Knowledge graph-based information content intelligent identification method and system

Abstract

The application relates to the technical field of electric digital data processing, and particularly discloses an information content intelligent identification method and system based on a knowledge graph, which comprises the following steps: acquiring to-be-identified information content, constructing a directed attribute graph with a time stamp mark and giving each node an initial semantic vector; extracting a time stamp sequence on each continuous path, generating a signature tensor through iterative integral transformation and compressing the signature tensor into a path order descriptor; performing neighborhood aggregation on the initial semantic vector with the path order descriptor as a topological constraint to obtain a node enhanced semantic vector; performing outer product fusion on the node enhanced semantic vector and the path order descriptor to generate a joint feature vector; performing logic compatibility calculation on the joint feature vector and a reference joint feature vector of a corresponding path in a standard knowledge graph, outputting an identification conclusion according to a compatibility value and identifying an incompatible path; and the application can effectively distinguish paths with opposite time sequence directions and improve the identification accuracy of time sequence sensitive information content.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically to a method and system for intelligent recognition of information content based on knowledge graphs. Background Technology

[0002] With the rapid development of information technology, the amount of unstructured information content generated on the Internet is growing exponentially. In the field of information content recognition, existing technologies fall into two categories: traditional recognition methods based on statistical features and keyword matching, which classify or distinguish text by calculating the similarity between the text to be recognized and a preset template; and recognition methods based on deep learning models, which use convolutional neural networks or recurrent neural networks to encode and classify text. Knowledge graphs, as a structured form of knowledge representation, are gradually being introduced into the field of information content recognition.

[0003] Existing knowledge graph-based information content recognition methods, when performing semantic parsing on temporal evolution paths with a clear sequence in the information subgraph to be identified, fail to distinguish paths with the same graph topology but completely opposite temporal directions because they only attach time information as a static attribute of nodes or edges and average it during feature compression. Furthermore, the logical compatibility calculation based on static attribute matching lacks the ability to accurately locate mismatched positions, resulting in insufficient accuracy in identifying time-sensitive information content and a lack of traceable evidence for the recognition results. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent identification of information content based on knowledge graphs, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] The knowledge graph-based intelligent information content recognition method includes the following steps:

[0007] S1: Obtain the information content to be identified, identify the entities and relationships within it, construct a directed attribute graph with timestamps, and assign an initial semantic vector to each node;

[0008] S2: Extract the timestamp sequence on each continuous path in the directed attribute graph, generate a signature tensor representing the path order structure through iterative integral transformation, and reduce the dimension of the signature tensor to a path order descriptor;

[0009] S3: Using the path order descriptor as a topological constraint, the initial semantic vector of the node is aggregated with neighborhood information under the topological constraint to obtain the node enhanced semantic vector carrying path temporal information.

[0010] S4: Perform an outer product fusion of the node enhanced semantic vector and the path order descriptor to generate a joint feature vector that fuses the node semantics and the path order;

[0011] S5: Calculate the logical compatibility between the joint feature vector and the reference joint feature vector of the corresponding path in the standard knowledge graph, output the identification conclusion based on the calculated compatibility value, and identify incompatible paths.

[0012] As a further aspect of the present invention: the process of assigning the initial semantic vector is as follows:

[0013] Semantic element extraction is performed on the information to be identified to obtain each semantic element and its associated labels;

[0014] An initial graph structure is constructed using each semantic element as a node and associated labels as directed edges, and a corresponding original timestamp is attached to each directed edge to form a directed attribute graph.

[0015] Map each node in the directed attribute graph to a preset concept space, and assign an initial semantic vector to each node based on the positional relationship of each node in the preset concept space.

[0016] Combine the initial semantic vector with the directed attributes Figure 1 And output it.

[0017] As a further aspect of the present invention: the process of obtaining the path order descriptor is as follows:

[0018] Extract the timestamp sequence carried by each continuous path from the directed attribute graph, and convert each timestamp into a cumulative time increment sequence with the path start point as the reference.

[0019] Based on the order of each term in the cumulative time increment sequence, perform multiple nested integral transformations on each term to generate a high-order signature tensor representing the path sequence structure;

[0020] Perform a logarithmic compression transformation on the higher-order signature tensor to obtain the path order descriptor.

[0021] As a further aspect of the present invention: the execution of multiple nested integral transforms specifically includes:

[0022] The cumulative time increment sequence is divided into multiple groups according to the path traversal order, and each group contains two adjacent time increments;

[0023] The two time increments in each group are arranged in order to form a directed binary pair;

[0024] The directed binary pairs generated from the previous group are cross-combined with the directed binary pairs from the next group to generate a higher-order nested combination sequence.

[0025] Numerical weights are assigned to each component based on its position in the higher-order nested combination sequence, forming a higher-order signature tensor.

[0026] As a further aspect of the present invention: the process of obtaining the node enhanced semantic vector is as follows:

[0027] Using the path order descriptor as a reference, determine the temporal position of the current node on each associated path;

[0028] The set of forward traceable nodes and the set of backward extendable nodes are defined by their temporal position;

[0029] Assign corresponding aggregation weights to the neighborhood nodes in each set based on the path order descriptor;

[0030] The initial semantic vector of the current node is combined with the initial semantic vectors of each neighboring node according to the aggregation weight to obtain the node enhanced semantic vector.

[0031] As a further aspect of the present invention: the process of allocating the aggregation weights is as follows:

[0032] The current node's position in each associated path is determined based on the path order descriptor, and the position is used as a reference.

[0033] For the set of forward traceable nodes, a first weight is assigned based on the distance between each node and the reference benchmark, with higher weights for closer nodes.

[0034] For the backward extendable node set, a second weight is assigned based on the extension span of each node from the reference benchmark, with a higher weight for smaller spans.

[0035] The first and second weights are rounded up to a uniform numerical range, and the rounded values ​​are output as the aggregate weights.

[0036] As a further aspect of the present invention: the process of generating the joint feature vector is as follows:

[0037] The node augmented semantic vector is expanded into multiple dimensions based on the path length recorded by the path order descriptor;

[0038] Each dimension is grouped and cross-combined with the path order descriptor to form a preliminary joint tensor;

[0039] Based on the numerical distribution of each component in the path order descriptor, the initial joint tensor is rearranged in position and normalized in value to generate a joint feature vector.

[0040] As a further aspect of the present invention: the identification of incompatible paths specifically includes:

[0041] Extract the preset reference paths and their reference joint feature vectors from the standard knowledge graph that correspond to the paths of the information content to be identified;

[0042] The joint feature vector and the reference joint feature vector are mapped to the same feature representation space to form the first feature distribution and the second feature distribution.

[0043] Compare the feature values ​​at corresponding positions in the first feature distribution and the second feature distribution in turn, and record the positions and number of feature value mismatches;

[0044] The compatibility value is determined based on the ratio of the number of mismatched positions to the total number of comparison positions, and the path corresponding to the ratio exceeding the set threshold is marked as an incompatible path.

[0045] As a further aspect of the present invention: the formation of the first feature distribution and the second feature distribution specifically includes:

[0046] Read the total number of dimensions and the numerical distribution range of the joint feature vector and the reference joint feature vector respectively;

[0047] The maximum dimension among the joint feature vector and the reference joint feature vector is taken as the target dimension for mapping, and the widest range of numerical distribution between the joint feature vector and the reference joint feature vector is taken as the target range for mapping.

[0048] Based on the mapping target dimension, the joint feature vector and the reference joint feature vector are extended to the corresponding dimension, and based on the mapping target interval, the values ​​of each dimension of the extended joint feature vector and the reference joint feature vector are proportionally adjusted to the mapping target interval.

[0049] The adjusted joint feature vector and the reference joint feature vector are used as the first feature distribution and the second feature distribution, respectively.

[0050] A knowledge graph-based intelligent information content recognition system includes:

[0051] Graph construction module: acquires the information content to be identified, identifies the entities and relationships within it, constructs a directed attribute graph with timestamps, and assigns an initial semantic vector to each node;

[0052] The time-series signature module extracts the timestamp sequence on each continuous path in the directed attribute graph, generates a signature tensor representing the path order structure through iterative integral transformation, and reduces the dimension of the signature tensor to a path order descriptor.

[0053] Temporal Constraint Aggregation Module: Using the path order descriptor as a topological constraint, the module aggregates neighborhood information of the initial semantic vectors of nodes under the topological constraint to obtain node-enhanced semantic vectors carrying path temporal information.

[0054] Joint Feature Fusion Module: This module fuses the node enhanced semantic vector with the path order descriptor through an outer product to generate a joint feature vector that integrates node semantics and path order.

[0055] Logical compatibility decision module: Calculates the logical compatibility between the joint feature vector and the reference joint feature vector of the corresponding path in the standard knowledge graph, outputs the identification conclusion based on the calculated compatibility value, and identifies incompatible paths.

[0056] The beneficial effects of this invention are:

[0057] (1) This invention extracts the timestamp sequence of each continuous path in the directed attribute graph and converts it into a cumulative time increment sequence, and then performs multiple nested integral transformations to generate a high-order signature tensor, so that the sequential structure of the path is completely preserved in a numerical manner. This scheme can effectively distinguish different paths with the same node sequence but opposite temporal directions, avoids semantic confusion caused by the averaging of temporal information in the traditional vector compression process, and improves the recognition accuracy of temporally sensitive information content.

[0058] (2) This invention generates a joint feature vector by fusing the node enhanced semantic vector with the path order descriptor through an outer product. The joint feature vector is then compared with the reference joint feature vector of the corresponding path in the standard knowledge graph. The compatibility value is determined based on the ratio of the number of mismatched positions to the total number of compared positions. This invention can directly complete the logical compatibility judgment based on the standard knowledge graph without relying on a pre-trained classification model. It can also accurately locate the specific mismatched position in the incompatible path, providing a traceable basis for the identification conclusion. Attached Figure Description

[0059] The invention will now be further described with reference to the accompanying drawings.

[0060] Figure 1 This is a flowchart of the method of the present invention;

[0061] Figure 2 This is a flowchart of the process of assigning the initial semantic vector in this invention;

[0062] Figure 3 This is a system block diagram of the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Please see Figure 1 As shown, this invention is an intelligent information content recognition method based on knowledge graphs, comprising the following steps:

[0065] S1: Obtain the information content to be identified, identify the entities and relationships within it, construct a directed attribute graph with timestamps, and assign an initial semantic vector to each node;

[0066] S2: Extract the timestamp sequence on each continuous path in the directed attribute graph, generate a signature tensor representing the path order structure through iterative integral transformation, and reduce the dimension of the signature tensor to a path order descriptor;

[0067] S3: Using the path order descriptor as a topological constraint, the initial semantic vector of the node is aggregated with neighborhood information under the topological constraint to obtain the node enhanced semantic vector carrying path temporal information.

[0068] S4: Perform an outer product fusion of the node enhanced semantic vector and the path order descriptor to generate a joint feature vector that fuses the node semantics and the path order;

[0069] S5: Calculate the logical compatibility between the joint feature vector and the reference joint feature vector of the corresponding path in the standard knowledge graph, output the identification conclusion based on the calculated compatibility value, and identify incompatible paths.

[0070] Please see Figure 2 As shown, in S1, the information content to be identified is obtained, the entities and relationships within it are identified, a directed attribute graph with timestamps is constructed, and an initial semantic vector is assigned to each node, specifically including:

[0071] The information to be identified is unstructured text data. By performing word segmentation and part-of-speech tagging on the information to be identified, the semantic elements and the associated tags between the semantic elements are obtained.

[0072] The word segmentation process employs a statistical-based word segmentation algorithm. This algorithm determines the word segmentation boundary based on the co-occurrence probability of adjacent characters in the corpus. The co-occurrence probability threshold is set to 0.3, and adjacent character combinations that exceed this threshold are used as word boundaries.

[0073] Part-of-speech tagging adopts a hidden state transition method, which uses the word's own features and the part-of-speech of the previous word as the basis for determining the part-of-speech of the current word. The transition probability values ​​in the transition probability matrix are determined by the ratio of the frequency of adjacent occurrences of various parts of speech in the corpus using the maximum likelihood estimation method.

[0074] Using the obtained semantic elements as nodes and the associated labels as directed edges, an initial graph structure is constructed. This initial graph structure contains the complete set of nodes and the complete set of directed edges.

[0075] In the process of constructing the initial graph structure, for each directed edge, the time information associated with the directed edge is obtained. This time information comes from the temporal expression of the information content to be identified that appears simultaneously with the semantic elements. This time information is converted into a raw timestamp in a unified format, and the corresponding directed edge is labeled with the raw timestamp to form a directed attribute graph with timestamp marking.

[0076] The unified format is an integer sequence of Gregorian calendar year, month, day, hour, and minute, with 4 digits representing the year, 2 digits representing the month, 2 digits representing the day, 2 digits representing the hour, and 2 digits representing the minute, totaling 12 integer digits to form a complete timestamp.

[0077] The preset concept space is a high-dimensional space with a fixed number of dimensions, which is set to 256. The origin of this high-dimensional space is the average position of all known semantic elements on the same dimension, and the unit step size in each dimension direction is equal and set to 1.

[0078] Mapping each node in the directed attribute graph with timestamps to the preset conceptual space involves: determining the initial coordinates of each node in the preset conceptual space based on the semantic features of its semantic elements. The semantic features are the corresponding row vectors of the node's words in the pre-trained word vector table. The dimension of these corresponding row vectors is the same as the dimension of the preset conceptual space, both being 256 dimensions. The values ​​of each dimension of these corresponding row vectors are used as the coordinate values ​​of each dimension of the node in the preset conceptual space, thereby obtaining the position of each node in the preset conceptual space.

[0079] Each node is assigned an initial semantic vector based on its positional relationship in the preset conceptual space. Specifically, for the current node, the displacement difference between the current node and the origin in the preset conceptual space is calculated based on the current node's position in the preset conceptual space to obtain the first displacement vector.

[0080] Calculate the displacement difference between this position and the position of the farthest endpoint of the diagonal in the preset conceptual space to obtain the second displacement vector;

[0081] Calculate the displacement difference between this position and the average position of all other nodes in the preset concept space to obtain the third displacement vector;

[0082] The first, second, and third displacement vectors are added together according to their corresponding dimensions, and then the average value of each dimension is taken to obtain the initial semantic vector of the current node. This initial semantic vector maintains a dimension of 256.

[0083] The initial semantic vectors of all nodes that have been assigned are then combined with the directed attributes marked with timestamps. Figure 1 And output it as input data for subsequent steps.

[0084] In S2, the timestamp sequence of each continuous path in the directed attribute graph is extracted, and a signature tensor representing the path order structure is generated through iterative integral transformation. The signature tensor is then reduced in dimension and compressed into a path order descriptor, specifically including:

[0085] Extract each continuous path from the directed attribute graph with timestamps. Each continuous path is formed by connecting multiple nodes and directed edges between adjacent nodes in the traversal direction, which is the direction from the starting point of the path along the directed edge to the ending point of the path.

[0086] For each extracted continuous path, obtain the original timestamps carried by each directed edge traversed by the path, and arrange the original timestamps into a timestamp sequence according to the order of the path traversal direction.

[0087] The timestamps in the timestamp sequence are converted into a cumulative time increment sequence with the original timestamp corresponding to the starting point of the path as the reference. The conversion method is as follows: taking the original timestamp of the starting point of the path as the reference zero value, the original timestamps corresponding to each subsequent node on the path are successively subtracted from the reference zero value to obtain the cumulative time increment value of each node. The cumulative time increment values ​​are arranged in order according to the path traversal direction to form the cumulative time increment sequence of the path. The length of the sequence is equal to the number of directed edges traversed by the path plus 1.

[0088] Based on the order of the terms in the cumulative time increment sequence, perform multiple nested integral transformations on each term.

[0089] The specific process of performing multiple nested integral transformation is as follows: the cumulative time increment sequence is split into multiple groups according to the path traversal order. Each group contains two adjacent time increments. That is, each pair of adjacent cumulative time increment values ​​in the cumulative time increment sequence is divided into one group. If the cumulative time increment sequence contains T cumulative time increment values, then it is split into T minus 1 groups. Each group is numbered sequentially from group 1 to group T minus 1 according to the path traversal direction.

[0090] The two time increments in each group are arranged in order to form a directed binary pair. That is, for any group, the first time increment value in the group is taken as the first component of the directed binary pair, and the second time increment value is taken as the second component of the directed binary pair, thus forming the directed binary pair corresponding to the group.

[0091] The directed binary pairs generated from the previous group are cross-combined with the directed binary pairs generated from the next group to generate a higher-order nested combination sequence. The cross-combination method is as follows:

[0092] For the directed binary pair corresponding to the j-th group and the directed binary pair corresponding to the j+1-th group, take the first and second components of each respectively. Combine the first component of the j-th group and the first component of the j+1-th group to form the first higher-order term, combine the first component of the j-th group and the second component of the j+1-th group to form the second higher-order term, combine the second component of the j-th group and the first component of the j+1-th group to form the third higher-order term, and combine the second component of the j-th group and the second component of the j+1-th group to form the fourth higher-order term. Arrange the first to fourth higher-order terms in the above order to form the higher-order nested combination sequence corresponding to the two groups. Traverse all adjacent group pairs and arrange the higher-order nested combination sequences corresponding to each adjacent group pair in the order of the group pairs to form a complete higher-order nested combination sequence.

[0093] Numerical weights are assigned to the components in the higher-order nested combination sequence based on their arrangement positions, forming a higher-order signature tensor. The specific method for assigning numerical weights is as follows:

[0094] For each higher-order term in the complete higher-order nested combination sequence, obtain the permutation number of the higher-order term in the complete higher-order nested combination sequence, use the permutation number as the weight index of the higher-order term, calculate 1 / 2 of the permutation number raised to the power of the permutation number, use the calculation result as the weight value of the higher-order term, multiply the weight value by the two time increment values ​​contained in the higher-order term to obtain the weighted component value of the higher-order term, arrange all the weighted component values ​​according to the permutation order of their respective higher-order terms in the complete higher-order nested combination sequence to form a higher-order signature tensor.

[0095] Specifically, for the higher-order item with the sequence number 1, the weight is 1 / 2 raised to the power of 1, which is 0.5; for the higher-order item with the sequence number 2, the weight is 1 / 2 raised to the power of 2, which is 0.25; for the higher-order item with the sequence number 3, the weight is 1 / 2 raised to the power of 3, which is 0.125; and so on, until the entire complete higher-order nested combination sequence is traversed.

[0096] Perform a logarithmic compression transformation on the above high-order signature tensor to obtain the path order descriptor. The specific method of the logarithmic compression transformation is as follows:

[0097] Take the absolute value of each weighted component in the higher-order signature tensor, take the logarithmic value with the natural constant e as the base of the absolute value, and replace the original weighted component value with each logarithmic value to obtain the logarithmically compressed higher-order signature tensor; when the absolute value of the weighted component is 0, directly replace the original 0 value with negative 10 to the power of 10 to avoid the meaninglessness of logarithmic operation.

[0098] The values ​​contained in the logarithmically compressed higher-order signature tensor are arranged in their original order and output as the path order descriptor corresponding to the path. Each continuous path corresponds to a unique path order descriptor, and the dimension of the path order descriptor is equal to the total number of higher-order terms in the complete higher-order nested combination sequence.

[0099] In S3, the path order descriptor is used as a topological constraint. Under this constraint, the initial semantic vector of a node is aggregated with neighborhood information to obtain an enhanced semantic vector carrying path temporal information, specifically including:

[0100] Extract the current node's position on each associated path from the path order descriptor.

[0101] The path order descriptor records the sequence of nodes contained in the path and the order of each node during the generation process. The order of nodes is numbered sequentially from the starting point of the path as the first position along the traversal direction.

[0102] For the current node, in each associated path, query the position number of the node in the node sequence, and use this number as the order position of the current node under that path.

[0103] If the current node is not in the node sequence of a certain path, then the path is ignored and will not participate in subsequent calculations.

[0104] Using the current node's position as a reference, define the set of forward traceable nodes and the set of backward extendable nodes on each associated path.

[0105] The forward traceable node set contains all nodes in the path whose order position is less than that of the current node, and the backward extendable node set contains all nodes in the path whose order position is greater than that of the current node.

[0106] Merge the forward nodes of the current node on all associated paths into a total forward set, and merge the backward nodes into a total backward set. During the merge, remove duplicate nodes, and keep only one copy of each node.

[0107] Aggregate weights are assigned to each neighboring node in the total forward set and the total backward set based on the path order descriptor.

[0108] The specific allocation process is as follows: For each node in the total forward set, calculate the distance between the node and the current node, where the distance is equal to the difference between the current node's position and the forward node's position; For each node in the total backward set, calculate the extension span between the node and the current node, where the extension span is equal to the difference between the backward node's position and the current node's position.

[0109] Calculate the unnormalized weights for each node. For the forward node, the expression for calculating its unnormalized weights is as follows: ;

[0110] in, Indicates the first in the total forward set The unnormalized weights of each node, It is a natural constant. Forward attenuation coefficient, This represents the distance between the preceding node and the current node. The value is set to 0.5.

[0111] For the backward node, the expression for calculating its unnormalized weight is: ;

[0112] in, Indicates the th element in the total backward set. The unnormalized weights of each node, The backward attenuation coefficient, This represents the span between the current node and the preceding node; The value is set to 0.3.

[0113] After calculating the unnormalized weights of all forward and backward nodes respectively, add all the unnormalized weights together to obtain the total weight.

[0114] Then, divide the unnormalized weight of each node by the sum of the weights to obtain the final aggregate weight of that node. The final aggregate weight is in the range of 0 to 1, and the sum of all the final aggregate weights is equal to 1.

[0115] For a forward node, its final aggregate weight is equal to its unnormalized weight divided by the sum of its weights; for a backward node, its final aggregate weight is equal to its unnormalized weight divided by the sum of its weights.

[0116] The initial semantic vector of the current node is combined and superimposed with the initial semantic vectors of each neighboring node in the total forward set and the total backward set according to their respective final aggregation weights. Specifically, the initial semantic vector of the current node is multiplied by a fixed coefficient of 1 to obtain the weighted vector of the current node.

[0117] Multiply the initial semantic vector of each neighboring node in the total forward set by its corresponding final aggregation weight, and add the resulting products along the corresponding dimensions to obtain the forward weighted sum vector;

[0118] Multiply the initial semantic vector of each neighboring node in the total backward set by its corresponding final aggregation weight, and add the resulting products along the corresponding dimensions to obtain the backward weighted sum vector;

[0119] The weighted vector, forward weighted sum vector, and backward weighted sum vector of the current node are added together according to their corresponding dimensions to obtain the original superimposed vector.

[0120] Divide the value of each dimension in the original overlay vector by the total number of nodes participating in the overlay plus 1 (i.e., the current node itself plus the total number of nodes in the forward set plus the total number of nodes in the backward set) to obtain the average value of each dimension. This average value vector is the node-enhanced semantic vector.

[0121] The augmented semantic vector for this node maintains the same dimensionality as the initial semantic vector, namely 256 dimensions. The augmented semantic vectors for all nodes are then output and used as input for subsequent steps.

[0122] In S4, the node-enhanced semantic vector and the path order descriptor are fused by an outer product to generate a joint feature vector that fuses node semantics and path order, specifically including:

[0123] The node augmented semantic vector is expanded into multiple dimensions based on the path length recorded in the path order descriptor. The node augmented semantic vector is a 256-dimensional numerical sequence, and the path order descriptor records the path length value contained in the path, which is equal to the number of directed edges traversed by the path.

[0124] Using path length as the grouping cardinality, the 256-dimensional node-enhanced semantic vector is uniformly divided into multiple dimensional groups, as follows:

[0125] Divide 256 by the sum of the path length value and 1, and use the quotient as the number of dimensions for each dimension group.

[0126] If 256 is not divisible by the sum, then take the integer quotient as the number of dimensions for each dimension group, and the remainder is added to the last dimension group.

[0127] Taking a path length of 2 as an example, the number of dimensions in each dimension group is 256 divided by 3, which is 85. The remainder 1 is added to the last dimension group. That is, the first two dimension groups each contain 85 dimension values, and the third dimension group contains 86 dimension values.

[0128] Each dimension is grouped and cross-combined with the path order descriptor to form a preliminary joint tensor.

[0129] The path order descriptor is a sequence of numbers, the number of which is equal to the total number of higher-order terms in a complete higher-order nested combination sequence.

[0130] For each dimension group, multiply each dimension value in that group by each value in the path order descriptor to obtain a product value; then, arrange all product values ​​in a cross-order manner, alternating between the order of the dimension values ​​within the dimension group and the order of the values ​​within the path order descriptor.

[0131] The product of the first dimension value in the dimension group and each value in the path order descriptor is arranged in its original order. The product of the second dimension value in the dimension group and each value in the path order descriptor is arranged in its original order, and so on, until the last dimension value in the dimension group has been processed.

[0132] Traverse all dimension groups and connect the product value sequences corresponding to each dimension group in the order of the dimension groups to form a preliminary joint tensor. The preliminary joint tensor is a numerical sequence whose total length is equal to the total number of dimension values ​​in each dimension group multiplied by the number of values ​​in the path order descriptor.

[0133] Based on the numerical distribution of each component in the path order descriptor, the initial joint tensor is rearranged in position and normalized in value to generate a joint feature vector.

[0134] The specific method for position rearrangement is as follows: obtain the absolute value of each value in the path order descriptor, sort the values ​​in ascending order of absolute value, and obtain the sorted position index sequence.

[0135] Using this position index sequence as a rearrangement mapping table, the product values ​​in the initial joint tensor are rearranged according to the same position exchange rule to obtain the rearranged joint tensor.

[0136] The specific method of numerical normalization is as follows:

[0137] Calculate the arithmetic mean and standard deviation of each value in the rearranged joint tensor. Using the arithmetic mean as the center and twice the standard deviation as the upper and lower limits, replace the upper limit value in the rearranged joint tensor that exceeds the range with the upper limit value, replace the lower limit value that is below the range with the lower limit value, and keep the values ​​within the range unchanged to obtain the normalized numerical sequence.

[0138] Calculate the minimum and maximum values ​​of the normalized numerical sequence, subtract the minimum value from each value and then divide by the difference between the maximum and minimum values, so that all values ​​fall within the numerical range of 0 to 1. The numerical sequence that falls within this range is used as the joint feature vector of the final output.

[0139] This joint feature vector maintains the same number of dimensions as the initial joint tensor and is used for subsequent logical compatibility calculations with the reference joint feature vectors of the corresponding paths in the standard knowledge graph.

[0140] In S5, the logical compatibility of the joint feature vector with the reference joint feature vector of the corresponding path in the standard knowledge graph is calculated. Based on the calculated compatibility value, the identification conclusion is output and incompatible paths are identified. Specifically, this includes:

[0141] Extract the preset reference paths and their reference joint feature vectors from the standard knowledge graph that correspond to the paths of the information content to be identified.

[0142] A standard knowledge graph is a pre-built structured knowledge base containing multiple preset reference paths, each of which stores a corresponding reference joint feature vector.

[0143] For each path in the information content to be identified, the node sequence contained in the path is used as the matching basis to search for a preset reference path containing the same node sequence in the standard knowledge graph. If the search is successful, the reference joint feature vector corresponding to the preset reference path is extracted.

[0144] If the search fails, the path is marked as an unmatched path and subsequent calculations are terminated.

[0145] The joint feature vector and the reference joint feature vector are mapped to the same feature representation space to form the first feature distribution and the second feature distribution.

[0146] The specific mapping process is as follows: First, read the total number of the first dimension and the first numerical distribution interval of the joint feature vector; then read the total number of the second dimension and the second numerical distribution interval of the reference joint feature vector.

[0147] Compare the total number of the first dimension with the total number of the second dimension, and take the larger value as the target dimension for mapping.

[0148] Compare the first and second numerical distribution intervals. The comparison method for numerical distribution intervals is to calculate the difference between the upper and lower endpoint values ​​of each interval, and take the interval corresponding to the larger difference as the mapping target interval.

[0149] If the total number of the first dimension of the joint feature vector is less than the target dimension, then multiple zero values ​​are padded to the end of the joint feature vector. The number of padded zero values ​​is equal to the difference between the target dimension and the total number of the first dimension, resulting in the expanded joint feature vector. If the total number of the first dimension is equal to the target dimension, then no expansion is needed, and the original joint feature vector is directly used as the expanded joint feature vector.

[0150] Similarly, the same expansion process is performed on the reference joint eigenvector to obtain the expanded reference joint eigenvector.

[0151] The expanded joint feature vector and the expanded reference joint feature vector are adjusted proportionally according to the mapping target interval. The specific adjustment method is as follows:

[0152] The lower endpoint of the target interval is denoted as the lower limit of the interval, the upper endpoint is denoted as the upper limit of the interval, and the width of the interval is equal to the upper limit of the interval minus the lower limit of the interval.

[0153] For the expanded joint feature vector, calculate the minimum and maximum values ​​of each dimension in the vector. The width of the vector is equal to the maximum value minus the minimum value. Using the lower limit of the interval as a reference, subtract the minimum value from each dimension value and then divide it by the width of the vector value. Multiply the quotient by the width of the target interval and finally add the lower limit of the interval to obtain the adjusted values ​​of each dimension. All the adjusted dimension values ​​form the first feature distribution.

[0154] Perform the same adjustment operation on the expanded reference joint feature vector to obtain the second feature distribution.

[0155] The first and second feature distributions have the same total number of dimensions and the same numerical distribution interval.

[0156] Compare the feature values ​​at corresponding positions in the first feature distribution and the second feature distribution in turn, and record the positions and number of feature value mismatches.

[0157] The comparison method is as follows: starting from the first dimension position, increment sequentially to the last dimension position. For each dimension position, calculate the absolute value of the difference between the value of the first feature distribution at that position and the value of the second feature distribution at that position. If the absolute value is greater than 0.1, then the position is determined to be a mismatch position, the position number is recorded, and the mismatch count is incremented by 1.

[0158] If the absolute value is less than or equal to 0.1, the position is determined to be a matching position, and neither the position is recorded nor the number of mismatches is increased.

[0159] After traversing all dimensions, the total number of mismatched positions is obtained.

[0160] The compatibility value is determined based on the ratio of the number of mismatched positions to the total number of comparison positions. The specific method for calculating the ratio is as follows:

[0161] The compatibility value is the quotient obtained by dividing the total number of mismatched positions by the total number of all comparison positions.

[0162] The threshold value is set to 0.3. When the ratio is less than or equal to 0.3, the path is deemed to have passed the logic compatibility test, and the compatibility value of the path is output as the ratio value.

[0163] When the ratio is greater than 0.3, the path is determined to have failed the logical compatibility test. The path is marked as an incompatible path, and the compatibility value of the path and the location numbers of all mismatched positions are output.

[0164] Finally, the identification results of all paths and the marked incompatible paths are output together.

[0165] Please see Figure 3 As shown, the knowledge graph-based intelligent information content recognition system includes:

[0166] Graph construction module: acquires the information content to be identified, identifies the entities and relationships within it, constructs a directed attribute graph with timestamps, and assigns an initial semantic vector to each node;

[0167] The time-series signature module extracts the timestamp sequence on each continuous path in the directed attribute graph, generates a signature tensor representing the path order structure through iterative integral transformation, and reduces the dimension of the signature tensor to a path order descriptor.

[0168] Temporal Constraint Aggregation Module: Using the path order descriptor as a topological constraint, the module aggregates neighborhood information of the initial semantic vectors of nodes under the topological constraint to obtain node-enhanced semantic vectors carrying path temporal information.

[0169] Joint Feature Fusion Module: This module fuses the node enhanced semantic vector with the path order descriptor through an outer product to generate a joint feature vector that integrates node semantics and path order.

[0170] Logical compatibility decision module: Calculates the logical compatibility between the joint feature vector and the reference joint feature vector of the corresponding path in the standard knowledge graph, outputs the identification conclusion based on the calculated compatibility value, and identifies incompatible paths.

[0171] The working principle of this invention is as follows: Unstructured information content to be identified is acquired; semantic elements and associated tags are extracted through word segmentation and part-of-speech tagging; semantic elements are used as nodes and associated tags as directed edges; an original timestamp is added to each directed edge to construct a directed attribute graph with timestamps; simultaneously, each node is mapped to a preset concept space and assigned an initial 256-dimensional semantic vector based on its positional relationship; timestamp sequences are extracted from each continuous path in the directed attribute graph and converted into cumulative time increment sequences; a high-order signature tensor is generated by performing multiple nested integral transformations on the cumulative time increment sequences and then logarithmically compressed to obtain a path order descriptor; using the path order descriptor as a topological constraint, the order position of the current node on each path is determined, thereby defining the set of forward traceable nodes and the set of backward extendable nodes. The node set is assigned aggregation weights according to the rule that the closer the distance between forward nodes, the higher the weight, and the smaller the extension span of backward nodes, the higher the weight. The initial semantic vectors are then superimposed according to the aggregation weights to obtain node-enhanced semantic vectors. The node-enhanced semantic vectors are expanded into multiple dimensions according to the path length and grouped, and cross-combined with the path order descriptor to form a preliminary joint tensor. After position rearrangement and numerical normalization, a joint feature vector is generated. The joint feature vectors and the reference joint feature vectors of the corresponding paths in the standard knowledge graph are mapped to the same feature expression space to form two feature distributions. The feature values ​​of each corresponding position are compared and the number of mismatched positions is counted. The ratio of the number of mismatched positions to the total number of compared positions is used as the compatibility value. Paths with compatibility values ​​exceeding the threshold are marked as incompatible paths, and the identification conclusions of all paths are output.

[0172] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A knowledge graph-based intelligent information content recognition method, characterized in that, Includes the following steps: S1: Obtain the information content to be identified, identify the entities and relationships within it, construct a directed attribute graph with timestamps, and assign an initial semantic vector to each node; S2: Extract the timestamp sequence on each continuous path in the directed attribute graph, generate a signature tensor representing the path order structure through iterative integral transformation, and reduce the dimension of the signature tensor to a path order descriptor; S3: Using the path order descriptor as a topological constraint, the initial semantic vector of the node is aggregated with neighborhood information under the topological constraint to obtain the node enhanced semantic vector carrying path temporal information. S4: Perform an outer product fusion of the node enhanced semantic vector and the path order descriptor to generate a joint feature vector that fuses the node semantics and the path order; S5: Calculate the logical compatibility between the joint feature vector and the reference joint feature vector of the corresponding path in the standard knowledge graph, output the identification conclusion based on the calculated compatibility value, and identify incompatible paths.

2. The intelligent information content recognition method based on knowledge graphs according to claim 1, characterized in that, The process of assigning the initial semantic vector is as follows: Semantic element extraction is performed on the information to be identified to obtain each semantic element and its associated labels; An initial graph structure is constructed using each semantic element as a node and associated labels as directed edges, and a corresponding original timestamp is attached to each directed edge to form a directed attribute graph. Map each node in the directed attribute graph to a preset concept space, and assign an initial semantic vector to each node based on the positional relationship of each node in the preset concept space. Output the initial semantic vector and the directed attribute graph together.

3. The intelligent information content recognition method based on knowledge graphs according to claim 1, characterized in that, The process of obtaining the path order descriptor is as follows: Extract the timestamp sequence carried by each continuous path from the directed attribute graph, and convert each timestamp into a cumulative time increment sequence with the path start point as the reference. Based on the order of each term in the cumulative time increment sequence, perform multiple nested integral transformations on each term to generate a high-order signature tensor representing the path sequence structure; Perform a logarithmic compression transformation on the higher-order signature tensor to obtain the path order descriptor.

4. The intelligent information content recognition method based on knowledge graphs according to claim 3, characterized in that, The execution of the multiple nested integral transform specifically includes: The cumulative time increment sequence is divided into multiple groups according to the path traversal order, and each group contains two adjacent time increments; The two time increments in each group are arranged in order to form a directed binary pair; The directed binary pairs generated from the previous group are cross-combined with the directed binary pairs from the next group to generate a higher-order nested combination sequence. Numerical weights are assigned to each component based on its position in the higher-order nested combination sequence, forming a higher-order signature tensor.

5. The intelligent information content recognition method based on knowledge graphs according to claim 1, characterized in that, The process of obtaining the node-enhanced semantic vector is as follows: Using the path order descriptor as a reference, determine the temporal position of the current node on each associated path; The set of forward traceable nodes and the set of backward extendable nodes are defined by their temporal position; Assign corresponding aggregation weights to the neighborhood nodes in each set based on the path order descriptor; The initial semantic vector of the current node is combined with the initial semantic vectors of each neighboring node according to the aggregation weight to obtain the node enhanced semantic vector.

6. The intelligent information content recognition method based on knowledge graphs according to claim 5, characterized in that, The process of allocating the aggregation weights is as follows: The current node's position in each associated path is determined based on the path order descriptor, and the position is used as a reference. For the set of forward traceable nodes, a first weight is assigned based on the distance between each node and the reference benchmark, with higher weights for closer nodes. For the backward extendable node set, a second weight is assigned based on the extension span of each node from the reference benchmark, with a higher weight for smaller spans. The first and second weights are rounded up to a uniform numerical range, and the rounded values ​​are output as the aggregate weights.

7. The intelligent information content recognition method based on knowledge graphs according to claim 1, characterized in that, The process of generating the joint feature vector is as follows: The node augmented semantic vector is expanded into multiple dimensions based on the path length recorded by the path order descriptor; Each dimension is grouped and cross-combined with the path order descriptor to form a preliminary joint tensor; Based on the numerical distribution of each component in the path order descriptor, the initial joint tensor is rearranged in position and normalized in value to generate a joint feature vector.

8. The intelligent information content recognition method based on knowledge graphs according to claim 1, characterized in that, The incompatible paths identified specifically include: Extract the preset reference paths and their reference joint feature vectors from the standard knowledge graph that correspond to the paths of the information content to be identified; The joint feature vector and the reference joint feature vector are mapped to the same feature representation space to form the first feature distribution and the second feature distribution. Compare the feature values ​​at corresponding positions in the first feature distribution and the second feature distribution in turn, and record the positions and number of feature value mismatches; The compatibility value is determined based on the ratio of the number of mismatched positions to the total number of comparison positions, and the path corresponding to the ratio exceeding the set threshold is marked as an incompatible path.

9. The intelligent information content recognition method based on knowledge graphs according to claim 8, characterized in that, The formation of the first feature distribution and the second feature distribution specifically includes: Read the total number of dimensions and the numerical distribution range of the joint feature vector and the reference joint feature vector respectively; The maximum dimension among the joint feature vector and the reference joint feature vector is taken as the target dimension for mapping, and the widest range of numerical distribution between the joint feature vector and the reference joint feature vector is taken as the target range for mapping. Based on the mapping target dimension, the joint feature vector and the reference joint feature vector are extended to the corresponding dimension, and based on the mapping target interval, the values ​​of each dimension of the extended joint feature vector and the reference joint feature vector are proportionally adjusted to the mapping target interval. The adjusted joint feature vector and the reference joint feature vector are used as the first feature distribution and the second feature distribution, respectively.

10. A knowledge graph-based intelligent information content recognition system, characterized in that, The method for intelligent information content recognition based on knowledge graphs as described in any one of claims 1-9 includes: Graph construction module: acquires the information content to be identified, identifies the entities and relationships within it, constructs a directed attribute graph with timestamps, and assigns an initial semantic vector to each node; The time-series signature module extracts the timestamp sequence on each continuous path in the directed attribute graph, generates a signature tensor representing the path order structure through iterative integral transformation, and reduces the dimension of the signature tensor to a path order descriptor. Temporal Constraint Aggregation Module: Using the path order descriptor as a topological constraint, the module aggregates neighborhood information of the initial semantic vectors of nodes under the topological constraint to obtain node-enhanced semantic vectors carrying path temporal information. Joint Feature Fusion Module: This module fuses the node enhanced semantic vector with the path order descriptor through an outer product to generate a joint feature vector that integrates node semantics and path order. Logical compatibility decision module: Calculates the logical compatibility between the joint feature vector and the reference joint feature vector of the corresponding path in the standard knowledge graph, outputs the identification conclusion based on the calculated compatibility value, and identifies incompatible paths.

Images