is The weight matrix of the layer is used to map the dimensions of the word vector to increase the deep features; It is After layer nonlinear coding The embedding vector of It is The bias vector of the layer provides translation parameters for the activation function; It is The activation function of the layer is used to introduce nonlinear features. Through multi-level nonlinear encoding, each word will obtain a deeply embedded feature vector , this vector has multi-level semantic information and is suitable for the input of subsequent sentence vector generation. is the total number of layers of nonlinear coding, which is determined according to the specific scenario.

[0048] Next, in order to further encode the semantic features of the entire sentence, a nested structured transformation is used to generate a sentence vector. This captures the overall semantic features of the sentence and builds the logical associations between the words in the sentence. First, the attention mechanism is introduced to generate the sentence vector to ensure that the semantic weights of different words in the sentence are fully represented. The attention weight can be calculated according to the following formula:

[0049] ,

[0050] in, Yes word The attention weight of , which indicates the relative importance of the word to the overall semantics of the sentence; Words The implicit vector of ; Words The implicit vector of ; , and They are the parameter matrix and bias of the attention mechanism respectively; is the overall context vector of the sentence, which is used to measure the contribution of each word to the overall semantics of the sentence. , the calculation formula is:

[0051] ,

[0052] in, is the number of words in the sentence, indicating the scope of the weighted sum operation of all word vectors in the sentence; It is the sentence vector, which is the overall vector expression of the sentence. It integrates the semantic information and weight of each word to form the overall feature expression of the sentence.

[0053] Furthermore, in order to make the semantic expression of sentence vector more hierarchical and structured, hierarchical transformation mapping is introduced. The sentence information is transformed into semantic expressions of different dimensions through the hierarchical structure. go through Layer transformation generates nested structured sentence vectors. The mapping process of each layer is as follows:

[0054] ,

[0055] in, is the initial value of the sentence vector, which represents the basic features of the sentence; It is The sentence vector expression of the layer indicates Sentence features after layer mapping; It is The sentence vector representation of the layer; It is The weight matrix of the layer maps the sentence vector; It is The bias term of the layer provides an offset for the nonlinear activation function; No. The sentence vector representation of the layer; is a hyperbolic tangent activation function, which increases nonlinear characteristics. Through hierarchical transformation, the sentence vector is further expressed in multiple levels. ,The mapping of each layer brings a combination of different semantic dimensions, making the expression of the sentence structure more three-dimensional.

[0056] Next, we will use the multi-level nested expression of the sentence vector Perform nonlinear dynamic mapping and feature amplification processing. First, expand the sentence structure and transform the sentence vector Generate diverse semantic expressions through nonlinear functions. Assume that the nonlinear expansion mapping formula is as follows:

[0057] ,

[0058] in, It is a sentence expression after nonlinear transformation, which is used to generate diverse semantic features; It is a parameter for extending the mapping, which is used to control the nonlinear changes of sentence expression. By combining nonlinear functions, a new sentence vector representation is generated. , so that semantic information has diverse expressions in different dimensions.

[0059] After the sentence expansion is completed, the sentence vector is further transformed into a structural one, and complex structural features are generated through the combination of multiple functions. The formula for structural transformation is:

[0060] ,

[0061] in, It is the converted sentence vector, which represents the semantics of the sentence with high complexity and multi-level expression; is a small constant used to prevent the denominator from being zero; represents the exponential function; yes Functional form, used to construct nonlinear transformations; are the terms of the polynomial expansion, is the order of expansion, and nonlinear features of different orders are generated through polynomial expansion to enrich the sentence vector The final generated It is a sentence feature with complex and multi-level expressions. yes of Power.

[0062] After the structural conversion is completed, feature fusion is performed to generate intelligently enhanced legal document data. The multi-level nested expression of the sentence vector is combined with the sentence vector after structural conversion, so that the final output of the intelligently enhanced legal document data contains both the original sentence structure and diversified nonlinear extension information. The feature fusion formula is as follows:

[0063] ,

[0064] in, is the fused sentence feature; Represents the sentence vector and expansion vector Take the average to ensure the weight balance between the two; By taking the absolute value of the difference, we ensure the difference between the features, and add a square term to amplify the difference information, thereby improving the distinguishability of the enhanced data. It not only contains the core semantics of the sentence after structural transformation, but also combines the diversified expressions after expansion, and merges the fusion features of all sentences to form intelligently enhanced legal document data. The core semantics is also combined with the expanded diversified expressions to merge the fusion features of all sentences to form intelligently enhanced legal document data. .

[0065] The legal document data after intelligent enhancement is subjected to feature classification processing using the existing multi-layer neural network model to obtain element classification data, including element category labels and corresponding feature data (i.e., legal document data after intelligent enhancement); the multi-layer neural network model includes an input layer, multiple hidden layers, and a classification output layer. The number of hidden layers is determined by the specific scenario and will not be elaborated here.

[0066] S2. Perform multi-level association recognition processing on the element classification data to generate a multi-level element association graph, perform structured processing based on the multi-level element association graph, and obtain the recognition results of the legal document elements.

[0067] Based on the feature classification data, a multi-level recursive association deduction and adaptive graph reconstruction algorithm is introduced to perform multi-level association identification processing and generate a multi-level feature association graph. The multi-level recursive association deduction and adaptive graph reconstruction algorithm maps the feature classification data into a multi-level association graph, and realizes the hierarchical and multi-dimensional association representation between the elements by continuously iteratively updating the feature vectors, association weights and graph structures of the elements. The specific implementation process is as follows:

[0068] First, the initial graph structure is constructed; each feature category label in the feature classification data is used as a feature node, and the features of each feature node are initialized to generate the initial feature vector and edge weight of each feature node. The initial association relationship of the features is captured in the graph structure by combining the semantic features, location features, and time features of the features. The initial feature vector expression is:

[0069] ,

[0070] in, Is a feature node The initial feature vector represents the basic semantic features of the feature node; It is the feature data corresponding to the feature node, that is, the feature data corresponding to the feature category label, which captures the semantic information of the feature; It is position information, indicating the relative position of an element in a legal document, such as a sentence, paragraph or chapter number, used to express the relative relationship between adjacent elements; It is a temporal feature, which indicates the time of an element or the time when an event occurs.

[0071] The edge weights are calculated by a nonlinear composite function to ensure that the edge weights not only reflect the semantic similarity of the features, but also take into account the differences in positional information between the features, so that features in adjacent positions have higher initial weights. The formula is as follows:

[0072] ,

[0073] in, Represents a feature node and its neighbor nodes The initial associated edge weight of , through which the association strength of each edge is initialized; is the adjustment parameter of the weight, which is used to control the sensitivity of the edge weight to the similarity of the feature nodes; The inner product between the initial feature vectors representing the feature nodes is used to measure the semantic similarity between the nodes; Indicates the difference in position of feature nodes. When the difference is small, the correlation is higher; Represents the total number of edges.

[0074] Furthermore, multi-level recursive deduction is performed on the initial graph structure. The feature vectors and edge weights of the feature nodes are recursively updated at each level to capture more complex associations between feature nodes layer by layer. The association strength is recalculated and adjusted at each level of recursion, so that the feature nodes and edge weights can reflect the hierarchical semantic relationships.

[0075] The feature recursive update formula of the feature node is as follows:

[0076] ,

[0077] in, It is Feature nodes after layer update The eigenvector of It is Edge weights of layer iterations, representing feature nodes and its neighbor nodes The strength of association; It is Feature nodes after layer update The eigenvector of Represents a feature node The set of neighbor nodes of is the weighted part of the initial feature vector, so that a part of the initial semantic information is always retained during the feature recursive update process. is the weight coefficient of the initial eigenvector; is the normalization term of the accumulated features, is a small constant used to prevent the eigenvalue from being too large; It is Feature nodes after layer update The eigenvector of is a nonlinear activation function (such as ), which is used to ensure that each layer of derivation can introduce nonlinear features; It is an adaptive reverse correlation matrix, which is used to adjust the correlation direction in each layer of derivation, and enhance the correlation depth of the feature nodes in the recursive derivation through the change of feature gradient; the adaptive reverse correlation matrix is ​​generated according to the gradient change of the feature of the previous layer of feature nodes, and can adapt to complex reverse correlation relationships. The formula is:

[0078] ,

[0079] in, Represents a feature node Adaptive inverse correlation matrix of ; is a function that generates a diagonal matrix, meaning that only the diagonal elements are filled and the other positions are filled with zeros; It is a tuning parameter used to control the sensitivity of feature changes of element nodes to reverse correlation; It is The characteristic gradient of the layer represents the rate of change of the feature node characteristics; represents bitwise multiplication of elements ( product), used to calculate the impact of changes in element characteristics dimension by dimension; It is the Euclidean distance between the current feature vector and the initial feature vector, ensuring that the feature change has a certain impact on the reverse correlation.

[0080] Furthermore, in order to make the edge weights and edge structures of each layer more in line with the requirements of recursive derivation, the edge weights are adaptively adjusted after the feature vectors of the element nodes of each layer are updated. Through the dynamic update of edge weights and edge reconstruction, the graph structure can reflect the recursive semantic hierarchical relationship.

[0081] The edge weight update formula is as follows:

[0082] ,

[0083] in, For the Feature nodes after layer update and its neighbor nodes The edge weights between them reflect the hierarchical semantic association; It is the edge weight sensitivity adjustment parameter, which is used to control the response of edge weight to the changes of feature node characteristics; It is the squared Euclidean distance of the feature vector of the feature node, which indicates the difference between the features of the feature nodes; It is Feature nodes after layer update The eigenvector of It is Feature nodes after layer update The eigenvector of represents the inner product of the adaptive inverse incidence matrix, which is used to introduce semantic relations in the recursive hierarchy; Represents a feature node Adaptive inverse correlation matrix of ; Represents a feature node The adaptive inverse incidence matrix of .

[0084] Furthermore, edge reconstruction is performed after each layer of recursion, and edges with edge weights below the threshold are replaced to ensure that the edge structure is adaptively updated to a structure that is more consistent with the hierarchical relationship. The edge reconstruction formula is as follows:

[0085] ,

[0086] in, Indicates The edge set after layer recursion; The edge reconstruction threshold is used to control whether edges with low correlation are retained and is set based on expert experience. is the feature node similarity threshold, which is used to ensure that the newly introduced edges are associated with feature nodes with high similarity, and is set according to the expert experience method; Represents a connected feature node and The edge of Represents a connected feature node and The edge of It is Feature nodes in the layer and The cosine similarity is used to represent the similarity of feature nodes. Finally, a multi-level feature association graph is formed. , Represents a collection of feature nodes.

[0087] Finally, the recursively generated element association graphs of each layer Perform hierarchical weighted fusion to generate a complete multi-level element association map , which is used to represent the deep relationship between elements. The fusion formula is:

[0088] ,

[0089] in, Represents the final structure of a multi-level element association graph; is the weight coefficient of each layer, which is used to control the contribution of different levels of information to the element association graph; Indicates the number of recursion levels.

[0090] Based on the multi-level element association graph, existing technologies such as knowledge graph, hierarchical clustering, and tree structure generation are used for structured processing to obtain the recognition results of legal document elements.

[0091] In summary, the method for identifying elements of legal documents based on big data has been completed.

[0092] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0093] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0094] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A method for identifying elements of legal documents based on big data, characterized in that: The following steps are involved: S1. Obtain legal document data to obtain original data, perform intelligent enhancement processing on the original data after preprocessing to obtain intelligently enhanced legal document data, perform feature classification processing on the intelligently enhanced legal document data to obtain element classification data; S2. Based on the feature classification data, a multi-level recursive association derivation and adaptive graph reconstruction algorithm is introduced to perform multi-level association recognition processing and generate a multi-level feature association graph; In the process of implementing the multi-level recursive association deduction and adaptive graph reconstruction algorithm, an initial graph structure is constructed, each feature category label in the feature classification data is used as a feature node, and the features of each feature node are initialized to generate the initial feature vector and edge weight of each feature node; on the initial graph structure, multi-level recursive deduction is performed, the feature vector and edge weight of the feature node are updated through each layer of recursion, and the association strength is recalculated and adjusted in each layer of recursion; The feature recursive update formula of the feature node is as follows: ， in, It is Feature nodes after layer update The eigenvector of It is Edge weights for layer iterations; It is Feature nodes after layer update The eigenvector of Represents a feature node The set of neighbor nodes of is the weight coefficient of the initial eigenvector; Is a feature node The initial eigenvector of is the normalization term; It is Feature nodes after layer update The eigenvector of Used to prevent eigenvalues ​​from being too large; is a nonlinear activation function; is the adaptive inverse incidence matrix; The edge weight update formula is as follows: ， in, For the Feature nodes after layer update and its neighbor nodes The edge weights between Tuning parameter for edge weight sensitivity; is the squared Euclidean distance of the feature node feature vector; It is Feature nodes after layer update The eigenvector of It is Feature nodes after layer update The eigenvector of represents the inner product of the adaptive inverse incidence matrix; Represents a feature node Adaptive inverse correlation matrix of ; Represents a feature node Adaptive inverse correlation matrix of ; Based on the multi-level element association graph, structured processing is performed to obtain the recognition results of legal document elements.

2. The method for identifying legal document elements based on big data according to claim 1 is characterized in that: The S1 specifically includes: The original data is preprocessed to obtain preprocessed legal document data; a legal document intelligent enhancement processing algorithm is introduced to enhance the preprocessed legal document data to obtain intelligently enhanced legal document data.

3. The method for identifying legal document elements based on big data according to claim 2 is characterized in that: The S1 specifically includes: In the process of implementing the intelligent enhancement processing algorithm for legal documents, each word in the preprocessed legal document data is semantically encoded through multi-level non-linear coding to obtain an embedding vector after non-linear coding; based on the embedding vector after non-linear coding, a nested structured transformation is used to generate a sentence vector; the sentence vector is hierarchically transformed and mapped to generate a nested structured sentence vector; the nested structured sentence vector is nonlinearly dynamically mapped and feature amplified to obtain the intelligently enhanced legal document data.

4. The method for identifying legal document elements based on big data according to claim 1 is characterized in that: The S2 specifically includes: In the process of implementing the multi-level recursive association derivation and adaptive graph reconstruction algorithm, edge reconstruction is performed after each level of recursion to generate element association graphs at each level; the element association graphs at each level are hierarchically weighted fused to generate a multi-level element association graph.

Images