Semantic association discovery method based on spatio-temporal-relation-interaction joint

Abstract

The application relates to the technical field of text data retrieval and multi-source heterogeneous data management, and discloses a semantic association discovery method based on space-time-relation-interaction combination, which comprises the following steps: obtaining initial feature vectors of to-be-processed objects and space-time environment data and constructing a space-time constraint adjacency tensor; when a discrete interaction behavior sequence is empty, combining background density distribution to calculate a generalized pseudo-inverse matrix as a topological compensation item and superimposing the generalized pseudo-inverse matrix to a tensor calculation flow to balance a sparse state; when the discrete interaction behavior sequence is not empty, converting into an intention bias transition matrix, and when variance exceeds a standard, calculating intention information entropy according to a discretized gradient, and zeroing high-frequency disturbance components, adjusting optimization weight distribution by point multiplication, and calculating a target semantic association feature vector by forward feed, wherein the application endogenously weaves environment constraints into a semantic vector space, effectively suppresses false semantic associations, avoids semantic drift, eliminates operation bottlenecks, and guarantees optimization efficiency.

Description

Technical Field

[0001] This invention relates to a semantic association discovery method based on spatiotemporal-relationship-interaction joint approach, belonging to the technical fields of text data retrieval and multi-source heterogeneous data management. Background Technology

[0002] In current text data retrieval practices, mainstream technical solutions typically employ language models to extract text feature vectors and collaborate with static topological networks to extend associated features. They then assess the content similarity between data entities by measuring topological distance in high-dimensional space. While such matching frameworks can maintain basic similarity matching outputs when handling routine retrieval tasks, in real-world retrieval scenarios with dynamic uncertainties, existing processing paths decouple semantic understanding of data content from external environmental features by default, ignoring physical spatiotemporal characteristics and the dynamic shifts in operator intent. This leads to a lag in timeliness perception when data retrieval systems respond to changing conditions, and in heterogeneous data flow paths, it is easy to generate false associated routes and cause intent deviations, resulting in decreased accuracy of retrieval outputs and data ambiguity in long-term operation.

[0003] To address issues such as delayed timeliness perception and data routing deviation, linear improvements such as increasing the number of language model network layers or expanding the feature dimensions of the topology network inevitably lead to an exponential increase in the overhead of high-dimensional tensor multiplication, increasing the processing latency of data optimization. This fails to meet the constraints of real-time business systems for ultra-low response latency, leaving the system in a state of mutual constraint between real-time perception of environmental features and low-latency computational response. Higher-order associations and recommendation control in multidimensional data flow also have shortcomings. For example, Chinese invention patent application CN116992131A discloses a tensor-based hybrid recommendation method and system. The method employs a dual-path extraction of latent factor matrices using tensor neural networks and method multivariate decomposition, followed by fusion and reconstruction. This control scheme relies on a structurally complete static scoring tensor and the classification attribute dimension of the rules. However, when faced with high-frequency text retrieval scenarios accompanied by transient variations, multi-source information modeling lacks the ability to actively weave and constrain the topology of the real-time physical spatiotemporal environment. In zero-interaction cold-start scenarios or interference scenarios accompanied by high-frequency random interaction noise, it cannot implement dynamic topology compensation for the sparse state of the input, lacks an adaptive gating anti-interference mechanism for sudden changes in intent, and is prone to causing divergence in high-dimensional spatial feature mapping calculations, leading to deterioration of retrieval output accuracy and response delay.

[0004] Therefore, how to combine the high-dimensional spatial topological distance metric function with real-time interactive behavior features to resolve the divergence of feature matrix calculation and suppress response delay in the feedforward flow path has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A semantic association discovery method based on spatiotemporal-relationship-interaction joint approach, comprising the following steps:

[0006] Step S1: Obtain the initial feature vector of the object to be processed and the spatiotemporal environment data corresponding to the object to be processed;

[0007] Step S2: Combine the spatiotemporal environment data and the logical link relationship of the object to be processed to construct a spatiotemporal constrained adjacency tensor;

[0008] Step S3: Capture the discrete interaction behavior sequence for the object to be processed within the current sliding time window; when the discrete interaction behavior sequence is empty, read the background context node density distribution under the global history section, calculate the generalized pseudo-inverse matrix of the spatiotemporal constraint adjacency tensor as a compensation term and superimpose it into the tensor mapping flow to balance the input sparse state and maintain structural convergence.

[0009] Step S4: When the discrete interaction behavior sequence is not empty, convert the discrete interaction behavior sequence into an intention bias transition matrix; when the variance of the discrete interaction behavior sequence exceeds the fluctuation benchmark value, calculate the discretized gradient of the discrete interaction behavior sequence to calculate the intention information entropy within the current sliding time window; when the intention information entropy exceeds the safety truncation threshold, truncate and set the high-frequency perturbation component in the intention bias transition matrix to zero, multiply the obtained intention bias transition matrix by the spatiotemporal constraint adjacency tensor, adjust the optimization weight distribution in the high-dimensional space, and thus calculate the target semantic association feature vector in a forward feedforward manner.

[0010] Preferably, step S3, where the generalized pseudo-inverse matrix is ​​superimposed as a compensation term onto the tensor mapping flow to maintain structural convergence, includes the following sub-steps: Step S31, the generalized pseudo-inverse matrix is ​​weighted using a constant adjustment factor, and the weighted generalized pseudo-inverse matrix is ​​added to the spatiotemporal constraint adjacency tensor to obtain the topology correction tensor; Step S32, the tensor product between the initial eigenvector and the topology correction tensor is calculated, and the tensor product result is normalized and scaled using a density scaling factor to eliminate sparsity divergence, and is output as the tensor mapping flow after the compensation term is superimposed.

[0011] Preferably, step S2, which combines spatiotemporal environment data and the logical link relationship of the object to be processed, constructs a spatiotemporal constrained adjacency tensor, including the following sub-steps: Step S21, parsing the spatiotemporal environment data to extract time feature intervals and geospatial coordinates; Step S22, calculating the time-sensitivity deviation between the time feature intervals and the sudden periodicity of the object to be processed, and simultaneously calculating the topological distance between the geospatial coordinates and known entity nodes; Step S23, based on the time-sensitivity deviation and topological distance, adjusting the entity association strength in the static knowledge graph through a collaborative decay method to generate a spatiotemporal constrained adjacency tensor.

[0012] Preferably, step S4, which converts the discrete interaction behavior sequence into an intent bias transition matrix and truncates and zeroes the high-frequency perturbation components, includes the following sub-steps: Step S41, mapping each historical interaction action in the discrete interaction behavior sequence to a state transition event in the state space, and counting the state transition frequency between adjacent state transition events; Step S42, normalizing the state transition frequency to construct an initial transition probability matrix; Step S43, when the intent information entropy exceeds the safety truncation threshold, identifying elements in the initial transition probability matrix whose values ​​are greater than the high-frequency perturbation threshold as high-frequency perturbation components; Step S44, setting the values ​​of the high-frequency perturbation components to zero to generate the intent bias transition matrix.

[0013] Preferably, step S4 involves dot-multiplying the obtained intention bias transfer matrix with the spatiotemporal constraint adjacency tensor to adjust the optimization weight distribution in the high-dimensional space, including the following sub-steps: Step S45, dot-multiplying the intention bias transfer matrix with each relation section matrix in the spatiotemporal constraint adjacency tensor to utilize the user's dynamic interactive intention to weight and modulate the spatiotemporal constraint relationship; Step S46, in the forward flow path, adjusting the feature mapping step size of each axis in the high-dimensional feature space according to the dot-multiply result to complete the adjustment of the optimization weight distribution.

[0014] Preferably, step S4 calculates the target semantic association feature vector in a forward feedforward manner, including the following sub-steps: step S47, inputting the initial feature vector into the forward propagation network constructed based on the adjusted optimization weight distribution; step S48, extracting cross-domain association features layer by layer through tensor mapping flow, and directly calculating the output target semantic association feature vector under the constraint of no backward gradient propagation update.

[0015] Preferably, step S3, which involves reading the background context node density distribution under the global historical section and calculating the generalized pseudo-inverse matrix, includes the following sub-steps: Step S33, extracting the historical association frequency of each logical node in the physical spatiotemporal coordinate system from the global historical section; Step S34, calculating the spatial kernel density distribution of the historical association frequency to generate the background context node density distribution; Step S35, performing singular value decomposition on the spatiotemporal constraint adjacency tensor, and multiplying the inverse of the non-zero singular values ​​with the transpose of the eigenvector matrix to construct the generalized pseudo-inverse matrix.

[0016] Preferably, in step S1, the initial feature vector of the object to be processed is obtained, and the declaration includes the following sub-steps: step S11, using the trained language model to semantically parse the text content of the object to be processed and extract the original semantic feature vector; step S12, reducing the dimensionality and normalizing the original semantic feature vector to generate the initial feature vector.

[0017] Preferably, the following post-processing steps are also included: Step S5, calculate the cosine similarity between the target semantic association feature vector and the feature vectors of each candidate entity in the static semantic library; Step S6, when the cosine similarity is greater than the security association threshold, determine that there is a semantic association between the object to be processed and the corresponding candidate entity, and output the semantic association discovery result.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] 1. In the joint semantic association discovery of spatiotemporal-relationship-interaction, the initial text feature vector and the spatiotemporal constraint adjacency tensor generated by the high-dimensional space topological distance metric function are imported into the tensor computation flow. The text features are then fed forward onto the spatiotemporal topological basis using the tensor orthogonal decoupling mapping rule. This generates a contextual collaborative high-dimensional tensor containing multi-dimensional environmental constraints. This approach directly weaves physical environmental constraints into the semantic vector space as structural features, changing the traditional approach of using spatiotemporal environment as a post-condition for filtering. This makes the data flow process have physical spatiotemporal topological specificity, thereby directly suppressing false semantic associations that are similar in content but spatiotemporally unrelated within the data computation, effectively avoiding the semantic drift problem within a specific time window.

[0020] 2. By capturing the discrete interaction behavior sequence of users within the current time window in real time and converting it into an intent bias transition matrix, the matrix multiplication algorithm is used to process the high-dimensional tensor directly in conjunction with the context. The internal weight distribution of the feature tensor is dynamically reorganized and the target semantic association feature vector is calculated. This method directly introduces dynamic interaction features as feature parameters into the feedforward flow path, constructing an adaptive closed loop in which feature parameters interfere with each other. This allows the semantic discovery results to generate monotonically deterministic bias correction as the operation converges. At the same time, because the entire process is based on forward logic, the complex backward gradient propagation update process is eliminated. While ensuring the speed of operation response, the semantic recognition system achieves adaptive optimization.

[0021] 3. In non-ideal cold start operation scenarios with missing historical user action data, the background context node density distribution under the global historical section is read through the spatiotemporal topological pseudo-inverse compensation rule. The generalized pseudo-inverse matrix of the spatiotemporal constraint adjacency tensor is calculated and superimposed on the tensor mapping flow as a compensation term. This method directly balances the input sparsity state caused by the lack of dynamic data through the bidirectional topological constraints of physical spatiotemporal and logical relationships. It ensures that the tensor mapping flow can still maintain structural convergence by relying on environmental prior information when there is no dynamic interaction data input, and avoids the paralysis of semantic recognition function due to computational divergence. This ensures the basic semantic discovery capability and output determinism of the system under incomplete working conditions. Attached Figure Description

[0022] Figure 1 This is a flowchart of the spatiotemporal-relationship-interaction joint semantic association discovery method of the present invention;

[0023] Figure 2 This is a diagram of the distributed logical node and bus connection architecture of the present invention.

[0024] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0026] A semantic association discovery method based on spatiotemporal-relational-interaction joint approach includes the following steps:

[0027] Step S1: Obtain the initial feature vector of the object to be processed and the spatiotemporal environment data corresponding to the object to be processed;

[0028] Step S2: Combine the spatiotemporal environment data and the logical link relationship of the object to be processed to construct a spatiotemporal constrained adjacency tensor;

[0029] Step S3: Capture the discrete interaction behavior sequence for the object to be processed within the current sliding time window; when the discrete interaction behavior sequence is empty, read the background context node density distribution under the global history section, calculate the generalized pseudo-inverse matrix of the spatiotemporal constraint adjacency tensor as a compensation term and superimpose it into the tensor mapping flow to balance the input sparse state and maintain structural convergence.

[0030] Step S4: When the discrete interaction behavior sequence is not empty, convert the discrete interaction behavior sequence into an intention bias transition matrix; when the variance of the discrete interaction behavior sequence exceeds the fluctuation benchmark value, calculate the discretized gradient of the discrete interaction behavior sequence to calculate the intention information entropy within the current sliding time window; when the intention information entropy exceeds the safety truncation threshold, truncate and set the high-frequency perturbation component in the intention bias transition matrix to zero, multiply the obtained intention bias transition matrix by the spatiotemporal constraint adjacency tensor, adjust the optimization weight distribution in the high-dimensional space, and thus calculate the target semantic association feature vector in a forward feedforward manner.

[0031] Preferably, step S3, where the generalized pseudo-inverse matrix is ​​superimposed as a compensation term onto the tensor mapping flow to maintain structural convergence, includes the following sub-steps: Step S31, the generalized pseudo-inverse matrix is ​​weighted using a constant adjustment factor, and the weighted generalized pseudo-inverse matrix is ​​added to the spatiotemporal constraint adjacency tensor to obtain the topology correction tensor; Step S32, the tensor product between the initial eigenvector and the topology correction tensor is calculated, and the tensor product result is normalized and scaled using a density scaling factor to eliminate sparsity divergence, and is output as the tensor mapping flow after the compensation term is superimposed.

[0032] Preferably, step S2, which combines spatiotemporal environment data and the logical link relationship of the object to be processed, constructs a spatiotemporal constrained adjacency tensor, including the following sub-steps: Step S21, parsing the spatiotemporal environment data to extract time feature intervals and geospatial coordinates; Step S22, calculating the time-sensitivity deviation between the time feature intervals and the sudden periodicity of the object to be processed, and simultaneously calculating the topological distance between the geospatial coordinates and known entity nodes; Step S23, based on the time-sensitivity deviation and topological distance, adjusting the entity association strength in the static knowledge graph through a collaborative decay method to generate a spatiotemporal constrained adjacency tensor.

[0033] Preferably, step S4, which converts the discrete interaction behavior sequence into an intent bias transition matrix and truncates and zeroes the high-frequency perturbation components, includes the following sub-steps: Step S41, mapping each historical interaction action in the discrete interaction behavior sequence to a state transition event in the state space, and counting the state transition frequency between adjacent state transition events; Step S42, normalizing the state transition frequency to construct an initial transition probability matrix; Step S43, when the intent information entropy exceeds the safety truncation threshold, identifying elements in the initial transition probability matrix whose values ​​are greater than the high-frequency perturbation threshold as high-frequency perturbation components; Step S44, forcibly clearing the values ​​of the high-frequency perturbation components to zero to generate the intent bias transition matrix.

[0034] Preferably, step S4 involves dot-multiplying the obtained intention bias transfer matrix with the spatiotemporal constraint adjacency tensor to adjust the optimization weight distribution in the high-dimensional space, including the following sub-steps: Step S45, dot-multiplying the intention bias transfer matrix with each relation section matrix in the spatiotemporal constraint adjacency tensor to utilize the user's dynamic interactive intention to weight and modulate the spatiotemporal constraint relationship; Step S46, in the forward flow path, adjusting the feature mapping step size of each axis in the high-dimensional feature space according to the dot-multiply result to complete the adjustment of the optimization weight distribution.

[0035] Preferably, step S4 calculates the target semantic association feature vector in a forward feedforward manner, including the following sub-steps: step S47, inputting the initial feature vector into the forward propagation network constructed based on the adjusted optimization weight distribution; step S48, extracting cross-domain association features layer by layer through tensor mapping flow, and directly calculating the output target semantic association feature vector under the constraint of no backward gradient propagation update.

[0036] Preferably, step S3, which involves reading the background context node density distribution under the global historical section and calculating the generalized pseudo-inverse matrix, includes the following sub-steps: Step S33, extracting the historical association frequency of each logical node in the physical spatiotemporal coordinate system from the global historical section; Step S34, calculating the spatial kernel density distribution of the historical association frequency to generate the background context node density distribution; Step S35, performing singular value decomposition on the spatiotemporal constraint adjacency tensor, and multiplying the inverse of the non-zero singular values ​​with the transpose of the eigenvector matrix to construct the generalized pseudo-inverse matrix.

[0037] Preferably, in step S1, the initial feature vector of the object to be processed is obtained, and the declaration includes the following sub-steps: step S11, using the trained language model to semantically parse the text content of the object to be processed and extract the original semantic feature vector; step S12, reducing the dimensionality and normalizing the original semantic feature vector to generate the initial feature vector.

[0038] Preferably, the following post-processing steps are also included: Step S5, calculate the cosine similarity between the target semantic association feature vector and the feature vectors of each candidate entity in the static semantic library; Step S6, when the cosine similarity is greater than the security association threshold, determine that there is a semantic association between the object to be processed and the corresponding candidate entity, and output the semantic association discovery result.

[0039] Example 1: The method provided by this invention is used in a distributed semantic retrieval system that includes cross-domain related data auditing. This system faces the situation of multiple heterogeneous text records flooding in and frequently generating sudden regional semantic variations. Since traditional word vector calculation is independent of the time window and geospatial topological constraints of data generation, heterogeneous records with highly similar content but unrelated spatiotemporal backgrounds cause data feature routing deviations in similarity matching, resulting in text semantic perception lag and semantic drift. If the drift is suppressed by simply increasing the number of layers in the pre-trained language model network, it will inevitably lead to an exponential increase in the overhead of high-dimensional tensor multiplication, resulting in degraded retrieval output response latency and failure to meet the 50ms real-time business response requirement.

[0040] For the specific task, the distributed semantic retrieval system receives the object to be processed and uses de-identification and anonymization techniques to remove sensitive personal features. It then calls a trained language model to parse the text content of the object, extracting the original semantic feature vector. Based on this, it calculates the initial feature vector using dimensionality reduction and normalization scaling. Simultaneously, the system analyzes the spatiotemporal environment data corresponding to the object to be processed to extract the time feature interval and geospatial coordinates. It calculates the timeliness deviation between the time feature interval and the sudden occurrence cycle of the object to be processed. Here, the sudden occurrence cycle refers to the discrete periodic time interval presented on the time axis by the data flow management module through statistical analysis of sudden word frequency jump events within the historical 30-day time range of the text category to which the object to be processed belongs. This time interval is instantiated as a set of discrete values ​​representing the second period in a specific register in the dynamic storage area. The main control processor directly reads this discrete value and calculates the center value of the currently acquired time feature interval and the most recent historical time of the sudden word frequency jump. The difference between interstamps yields the topological distance between simultaneously calculated geospatial coordinates and known entity nodes, consisting of a defined number of seconds. Based on the timeliness deviation and the topological distance, the entity association strength in the static knowledge graph is adjusted using a corresponding attenuation method. Specifically, this collaborative attenuation method involves the main control processor executing multiplication instructions composed of digital logic circuits to perform concatenated multiplication operations on the floating-point number representing the original entity association strength of the static graph node with the normalized first-level time attenuation component and the second-level spatial attenuation component, making the output association value monotonically converge to a floating-point number between 0 and 1, thereby generating a spatiotemporal constraint adjacency tensor. Specifically, the collaborative decay method normalizes and combines time-dependent deviation and topological distance. In physics and graph theory, the strength of entity associations decays nonlinearly and exponentially with the increase of spatial distance over time. This conforms to the adaptive law of the composite distribution of spatial physical field strength and the classical time forgetting curve. Based on the objective correlation of spatiotemporal physical field decay, the system constructs a collaborative decay calculation path. It transforms the implicit physical field distribution law into a data matrix processing step using a purely textual algebraic mapping method, isolating the traditional purely static logical routing. In actual calculation, the time-dependent deviation is divided by a preset time standard base period constant, such as 3600 seconds, to obtain a dimensionless time decay variable. An exponential function of the negative value of this dimensionless variable is then calculated with the natural constant as the base to generate the first-level time decay component.

[0041] Simultaneously, the topological distance between the geospatial coordinates and the known entity nodes is divided by a preset spatial division constant, such as 500 meters, to obtain a dimensionless spatial attenuation variable. The exponential function of the negative value of this dimensionless variable is then calculated in the same way to generate the second-level spatial attenuation component. Finally, the original entity association strength values ​​in the static knowledge graph are multiplied by the first-level temporal attenuation component, and then by the second-level spatial attenuation component. This joint multiplication achieves coordinated attenuation modulation, transforming the initial feature vector... Adjacency tensor with spatiotemporal constraints Importing the tensor computation stream, specifically, refers to a serial pipelined multiply-accumulate operation queue built by the main control processor at a 3.5GHz clock frequency in the cache. This queue consists of a 128-dimensional one-dimensional floating-point array and a 128x128x128 three-dimensional floating-point tensor. Feature components are sequentially read and written into the multiply-accumulate general-purpose registers via the system bus. Using this, the tensor orthogonal projection rule is applied to project the text features onto the spatiotemporal topological basis, calculating the contextual co-occurrence high-dimensional tensor as a high-dimensional feature cross-data array. Contextual co-occurrence high-dimensional tensor The calculation formula is ,in, For contextual co-occurrence high-dimensional tensors, The initial feature vector, For spatiotemporally constrained adjacency tensors, For tensor product operations, The pre-defined spatiotemporal topological attenuation factor is used to weave multidimensional physical environment constraints into the semantic feature space as structural features to establish nonlinear constraint boundaries.

[0042] Based on this, the distributed semantic retrieval system continuously captures discrete interaction behavior sequences for the object to be processed within the current sliding time window to adjust the dynamic feedback. When the system detects that the discrete interaction behavior sequence within the current sliding time window is empty, it triggers a spatiotemporal topological pseudo-inverse compensation mechanism. The system retrieves the historical association frequencies of each logical node in the global historical section under the physical spatiotemporal coordinate system and calculates the spatial kernel density distribution of the historical association frequencies to output the background context node density distribution, decomposing the spatiotemporal constraint adjacency tensor. The inverse of each non-zero singular value is multiplied by the transpose of the eigenvector matrix to construct a generalized pseudo-inverse matrix. As a compensation term, it is superimposed on the tensor computation flow to balance the input sparsity state, at which point the cold-start output tensor... The calculation formula is ,in, This is the context-coordinated high-dimensional tensor output under cold start conditions. The initial feature vector, For spatiotemporally constrained adjacency tensors, The preset constant adjustment factor, It is the generalized pseudo-inverse matrix of the spatiotemporal constrained adjacency tensor. For tensor product operations, This is the preset spatiotemporal topological decay factor.

[0043] To avoid the risk of computational divergence during high-frequency step transitions between empty and non-empty states, this invention sets up a temporal smoothing buffer with a length of 3 sliding window periods during the transition from an empty to a non-empty state. When the system detects that the discrete interaction behavior sequence has changed from empty to non-empty, it does not completely remove the additive compensation term of the generalized pseudo-inverse matrix. Instead, it uses a weight coefficient that decreases linearly with time to reduce the generalized pseudo-inverse matrix. At the same time, it assigns a dynamic weight that increases linearly with time to the newly generated intention bias transition matrix to ensure the numerical continuity of the feature transformation step size during mode transition. When the system detects that the discrete interaction behavior sequence within the current sliding time window is non-empty, it converts each historical interaction action in the discrete interaction behavior sequence into a state transition event in the state space, counts the state transition frequency between adjacent state transition events, normalizes the state transition frequency, and constructs an initial transition probability matrix.

[0044] When the variance of the discrete interaction behavior sequence exceeds the fluctuation baseline value, the discretized gradient of the discrete interaction behavior sequence is statistically analyzed to calculate the intent information entropy within the current sliding time window. The specific calculation logic is as follows: The main control processor extracts the difference in the frequency of discrete interaction actions between two adjacent sampling points within the current sliding time window to generate a discrete gradient array. The absolute value of each gradient component in this array is calculated, and each absolute value is divided by the sum of the absolute values ​​of all components to complete probability normalization, resulting in a floating-point sequence representing the probability distribution of intent variation. Then, each non-zero floating-point number in the sequence is traversed, and its base-2 logarithmic result is obtained by looking up the logarithmic mapping table preset in static memory. The logarithmic result is multiplied by the floating-point number itself and inverted. The product terms of the entire sequence are then summed, and finally, a dimensionless floating-point value between 0 and 8 is output to the main control register as the intent information entropy. When the safety truncation threshold is exceeded, elements in the initial transition probability matrix whose values ​​are greater than the high-frequency perturbation threshold are identified as high-frequency perturbation components. The gating unit forcibly clears the values ​​of the high-frequency perturbation components to zero in order to calculate the intention bias transition matrix and lock the current dynamic update step. The obtained intention bias transition matrix is ​​then compared with the spatiotemporal constraint adjacency tensor. In the dot product of each relation aspect matrix, due to the inconsistency in the initial order and dimension between the intention bias transition matrix and the relation aspect matrix during the specific alignment and dot product operations, this invention establishes a mapping index chain between the state space and entity nodes before performing the dot product. The specific white-box mapping logic is as follows: the main control processor traverses the two-dimensional integer mapping array from the action number to the entity node identifier in the knowledge graph, which is pre-stored in the configuration register, to establish a unique correspondence address between the discrete interactive action state and the candidate entity pair. It also allocates a continuous memory addressing offset for each legal migration path component. By reading the preset entity behavior mapping matrix, the state transition probability corresponding to the discrete interactive action is adaptively projected onto the candidate entity node pair in the static knowledge graph. For those not containing direct action mappings... The graph nodes are extended by padding the corresponding matrix elements with zeros, thereby transforming the intention bias transfer matrix into an alignment transformation matrix that is completely equal in order to the relational section matrix of the spatiotemporal constraint adjacency tensor in terms of algebraic structure. This ensures that the number of rows and columns of the two sets of matrices are completely consistent, achieving precise multiplication of the Hadamard product between feature points in the high-dimensional space. In the forward flow path, the feature transformation step size of each axis in the high-dimensional feature space is adjusted according to the dot product result to adjust the optimization weight distribution in the high-dimensional space. In this process, the specific logic for adjusting the feature transformation step size of each axis is as follows: the main control processor extracts the matrix trace or eigenvalue vector of the dot product result matrix projected onto each feature axis, normalizes and scales the eigenvalue corresponding to each axis, and maps it to a preset step size adjustment range, such as between 0.1 and 2.0.

[0045] When the dot product projection value on a specific axis increases, it indicates that the user's dynamic interaction intent has a higher focus on that feature axis. The system proportionally increases the feature transformation step size on that axis, allowing it to converge quickly in the forward flow; conversely, it proportionally decreases the feature transformation step size, thereby completing the precise adjustment of the optimization weight distribution and changing the initial feature vector. In the forward propagation network constructed based on the adjusted optimization weight distribution, cross-domain correlation features are extracted layer by layer through tensor computation flow. Under the constraint of updating without backpropagation gradient propagation, the output target semantic correlation feature vector is calculated in a feedforward manner. In multidimensional linear transformation and hierarchical matrix mapping, the feature vector is multiplied through multiple layers of forward concatenated matrices to achieve nonlinear orthogonal projection of the high-dimensional feature space and cross-domain feature extraction. The forward propagation network includes an input layer, hidden layers, and an output layer. The number of nodes in the input layer and the initial feature vector... The dimensions are consistent at 128; the hidden layer has 512 nodes and uses a linear rectified function as the activation function; the output layer has 128 nodes and outputs the target semantic association feature vector. In the forward feature mapping flow, the main control processor adjusts the axial feature transformation step size of the high-dimensional feature space based on the dot product result, and uses the adjusted optimization weight distribution as the weight modulation coefficient in the form of a diagonal matrix. The bottom-level array algebra reconstruction step is as follows: the main control processor directly uses the scaled floating-point numbers on each feature axis as the main diagonal elements and fills them into a 128x128 dimensional diagonal weight array. In the process, the hardware multiply-accumulate matrix operator is called to perform matrix dot product between the diagonal weight array and the 512-by-128 dimensional original connection matrix preloaded in the static dynamic storage area during network initialization. The new matrix data stream obtained by multiplication is written to the weight addressing register, thereby acting on the hidden layer connection weight matrix. The hidden layer connection weight matrix is ​​updated by calculating the product of the basic initial matrix and the diagonal matrix, changing the algebraic transformation path inside the forward propagation network, constraining the multiply-accumulate calculation unit of the bottom dual-processor, and realizing the feature flow of feature vectors without the constraint of backward gradient propagation update.

[0046] Based on the discretized algebraic computation of forward topological flow and interactive adaptive modulation, the distributed semantic retrieval system calculates the cosine similarity between the target semantic association feature vector and the feature vectors of each candidate entity in the static semantic database. When the cosine similarity is greater than the safe association threshold, it determines that there is a semantic association between the object to be processed and the corresponding candidate entity and outputs the semantic association discovery result. Thus, without relying on the backpropagation of the fully connected layer of the deep neural network, the retrieval latency of the data processing flow is stably controlled within 50ms. Since the spatiotemporal constraints are woven into the semantic vector space as structural features, false semantic associations that are similar in content but spatiotemporally unrelated in the multi-source heterogeneous data flow path are suppressed in situ during the data flow stage. Moreover, in the cold start condition with zero interaction or in the strong noise interference environment accompanied by high frequency mutation and misoperation, the structural convergence and operational stability of the feature matrix calculation are maintained by generalized pseudo-inverse topology compensation and information entropy gating truncation, maintaining the balance between real-time perception of system environmental features and ultra-low latency computation response.

[0047] Example 2: When the system faces a high-frequency influx of heterogeneous text records from multiple sources and a gradient change in the intensity of sudden regional semantic variations, the method claimed in this invention is validated on a distributed semantic retrieval test platform equipped with dual processors, each with 32 TFLOPS of floating-point computing power and 128 GB of system memory. The platform uses a publicly available general multidimensional text retrieval benchmark dataset as the underlying data source. To simulate industrial electromagnetic environments and feature distortions caused by network transmission jitter, Gaussian white noise with a signal-to-noise ratio of 20 dB and power frequency interference harmonics with a frequency of 50 Hz are actively superimposed on the input raw semantic stream. During operation, the test platform faces multidimensional performance testing of the feature image, among which it is necessary to determine the key control... The current sliding time window length is determined by factors including the arrival frequency of the original text sequence and the burst cycle of historical intent. The technical consideration is to achieve a balance between the sensitivity of dynamic intent capture and system memory and computing load. The decision rule is based on the fact that when the arrival frequency varies in the range of 1000Hz to 5000Hz, if the burst cycle of historical intent is shortened, the current sliding time window length shrinks towards the lower limit of its range to avoid capture failure, and vice versa. According to this rule, for the typical working condition of 3000 records per second and a burst cycle of 10s corresponding to this experiment, the current sliding time window length is set to 5s as a non-limiting engineering example.

[0048] During the experimental verification phase, a control group consisting of a conventional pre-trained language model vector retrieval without injected spatiotemporal constraints and dynamic interaction features was set up, along with the present invention sample group consisting of the method claimed in this invention. Semantic mutation intensity was divided into three test gradients: low, medium, and high. When in the low mutation intensity gradient (i.e., when the number of mutated entities per hour is less than 5), the initial semantic retrieval accuracy of the control group was 85.3%, while the present invention sample group constructed a spatiotemporal constraint adjacency tensor containing multidimensional floating-point numbers based on collected spatiotemporal environment data. And calculate the initial feature vector. If the discrete interaction sequence within the current sliding time window is empty, the system automatically triggers the spatiotemporal topological pseudo-inverse compensation mechanism to adjust the spatiotemporal constraint adjacency tensor. Expand the singular value decomposition and calculate the generalized pseudoinverse matrix. The cold start output tensor is calculated using the formula. The specific calculation formula is expressed as follows: ,in, This is the context-coordinated high-dimensional tensor output under cold start conditions. The initial feature vector, For spatiotemporally constrained adjacency tensors, This is a preset constant adjustment factor, set to 0.15 in this example. It is the generalized pseudo-inverse matrix of the spatiotemporal constrained adjacency tensor. For tensor product operations, The preset spatiotemporal topological decay factor is set to 1.2, and the generalized pseudo-inverse matrix is ​​used. As a physical topology compensation term, the in-situ balancing of the input sparse state ensures that the final association accuracy of the sample group of this invention is stable at 94.6%, and the retrieval response latency is kept at a low level of 32.5ms due to the elimination of backpropagation updates in the deep neural network.

[0049] As the problem severity escalates to a medium mutation intensity gradient, meaning the number of mutated entities per hour is between 10 and 20 zones, and the discrete interaction behavior sequence is not empty, high-frequency mutation errors occur in the interaction sequence due to injected Gaussian white noise interference. The variance of the interaction sequence reaches 18.4, exceeding the preset fluctuation baseline value of 10.0. The system automatically activates the intent information entropy gating threshold constraint rule, statistically analyzes the discretized gradient of the interaction behavior sequence, and calculates the intent information entropy within the current sliding time window. The measured value was 2.45. Since this measured value exceeded the preset safety cutoff threshold of 1.80, the gating unit forcibly set the components in the intent bias transition matrix that were greater than the high-frequency disturbance threshold of 0.50 to zero to eliminate interference noise. The internal optimization weight distribution of the calculated feature tensor was dynamically reorganized, so that the semantic association discovery accuracy of the sample group under the medium gradient remained at 93.1%, and the retrieval latency was 36.8ms. In contrast, the accuracy of the control group deteriorated sharply to 62.4% due to feature routing deviation caused by noise, and the retrieval latency increased to 142.3ms. When the severity of the problem further increased to the high mutation intensity gradient, that is, the extreme condition where the number of mutated entities per hour reached more than 50, the intent information entropy of the sample group of the present invention... Increasing it to 4.12, the feature matrix calculation exhibits a performance inflection point with nonlinear effects. If the safety truncation threshold is arbitrarily widened to 4.50, thus exceeding the upper limit of the protection range defined by this invention, the data shows that the gating unit cannot truncate the excessive disturbance components, causing the optimization weight distribution during forward feedforward calculation to diverge. The retrieval response delay deteriorates sharply to 185.6ms and the accuracy drops to 55.2%, indicating that the safety truncation threshold range defined by this invention belongs to the parameter range that hinders calculation divergence and maintains the physical boundary of the multidimensional environment constraint.

[0050] Example 3: This example combines Figures 1 to 2 The method for semantic association discovery based on spatiotemporal-relationship-interaction joint analysis is explained, such as... Figure 1As shown, step S1 obtains the initial feature vector of the object to be processed and the spatiotemporal environment data corresponding to the object. Step S2 combines the spatiotemporal environment data and the logical link relationship of the object to be processed to construct a spatiotemporal constraint adjacency tensor. Then, step S3 captures the discrete interaction behavior sequence for the object to be processed within the current sliding time window. When the discrete interaction behavior sequence is empty, the background context node density distribution under the global history section is read to calculate the generalized pseudo-inverse matrix of the spatiotemporal constraint adjacency tensor as a compensation term and superimposed on the tensor mapping stream to balance the input sparsity state and maintain the structural contraction. The convergence process is as follows: Step S4 is executed to convert the discrete interaction behavior sequence into an intent bias transition matrix when the discrete interaction behavior sequence is not empty. When the variance of the discrete interaction behavior sequence exceeds the fluctuation benchmark value, the discretized gradient of the discrete interaction behavior sequence is calculated to calculate the intent information entropy within the current sliding time window. When the intent information entropy exceeds the safety truncation threshold, the high-frequency perturbation components in the intent bias transition matrix are truncated to zero. The obtained intent bias transition matrix is ​​then multiplied with the spatiotemporal constraint adjacency tensor to adjust the optimization weight distribution in the high-dimensional space, thereby calculating the target semantic association feature vector in a forward feedforward manner.

[0051] like Figure 2 As shown, the top data bus contains heterogeneous text records input from the data bus to the cache memory and passed down to the distributed logical nodes. Inside the distributed logical nodes, the front-end cache memory on the left initializes and loads the baseline parameters for the queue to be processed, and passes the parameters to the dual-processor that is closed at a main frequency of 3500rpm. At the same time, the dynamic storage area on the right scans the timestamps attached to each background context data record in the dynamic storage area, and passes the data to the master processor that reads the computing core load status of the current logical node from the system control bus at a period of 100ms. Both the dual-processor and the master processor are connected down to the system control bus to read the computing core load status of the current logical node and the spatial parameters of the dynamic storage area. Finally, the system control bus is connected down to the static semantic library at the bottom to calculate the cosine similarity between the target semantic association feature vector and the feature vectors of each candidate entity in the static semantic library.

[0052] Example 4: When a distributed semantic retrieval system faces fluctuations in input data stream density caused by uneven hardware processing capabilities of distributed logical nodes, the instantaneous arrival frequency of heterogeneous text records input from the data bus to the cache memory frequently jumps between 1000Hz and 5000Hz. Since the fixed parameter determination boundary cannot match the dynamically changing processor cycle overhead, the feature vector queue to be processed in the cache memory is prone to read / write pointer overflow, which in turn causes memory addressing conflicts and divergence in relation aspect matrix calculation, resulting in the overall retrieval processing latency of the system deviating from the preset target of less than 50ms.

[0053] To constrain transient fluctuations in the data bus, the master processor in the distributed semantic retrieval system reads the current logical node's computational core load status and dynamic storage area space parameters from the system control bus every 100ms. It then dynamically fine-tunes the safety truncation threshold based on resource load. The safety truncation threshold is expressed as... ,in,

[0054] To safely truncate the threshold, The basic cutoff threshold is fixed at 1.80. This is the resource load adjustment coefficient, and its value is set to 0.05. This is the normalized frequency variance of text records input from the data bus within the current sliding time window of 5 seconds. Simultaneously, to eliminate the risk of baseline drift caused by the accumulation of old historical data in the background context node density distribution under the global historical section, the main control processor periodically scans the timestamps attached to each background context data record in the dynamic storage area. When the residence time indicated by the timestamp exceeds 24 hours, the corresponding old historical data record is erased from the physical memory address. This automatically initiates the spatiotemporal constraint adjacency tensor under cold start conditions without dynamic interactive data input. The topological baseline is reconstructed, and the generalized pseudo-inverse matrix is ​​recalculated using a dual-processor architecture. To maintain the continuous causal relationship in the mapping of data feature vectors.

[0055] Under the synergistic effect of the aforementioned time-sensitive replacement loop and dynamic threshold adjustment, the input high-frequency mutation interference records are truncated in real time at the gating unit. The queue length of the heterogeneous data stream in the dynamic storage area is always maintained within the preset 512 data points. During this process, the specific control logic for maintaining the queue length not exceeding 512 data points is as follows: when the main control processor detects that the influx of records to be processed in the high-speed cache is too fast, causing the current queue length to reach the critical upper limit of 512 data points, the system automatically starts a rolling over mechanism based on the first-in-first-out strategy, removing the oldest data records moved into the queue and releasing their corresponding memory addresses, prioritizing the real-time writing of the latest heterogeneous text records, thereby effectively preventing read errors. Write pointer overflow and memory addressing conflicts were tested. Test data showed that when the frequency of input records arriving jumped from 1000Hz to 5000Hz and was accompanied by Gaussian white noise, the distributed semantic retrieval system maintained a data processing flow retrieval latency between 34.2ms and 41.5ms due to the nonlinear correction that matched the resource load by the feature vector transformation step size. No read / write stalls occurred due to computational divergence. Furthermore, the dynamic reconstruction of the spatiotemporal topology balanced the data sparsity, and the routing of heterogeneous texts with similar content but unrelated spatiotemporal backgrounds was accurately corrected. The error matching rate of semantic association discovery results dropped to below 0.85%, maintaining a balance between real-time perception of system environmental features and ultra-low latency computation response.

[0056] Example 5: When the system faces the initial deployment of distributed nodes or changes in computing card hardware parameters, the main control processor initiates a pre-calibration procedure to determine the initial parameter baseline before accessing the real heterogeneous text recording stream. The main control processor calls the test high-speed cache to import a reference corpus stream containing 10,000 standard specification texts. Under stable conditions without external spatiotemporal interference, the language model feeds forward to parse and extract standard feature vectors. When the processing overhead allocated by the control bus is at a constant 20% load baseline, the vector similarity variance sequence within 50 time panes is continuously collected. The basic truncation threshold is established by calculating the mean of the variance sequence. The initial solution point was determined, and a simulated load flow with a throughput step change from 1000Hz to 5000Hz was artificially applied to the test platform to induce read / write pointer displacement in the feature queue to be processed. The frequency variance response slope when the retrieval delay crossed the critical state of 50ms was recorded. The frequency variance response slope was used to inversely balance the load disturbance of the system memory, thereby calculating the resource load adjustment coefficient. The convergence point.

[0057] Applying the aforementioned pre-calibration procedure to the hardware environment corresponding to this test group, and after decomposition by a dual-processor at a clock frequency of 3500 rpm, the static convergence center value of the information entropy of the reference corpus stream in the standard forward propagation path was measured to be 1.80. This value was determined as the basic truncation threshold. The input quantity was determined, and when the simulated load flow reached a step peak of 5000Hz and intruded into the extreme value of memory addressing, the proportionality coefficient of the change in processing latency to the change in frequency variance was measured to be 0.05. This proportionality coefficient is used as the resource load adjustment coefficient. The fixed parameters are written into the configuration address of the main control memory. The selection of the fluctuation baseline value of 10.0 and the high-frequency disturbance threshold of 0.50 is determined by collecting the disturbance variance sequence of 1000 normal user misoperations. When the variance obtained by the test is lower than the lower limit boundary value of 10.0, it indicates that the fluctuation of the current interaction sequence is a normal input jitter and there is no need to trigger the gating cutoff mechanism. If it is forcibly triggered, it will lead to the loss of useful semantic features. When the amplitude of the high-frequency disturbance component is lower than 0.50, it does not pose a threat to the overall convergence of the state transition probability. Only when the transition probability of a specific element exceeds the upper limit of the high-frequency disturbance threshold of 0.50 is it judged as a sudden misoperation and hard-cleared to zero. In this way, while suppressing abnormal mutation noise, it avoids excessive smoothing of normal dynamic interaction intentions. After the waiting queue in the front-end high-speed cache memory completes the initialization and loading of the above baseline parameters, the entire distributed semantic retrieval system enters a stable monitoring state of continuous operation.

[0058] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A semantic association discovery method based on spatiotemporal-relational-interaction joint approach, characterized in that, Includes the following steps: Step S1: Obtain the initial feature vector of the object to be processed and the spatiotemporal environment data corresponding to the object to be processed; Step S2: Combine the spatiotemporal environment data and the logical link relationship of the object to be processed to construct a spatiotemporal constrained adjacency tensor; Step S3: Capture the discrete interaction behavior sequence for the object to be processed within the current sliding time window; when the discrete interaction behavior sequence is empty, read the background context node density distribution under the global history section, calculate the generalized pseudo-inverse matrix of the spatiotemporal constraint adjacency tensor as a compensation term and superimpose it into the tensor mapping flow to balance the input sparse state and maintain structural convergence. Step S4: When the discrete interaction behavior sequence is not empty, convert the discrete interaction behavior sequence into an intent bias transition matrix. When the variance of the discrete interaction behavior sequence exceeds the fluctuation baseline value, the discretized gradient of the discrete interaction behavior sequence is statistically analyzed to calculate the intention information entropy within the current sliding time window; When the intent information entropy exceeds the safety truncation threshold, the high-frequency perturbation components in the intent bias transition matrix are truncated to zero. The resulting intent bias transition matrix is ​​then multiplied by the spatiotemporal constraint adjacency tensor to adjust the optimization weight distribution in the high-dimensional space, thereby calculating the target semantic association feature vector in a forward-feedback manner.

2. The semantic association discovery method based on spatiotemporal-relationship-interaction joint method according to claim 1, characterized in that, In step S3, the generalized pseudo-inverse matrix is ​​superimposed as a compensation term onto the tensor mapping flow to maintain structural convergence. This includes the following sub-steps: Step S31, the generalized pseudo-inverse matrix is ​​weighted using a constant adjustment factor, and the weighted generalized pseudo-inverse matrix is ​​added to the spatiotemporal constraint adjacency tensor to obtain the topology correction tensor; Step S32, the tensor product between the initial eigenvector and the topology correction tensor is calculated, and the tensor product result is normalized and scaled using a density scaling factor to eliminate sparsity divergence, and is output as the tensor mapping flow after the compensation term is superimposed.

3. The semantic association discovery method based on spatiotemporal-relationship-interaction joint method according to claim 1, characterized in that, Step S2 involves constructing a spatiotemporal constrained adjacency tensor by combining spatiotemporal environment data and the logical link relationship of the object to be processed. This includes the following sub-steps: Step S21: Parse the spatiotemporal environment data to extract time feature intervals and geospatial coordinates; Step S22: Calculate the time-sensitivity deviation between the time feature intervals and the sudden periodicity of the object to be processed, and simultaneously calculate the topological distance between the geospatial coordinates and known entity nodes; Step S23: Based on the time-sensitivity deviation and topological distance, adjust the entity association strength in the static knowledge graph through a collaborative decay method to generate the spatiotemporal constrained adjacency tensor.

4. The semantic association discovery method based on spatiotemporal-relationship-interaction joint method according to claim 1, characterized in that, Step S4 converts the discrete interaction behavior sequence into an intention bias transition matrix and truncates and sets the high-frequency disturbance components to zero, including the following sub-steps: Step S41, maps each historical interaction action in the discrete interaction behavior sequence to a state transition event in the state space, and counts the state transition frequency between adjacent state transition events. Step S42: Normalize the state transition frequency to construct an initial transition probability matrix; Step S43: When the intent information entropy exceeds the security truncation threshold, identify the elements in the initial transition probability matrix whose values ​​are greater than the high-frequency perturbation threshold as high-frequency perturbation components; Step S44: Forcibly clear the values ​​of the high-frequency perturbation components to generate an intent bias transition matrix.

5. The semantic association discovery method based on spatiotemporal-relationship-interaction joint method according to claim 1, characterized in that, In step S4, the obtained intention bias transfer matrix is ​​multiplied by the spatiotemporal constraint adjacency tensor to adjust the optimization weight distribution in the high-dimensional space, including the following sub-steps: Step S45, the intention bias transfer matrix is ​​multiplied by each relation section matrix in the spatiotemporal constraint adjacency tensor to utilize the user's dynamic interactive intention to weight modulate the spatiotemporal constraint relationship. Step S46: In the forward flow path, adjust the feature mapping step size of each axis of the high-dimensional feature space according to the dot product result to complete the adjustment of the optimization weight distribution.

6. The semantic association discovery method based on spatiotemporal-relationship-interaction joint method according to claim 1, characterized in that, Step S4 calculates the target semantic association feature vector using a feedforward approach, including the following sub-steps: Step S47, inputting the initial feature vector into the feedforward network constructed based on the adjusted optimization weight distribution; Step S48, extracting cross-domain association features layer by layer through tensor mapping flow, and directly calculating the output target semantic association feature vector without the constraint of updating through backward gradient propagation.

7. The semantic association discovery method based on spatiotemporal-relationship-interaction joint method according to claim 1, characterized in that, Step S3 involves reading the background context node density distribution under the global historical section and calculating the generalized pseudo-inverse matrix, including the following sub-steps: Step S33: Extract the historical association frequency of each logical node in the physical spatiotemporal coordinate system from the global historical section; Step S34: Calculate the spatial kernel density distribution of the historical association frequency to generate the background context node density distribution; Step S35: Perform singular value decomposition on the spatiotemporal constraint adjacency tensor, and multiply the inverse of the non-zero singular values ​​by the transpose of the eigenvector matrix to construct the generalized pseudo-inverse matrix.

8. The semantic association discovery method based on spatiotemporal-relationship-interaction joint method according to claim 1, characterized in that, Step S1 obtains the initial feature vector of the object to be processed, and the declaration includes the following sub-steps: Step S11, use the trained language model to semantically parse the text content of the object to be processed and extract the original semantic feature vector; Step S12, reduce the dimensionality and normalize the original semantic feature vector to generate the initial feature vector.

9. The semantic association discovery method based on spatiotemporal-relationship-interaction joint method according to claim 1, characterized in that, It also includes the following post-processing steps: Step S5, calculate the cosine similarity between the target semantic association feature vector and the feature vectors of each candidate entity in the static semantic library; Step S6, when the cosine similarity is greater than the security association threshold, determine that there is a semantic association between the object to be processed and the corresponding candidate entity, and output the semantic association discovery result.

Images