An artificial intelligence-based multi-level context intelligent construction method applied to a text scene

Abstract

The application relates to the technical field of multi-level context intelligent construction based on artificial intelligence, and discloses a multi-level context intelligent construction method based on artificial intelligence applied to a text scene. A complete context intelligent modeling system is constructed, which covers processes such as normalized feature construction, adaptive spatial topology establishment, prototype state division, single-point and edge potential calculation, time transition modeling, energy function construction and improved ICM reasoning. The method starts from original context data and finally realizes state sequence recognition of nodes changing with time, and is strong in systematicness. Through fusion of three types of potential energy of nodes, edges and time, and cooperation of a point-by-point optimization reasoning mechanism, the recognition precision and operation efficiency are improved. The method has good adaptability, interpretability and wide application potential, and is suitable for context perception tasks in multi-source, multi-node and multi-time scene.

Description

Technical Field

[0001] This invention relates to the field of multi-level context intelligent construction technology based on artificial intelligence, specifically to a multi-level context intelligent construction method based on artificial intelligence applied to text scenarios. Background Technology

[0002] With the rapid development of cloud computing, big data, and deep learning, intelligent perception and efficient organization of massive, multi-source, and heterogeneous contextual information has become an important research direction in the field of artificial intelligence. In practical applications, downstream tasks (such as accurate recommendation, intelligent decision-making, and visualization analysis) often rely on a multi-level and multi-perspective understanding of the contextual environment.

[0003] First, traditional context representation methods often rely on pre-built static graphs or rule-based templates, typically capturing only relationships between a single level or a few attributes, making it difficult to reflect the dynamic evolution of context instances at different times or from different perspectives. Even when using knowledge graphs, they are often based on manually maintained nodes and edges, unable to handle real-time changes in node features and the self-organizing topological structures among massive numbers of nodes. Second, existing modeling methods for temporal information, such as Markov chains and Hidden Markov Models, while capable of characterizing the state transitions of a single node in the time dimension, are usually limited to chain structures, ignoring the complex spatial relationships that may exist between different nodes at the same time. Some spatial modeling methods based on Markov random fields are mostly used for image segmentation and annotation, focusing on spatial smoothness constraints in static scenes, and rarely involve in-depth research combining them with temporal evolution. Third, regarding state discretization, common clustering algorithms (such as K-means, spectral clustering, etc.) require pre-setting the number of clusters and are sensitive to initial centers, making it difficult to adapt to dynamic changes in the number of context instances; furthermore, these methods are mostly based on global statistical features for partitioning, making it difficult to account for the volatility of feature distribution at different times. When the number of clusters is fixed while the size of context instances increases, state partitioning can easily become too fine or too coarse, affecting the accuracy and efficiency of subsequent inference. Furthermore, while unified energy optimization frameworks based on potential functions have achieved some success in image processing, annotation, and knowledge reasoning, the design of potential functions often relies on experience or extensive hyperparameter tuning, and mostly considers only a single type of potential (such as node potential or edge potential), lacking a complete scheme that unifies nodes, spatial edges, and temporal transitions into a single energy model. Even those few studies that attempt to construct spatiotemporal joint energy functions typically optimize them piecewise, making it difficult to truly achieve globally consistent optimal state allocation.

[0004] Therefore, this case aims to propose a multi-level context intelligent construction method based on artificial intelligence for text scenarios. It formalizes the context modeling problem into a matching problem between nodes and states in a multi-dimensional space, and uses a potential energy mechanism and a variant ICM algorithm to iteratively find the optimal state sequence in the three-dimensional space of "node-time-state", thereby achieving high-precision and high-resolution context recognition and prediction. Summary of the Invention

[0005] This invention provides an AI-based multi-level context intelligent construction method for text scenarios, which helps to solve the problems mentioned in the background art.

[0006] This invention provides the following technical solution: a multi-level context intelligent construction method based on artificial intelligence applied to text scenarios, comprising:

[0007] Collect raw context metric data and perform normalization processing to construct normalized feature vectors of context instance nodes at each time step;

[0008] The original contextual measurement data is the keyword frequency;

[0009] Based on the normalized context feature vector, the spatial topology is adaptively constructed at each time step;

[0010] The number of prototype states is adaptively divided by binary clustering based on the main variance dimension of the context features.

[0011] Based on the distance between the normalized feature vector and various prototype vectors, the single-point potential energy of the node in a specific state at a specific time is calculated.

[0012] Calculate the potential energy of spatial connection edges based on the feature distance and state label between nodes at the current time.

[0013] Set a penalty term for the state change of the same node between adjacent time points;

[0014] A global energy function is constructed based on the single-point potential energy, the potential energy of the spatial connecting edges, and the penalty term, and minimizing this energy function is taken as the global optimal objective.

[0015] An improved iterative conditional pattern algorithm is used for point-by-point convergence inference to obtain the optimal state sequence of each node over time.

[0016] Optionally, the step of collecting the original context metric data and performing normalization processing to construct the normalized feature vector of the context instance node at each time step specifically includes:

[0017] Let the total set of context instance nodes be ;in, For the first One context instance node; This represents the total number of context instance nodes.

[0018] Let the complete set of time indices be ;in, For time indexing;

[0019] Get Nodes At any moment The Original Measure ;in, For measuring latitude index; This represents the total number of original latitude measurements.

[0020] Calculate the normalized metric ;

[0021] Construct nodes At any moment Normalized eigenvectors .

[0022] Optionally, the adaptive construction of the spatial topology structure based on the normalized context feature vector at each time step specifically includes:

[0023] compute nodes and At any moment Feature distance Specifically:

[0024] ;in, It is a Euclidean norm;

[0025] Set the global distance threshold to ;

[0026] in, To find the midvalue function of a set;

[0027] Construction time The set of undirected edges , .

[0028] Optionally, the step of adaptively dividing the number of prototype states by binary clustering based on the main variance dimension of the context features specifically includes:

[0029] Set the total number of discrete states to ;in, To round down;

[0030] Perform the following steps to perform binary clustering based on the main variance dimension:

[0031] S110, Initialize the family set and the initial family containing all eigenvectors :

[0032] , ;

[0033] S120, iterative splitting, until... :

[0034] S121, For each clan ,calculate , ;in, For the clan In the middle, the first The variance of all sample values ​​in the dimension; This is the current set of vectors with splitting. For the clan A normalized eigenvector, that is, the value of a node at a certain time. ; For this vector In the The components in the dimension correspond to the normalized 1st dimension. Item measurement ;

[0035] Elected The largest clan, denoted as ;

[0036] S122. Calculate the principal variance dimension: ;

[0037] S123. Calculate the median of this dimension. ;

[0038] S124, split into , ;in, A subset whose values ​​on the principal variance dimension do not exceed the median; This refers to the subset whose values ​​on the principal variance dimension exceed the median.

[0039] S125, Update ;

[0040] Calculate the first Prototype vector of class state , ;in, The first cluster obtained after variance-based binary clustering Individuals.

[0041] Optionally, the calculation of the single-point potential energy of a node at a specific moment, based on the distance between the normalized feature vector and various prototype vectors, when the node is assigned a specific state, specifically includes:

[0042] compute nodes At any moment State assigned single-point potential Specifically:

[0043] , ;in, This represents the global maximum distance.

[0044] Optionally, the step of calculating the potential energy of spatial connection edges based on the feature distances and state labels between nodes at the current time specifically includes:

[0045] Calculate normalized distance ;

[0046] Set edge At any moment Spatial potential ;in, For nodes The status label assigned at the current moment; For nodes The state label assigned at the current moment; both labels take values ​​from , corresponding to different prototype vector clusters.

[0047] Optionally, the penalty term for state changes of the same node between adjacent time points specifically includes:

[0048] Set the penalty for state transitions of the same node at adjacent time steps as follows: Specifically:

[0049] ;in, For a certain node at time... Status labels; For the same node at time Status labels.

[0050] Optionally, the step of constructing a global energy function based on the single-point potential energy, the potential energy of the spatial connecting edges, and the penalty term, and using minimizing this energy function as the global optimal objective, specifically includes:

[0051] Calculate global energy Specifically:

[0052] ;in, For nodes At any moment The final optimal state label;

[0053] , ;in, This is the optimal state allocation scheme that includes all nodes and all time points.

[0054] Optionally, the step of employing an improved iterative conditional pattern algorithm for point-by-point convergence inference to obtain the optimal state sequence of each node over time specifically includes:

[0055] initialization , ;

[0056] Update point by point until all No longer changing:

[0057] S201, Enumerate each pair ;

[0058] S202, Order

[0059] ;

[0060] S203, Boundary Time or When, omit non-existent elements. item;

[0061] After convergence, the final state sequence of each node over time is obtained. .

[0062] The present invention has the following beneficial effects:

[0063] 1. Starting from the original contextual measurement data, this approach fully considers the differences in dimensions and scales among different indicators. Through normalization, various heterogeneous data are uniformly mapped to comparable intervals (e.g., [0,1]), thereby constructing feature vectors with unified dimensions and scale independence. This approach not only improves the robustness of subsequent algorithms (such as distance calculation and cluster analysis) but also lays the foundation for the generalization of algorithms in different application scenarios. In reality, especially when processing multi-source data such as text semantics (e.g., word frequency) and sensor networks (e.g., temperature, voltage, noise), the inconsistency in the scale of features with different dimensions directly affects similarity calculation and model stability.

[0064] 2. This invention overcomes the limitations of fixed adjacency rules or static graph models by proposing a "time-independent adaptive undirected graph" construction mechanism. Specifically, the algorithm adaptively sets a distance threshold controlled by the median based on the Euclidean distance between normalized feature vectors at the current time, thereby generating an edge set that "retains only key similarities." This approach avoids topological shifts caused by outliers and improves the representation efficiency and generalization ability of the graph structure. In real-world applications, such as building user profiles for social networks and sensor collaborative modeling, static connections or preset thresholds can easily lead to underfitting or overfitting, while the method of this invention can adjust the structure in real time and has good dynamic adaptability.

[0065] 3. The "variance-driven binary clustering" mechanism adaptively determines the size of the state space, making it a non-parametric partitioning method that combines interpretability and efficiency. Its key advantage lies in dynamically identifying the most significantly different dimensions in the data and using the median as the split point to achieve effective data splitting, ultimately constructing multi-class, highly discriminative state prototypes.

[0066] 4. Potential energy modeling is one of the core innovations of this invention. It comprehensively quantifies the system's "information stability and structural rationality" from three dimensions: node potential (single-point distance), spatial edge potential (consistency of adjacent states), and temporal transition potential (state smoothness). This idea draws on Markov random fields and energy minimization mechanisms in graph models, but further enhances the spatiotemporal joint modeling capability. In particular, by introducing the matching potential between state labels and prototype vectors, a spatial penalty mechanism based on similarity decay, and a time smoothing term with fixed weights, it not only effectively improves classification robustness but also achieves a natural simulation of the continuous evolution of context states.

[0067] 5. The entire state recognition problem is formalized as a global energy function minimization problem, embodying the design philosophy of "from local constraints to global optimality." This energy function organically integrates node potential, spatial potential, and temporal potential, with adjustable weights for each component, offering good interpretability and scalability. This global perspective ensures that node state determination no longer relies solely on its own characteristics but comprehensively considers similarity to neighboring nodes and historical state change trends, thus obtaining more stable and reliable recognition results. Compared to traditional independent classifier models or linear sliding window strategies, this energy optimization strategy is better suited to the continuous, dependent, and uncertain characteristics of complex real-world scenarios.

[0068] 6. This method employs a variant of the ICM inference algorithm. Compared to the global synchronous update approach of traditional ICM, this method uses a point-by-point convergence update strategy, updating only the state of a single node-time combination at a time and evaluating its impact on global energy in real time, thus converging to a local optimum more efficiently. This mechanism significantly improves inference speed and practicality, making it particularly suitable for scenarios with large-scale nodes and long-term time-series data. Furthermore, by appropriately setting boundary conditions and convergence criteria, the algorithm is ensured not to fall into unstable loops. Compared to traditional time-series models such as HMM and CRF, this inference method requires no training phase, has lower computational complexity, and is more practical in unlabeled supervised scenarios. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0070] 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.

[0071] Example, refer to Figure 1 A multi-level context intelligent construction method based on artificial intelligence for text scenarios, comprising:

[0072] Collect raw context metric data and perform normalization processing to construct normalized feature vectors of context instance nodes at each time step;

[0073] The original contextual measurement data is the keyword frequency;

[0074] Based on the normalized context feature vector, the spatial topology is adaptively constructed at each time step;

[0075] The number of prototype states is adaptively divided by binary clustering based on the main variance dimension of the context features.

[0076] Based on the distance between the normalized feature vector and various prototype vectors, the single-point potential energy of the node in a specific state at a specific time is calculated.

[0077] Calculate the potential energy of spatial connection edges based on the feature distance and state label between nodes at the current time.

[0078] Set a penalty term for the state change of the same node between adjacent time points;

[0079] A global energy function is constructed based on the single-point potential energy, the potential energy of the spatial connecting edges, and the penalty term, and minimizing this energy function is taken as the global optimal objective.

[0080] An improved iterative conditional pattern algorithm is used for point-by-point convergence inference to obtain the optimal state sequence of each node over time.

[0081] By designing an overall process consisting of eight clearly defined steps, a complete intelligent modeling system is realized, starting from raw context data and ultimately generating a sequence of context states. These steps include normalized feature construction, adaptive topology establishment, prototype state partitioning, potential energy calculation, multi-source information fusion (nodes, edges, time), and energy function construction and optimization reasoning. This method addresses the problems of traditional context modeling, such as insufficient ability to handle complex context structures, lack of modeling for spatiotemporal dependencies, coarse state classification, and difficulty in reasoning for the globally optimal state. On the one hand, it allows the automatic establishment of logical and evolutionary relationships between contexts in multi-source, multi-node, and multi-time-stack scenarios, enhancing the model's adaptability and interpretability. On the other hand, the use of multi-potential term joint optimization and improved reasoning algorithms not only ensures model computational efficiency but also improves final recognition accuracy. The overall solution possesses advantages such as reasonable structure, strong adaptability, and wide applicability, providing effective support for intelligent context perception in complex scenarios.

[0082] The process of collecting raw context metric data and performing normalization to construct normalized feature vectors for context instance nodes at each time step specifically includes:

[0083] Let the total set of context instance nodes be ;in, For the first One context instance node; This represents the total number of context instance nodes.

[0084] Let the complete set of time indices be ;in, For time indexing;

[0085] Clearly define the analysis object and time range to build the index foundation for all subsequent operations;

[0086] Get Nodes At any moment The Original Measure If it is a text-based scenario, then For the first The frequency of each keyword; if it is a sensing scenario, then... For the first The readings of sensor number 1; among which... For measuring latitude index; The original metric is the total number of latitudes; obtain each node. At any moment The original observation data provides raw materials for subsequent feature construction;

[0087] Calculate the normalized metric Eliminate the dimensional differences between different measures and unify all measures to [a certain standard]. The interval ensures comparability in subsequent distance calculations;

[0088] Construct nodes At any moment Normalized eigenvectors The multidimensional normalized metric is integrated into a vector form, which facilitates subsequent matrix operations and distance evaluation.

[0089] By collecting and normalizing the original contextual metric data, the problems of inconsistent scales across different feature dimensions and uneven data quality were resolved. Normalization unified the units of measurement, making subsequent distance calculations and vector comparisons comparable and avoiding clustering errors or misclassifications caused by "unit interference." Simultaneously, integrating features from multiple dimensions into normalized feature vectors not only improved data organization efficiency but also laid a solid mathematical foundation for subsequent graph structure construction and cluster analysis. Compared to traditional methods of directly using raw values ​​or empirical standardization, this step provides uniformity, stability, and repeatability, effectively enhancing the model's generalization ability across domains or multi-source heterogeneous data. The adaptive construction of the spatial topology structure based on the normalized contextual feature vectors at each time step specifically includes:

[0090] compute nodes and At any moment Feature distance Specifically:

[0091] ;in, It is a Euclidean norm; it quantifies the feature difference between any two nodes at the same time, serving as a fundamental metric for edge generation.

[0092] Set the global distance threshold to ;

[0093] in, To find the median function of the set, a robust threshold is selected using the median, which automatically adapts to the data distribution and avoids the influence of outliers.

[0094] Construction time The set of undirected edges , Generate the undirected graph topology at each time step, representing the significant similarity relationships between nodes.

[0095] By constructing a spatial topology structure based on normalized features, this method solves the problem of modeling node relationships in multi-node systems. Its innovation lies in adaptively constructing an undirected edge set based on feature similarity and setting a distance threshold using the median to eliminate outlier data. This approach breaks through the limitations of traditional graph construction methods that rely on static or fully connected methods, dynamically adapting to distribution characteristics under different scenarios and realistically reflecting the local correlations between nodes. This dynamic and adjustable graph structure greatly enhances the realism of context representation and the ability to recognize local patterns, enabling the model to not only see "points" but also reasonably see "areas," that is, to reasonably model adjacency relationships, providing important support for subsequent state partitioning and reasoning.

[0096] The method of adaptively dividing the number of prototype states by binary clustering based on the main variance dimension of context features specifically includes:

[0097] Set the total number of discrete states to ;in, To round down; the number of discrete states is dynamically set to ensure that the state space expands appropriately as the number of nodes increases;

[0098] Perform the following steps to perform binary clustering based on the main variance dimension:

[0099] S110, Initialize the family set and the initial family containing all eigenvectors :

[0100] , Treat all samples as a whole to facilitate recursive splitting;

[0101] S120, iterative splitting, until... :

[0102] S121, For each clan ,calculate , ;in, For the clan In the middle, the first The variance of all sample values ​​in the dimension; This is the current set of vectors with splitting. For the clan A normalized eigenvector, that is, the value of a node at a certain time. ; For this vector In the The components in the dimension correspond to the normalized 1st dimension. Item measurement ;

[0103] Elected The largest clan, denoted as Prioritize refining the clusters with the greatest internal differences to improve clustering results;

[0104] S122. Calculate the principal variance dimension: ;

[0105] S123. Calculate the median of this dimension. ;

[0106] Clearly define in which dimension and where it is divided into two;

[0107] S124, split into , ;in, A subset whose values ​​on the principal variance dimension do not exceed the median; This refers to the subset whose values ​​on the principal variance dimension exceed the median.

[0108] S125, Update ;

[0109] Implement the binary steps of clustering and recursively construct... Individual ethnic groups;

[0110] Calculate the first Prototype vector of class state , ;in, The first cluster obtained after variance-based binary clustering Each state cluster is divided into several sub-families; a representative vector is generated for each state cluster to facilitate subsequent potential function calculation.

[0111] By introducing "primary variance-driven binary clustering," this invention provides a highly efficient and interpretable adaptive partitioning strategy for state space partitioning. This method calculates the dimensional variance of each sample class and selects the most dissimilar dimension for binary partitioning, ensuring that each partition maximizes inter-cluster discriminability. This approach avoids the problems of traditional methods such as K-Means or DBSCAN, such as the need to pre-determine the number of clusters, hyperparameter sensitivity, and unstable results, achieving an intelligent partitioning mechanism where "the number of clusters moves with the data." The resulting state prototype has the advantages of strong representativeness and clear visualization, making it particularly suitable for subsequent use as potential contextual state labels, effectively supporting efficient evolutionary modeling of intelligent perception systems.

[0112] The calculation of the single-point potential energy of a node at a specific moment, based on the distance between the normalized feature vector and various prototype vectors, and the state of the node, specifically includes:

[0113] compute nodes At any moment State assigned single-point potential Specifically:

[0114] , ;in, The global maximum distance; measures the state to which a node is assigned. The degree of non-compliance is dimensionless after normalization.

[0115] By defining node potential energy, the rationality of assigning each node to a particular state is transformed into a numerical evaluation index, quantifying its similarity to the prototype state vector. The normalized form of this potential energy function ensures the dimensionless nature of the measurement, eliminating the error accumulation problem caused by differences in data type and scale. Through this step, the system can evaluate the "belonging degree" of a node based on its own characteristics, providing a refined, data-driven state matching mechanism. This single-point potential function provides the basic building block for the overall energy function, supporting subsequent optimization algorithms to precisely control the overall state allocation and ultimately improving the consistency and reliability of state recognition.

[0116] The calculation of the potential energy of spatial connection edges based on the feature distance and state label between nodes at the current time specifically includes:

[0117] Calculate normalized distance ;

[0118] Set edge At any moment Spatial potential ;in, For nodes The status label assigned at the current moment; For nodes The state label assigned at the current moment; both labels take values ​​from This corresponds to different prototype vector clusters; it encourages similar nodes to take the same state and penalizes different states, with the intensity varying with similarity.

[0119] The introduction of spatial edge potential allows for explicit modeling of interactions between nodes. By calculating the consistency of state labels among nodes with similar characteristics and assigning penalty weights to inconsistent state configurations, the system effectively guides similar nodes to converge to the same state, effectively suppressing the discreteness and randomness of classification. This is particularly suitable for scenarios with strong adjacency semantics, such as opinion propagation modeling in social networks and adjacent pixel consistency determination in image segmentation. This step transforms local connectivity into part of the global state potential, significantly improving the smoothness and semantic rationality of state assignment in spatial structure.

[0120] The penalty term for setting the state change of the same node between adjacent time points specifically includes:

[0121] Set the penalty for state transitions of the same node at adjacent time steps as follows: Specifically:

[0122] ;in, For a certain node at time... Status labels; For the same node at time The state label; apply a fixed penalty to the state changes of the same node at adjacent times to promote time sequence smoothing.

[0123] The temporal transition potential addresses the issues of frequent state abrupt changes and lack of stability in traditional context recognition. By designing a penalty mechanism for state changes of the same node at consecutive time steps, the model is guided to maintain state stability during state allocation. This design is well-suited for handling temporal problems with "gradual" characteristics, such as user interest modeling, emotion evolution, and environmental state monitoring. It significantly reduces meaningless state jumps and enhances the coherence and interpretability of the context evolution path. Compared to ignoring temporal continuity or using simple sliding window mechanisms, this method strengthens the natural logic of state evolution over time at the model level.

[0124] The construction of a global energy function based on the single-point potential energy, the potential energy of the spatially connected edges, and the penalty term, with minimizing this energy function as the global optimal objective, specifically includes:

[0125] Calculate global energy Specifically:

[0126] ;in, For nodes At any moment The final optimal state label;

[0127] , ;in, This is the optimal state allocation scheme that includes all nodes and all time points;

[0128] By integrating three types of potential functions—single-point, spatial, and temporal—we can find the globally optimal state allocation.

[0129] This invention proposes a globally unified state optimization framework by constructing a global energy function that integrates single-point potential, spatial boundary potential, and temporal transition potential. This energy function is not only clear in form and rigorous in structure, but also provides quantitative evaluation indicators for system performance, facilitating practical deployment and tuning. Minimizing this energy function essentially involves finding the most reasonable state allocation method across time, space, and feature dimensions, effectively coordinating the balance between local information and global constraints. Compared to traditional decentralized methods of point-by-point classification and time-by-time prediction, this method possesses advantages such as global optimality, strong robustness, and good controllability, making it suitable for unified modeling in large-scale dynamic context systems.

[0130] The improved iterative conditional pattern algorithm is used for point-by-point convergence inference to obtain the optimal state sequence of each node over time, specifically including:

[0131] initialization , Provides an initial state estimate for each node at each time step, simplifying the starting point for subsequent optimization;

[0132] Update point by point until all No longer changing:

[0133] S201, Enumerate each pair ;

[0134] S202, Order

[0135] ;

[0136] S203, Boundary Time or When, omit non-existent elements. item;

[0137] With the states of other nodes fixed, we continuously search for the optimal state of the current node to gradually reduce the global energy.

[0138] After convergence, the final state sequence of each node over time is obtained. The state assignment is completed, and the evolution path of each node over time is obtained, which can be used for subsequent analysis or decision-making.

[0139] By employing a variant of ICM for inference, the global optimal state allocation problem is transformed into a pointwise energy descent process, effectively reducing computational complexity and improving convergence speed. This method not only avoids the problem of traditional global search easily getting trapped in local optima, but also ensures the monotonically decreasing trend of the overall energy function by optimizing the local energy function of the current node state.

[0140] Example 2: A multi-level contextual intelligence construction method based on artificial intelligence applied to text scenarios, comprising:

[0141] by For example.

[0142] S1. Raw metric acquisition and normalization:

[0143]

[0144] Normalized denominator:

[0145] ;

[0146] ;

[0147]

[0148] S2, Topology construction within time step:

[0149] distance ;

[0150]

[0151] Global threshold:

[0152] ;

[0153] Edge set :

[0154] :only satisfy ;

[0155] : , The conditions are met.

[0156] S3, Adaptive partitioning of state prototypes:

[0157] Number of states ;

[0158] The complete set of feature vectors consists of 6 points;

[0159] Calculate the variance of each dimension: Dimension 1 variance Dimension 2 variance 1;

[0160] Median split: Dimension 1 value sorting , median ;

[0161] Cluster partitioning:

[0162] : correspond ;

[0163] Remaining ;

[0164] Prototype vector:

[0165] ;

[0166] ;

[0167] S4, nodal potential :

[0168] , ;

[0169]

[0170] S5, Spatial Potential :

[0171] ;

[0172] For example ,side : , ,

[0173] .

[0174] S6, Time Transition Potential :

[0175] .

[0176] S7, Global Energy :

[0177] ;

[0178] Example: Get initial allocation The values ​​in the table can be calculated. .

[0179] S8, Point-by-Point ICM Inference:

[0180] Initialization: See the initialization table. ;

[0181] Point-by-point update: for each ,make

[0182] ;

[0183] For example, updating When considering Spatial potential Leap ;

[0184] Converged output: Final state sequence

[0185] .

[0186] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0187] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-level context intelligent construction method based on artificial intelligence for text scenarios, characterized in that, include: Collect raw context metric data and perform normalization processing to construct normalized feature vectors of context instance nodes at each time step; The original contextual measurement data is the keyword frequency; Based on the normalized context feature vector, the spatial topology is adaptively constructed at each time step; The number of prototype states is adaptively divided by binary clustering based on the main variance dimension of the context features. Based on the distance between the normalized feature vector and various prototype vectors, the single-point potential of the node at a specific moment is assigned to a specific state. Calculate the potential energy of spatial connection edges based on the feature distance and state label between nodes at the current time. Set a penalty term for the state change of the same node between adjacent time points; A global energy function is constructed based on the single-point potential energy, the potential energy of the spatial connecting edges, and the penalty term, and minimizing this energy function is taken as the global optimal objective. An improved iterative conditional pattern algorithm is used for point-by-point convergence inference to obtain the optimal state sequence of each node over time.

2. The method for constructing multi-level contextual intelligence based on artificial intelligence for text scenarios according to claim 1, characterized in that, The process of collecting raw context metric data and performing normalization to construct normalized feature vectors for context instance nodes at each time step specifically includes: Let the total set of context instance nodes be ;in, For the first One context instance node; This represents the total number of context instance nodes. Let the complete set of time indices be ;in, For time indexing; Get Nodes At any moment The Original Measure ;in, For measuring latitude index; This represents the total number of original latitude measurements. Calculate the normalized metric ; Construct nodes At any moment Normalized eigenvectors .

3. The method for constructing multi-level contextual intelligence based on artificial intelligence for text scenarios according to claim 2, characterized in that, The adaptive construction of the spatial topology structure based on the normalized context feature vector at each time step specifically includes: compute nodes and At any moment Feature distance Specifically: ;in, It is a Euclidean norm; Set the global distance threshold to ; in, To find the midvalue function of a set; Construction time The set of undirected edges , .

4. The method for constructing multi-level contextual intelligence based on artificial intelligence for text scenarios according to claim 3, characterized in that, The method of adaptively dividing the number of prototype states by binary clustering based on the main variance dimension of context features specifically includes: Set the total number of discrete states to ;in, To round down; Perform the following steps to perform binary clustering based on the main variance dimension: S110, Initialize the family set and the initial family containing all eigenvectors : ， ； S120, iterative splitting, until... : S121, For each clan ,calculate , ;in, For the clan In the middle, the first The variance of all sample values ​​in the dimension; This is the current set of vectors with splitting. For the clan A normalized eigenvector, that is, the value of a node at a certain time. ; For this vector In the The components in the dimension correspond to the normalized 1st dimension. Item measurement ; Elected The largest clan, denoted as ; S122. Calculate the principal variance dimension: ; S123. Calculate the median of this dimension. ; S124, split into , ;in, A subset whose values ​​on the principal variance dimension do not exceed the median; This refers to the subset whose values ​​on the principal variance dimension exceed the median. S125, Update ; Calculate the first Prototype vector of class state , ;in, The first cluster obtained after variance-based binary clustering Individuals.

5. The method for constructing multi-level contextual intelligence based on artificial intelligence for text scenarios according to claim 4, characterized in that, The calculation of the single-point potential energy of a node at a specific moment, based on the distance between the normalized feature vector and various prototype vectors, and the state of the node, specifically includes: compute nodes At any moment State assigned single-point potential Specifically: , ;in, This represents the global maximum distance.

6. The method for constructing multi-level contextual intelligence based on artificial intelligence for text scenarios according to claim 5, characterized in that, The calculation of the potential energy of spatial connection edges based on the feature distance and state label between nodes at the current time specifically includes: Calculate normalized distance ; Set edge At any moment Spatial potential ;in, For nodes The status label assigned at the current moment; For nodes The state label assigned at the current moment; both labels take values ​​from , corresponding to different prototype vector clusters.

7. The method for constructing multi-level contextual intelligence based on artificial intelligence for text scenarios according to claim 6, characterized in that, The penalty term for setting the state change of the same node between adjacent time points specifically includes: Set the penalty for state transitions of the same node at adjacent time steps as follows: Specifically: ;in, For a certain node at time... Status labels; For the same node at time Status labels.

8. The method for constructing multi-level contextual intelligence based on artificial intelligence for text scenarios according to claim 7, characterized in that, The construction of a global energy function based on the single-point potential energy, the potential energy of the spatially connected edges, and the penalty term, with minimizing this energy function as the global optimal objective, specifically includes: Calculate global energy Specifically: ;in, For nodes At any moment The final optimal state label; , ;in, This is the optimal state allocation scheme that includes all nodes and all time points.

9. A multi-level context intelligent construction method based on artificial intelligence applied to text scenarios according to claim 8, characterized in that, The improved iterative conditional pattern algorithm is used for point-by-point convergence inference to obtain the optimal state sequence of each node over time, specifically including: initialization , ; Update point by point until all No longer changing: S201, Enumerate each pair ; S202, Order ； S203, Boundary Time or When, omit non-existent elements. item; After convergence, the final state sequence of each node over time is obtained. .

Images