Intelligent retrieval precision matching method and system based on multi-level semantic decomposition

Abstract

The application provides an intelligent retrieval accurate matching method and system based on multi-level semantic decomposition, relates to the technical field of information retrieval, and comprises the following steps: generating a hierarchical semantic representation tree by performing semantic decomposition on a text to be retrieved; constructing a semantic propagation graph and optimizing the same by using an adaptive probability sampling strategy; extracting a semantic feature vector with context dependency relationship fused to establish an index; and analyzing a retrieval request and matching the same with the semantic propagation graph to generate a result. The application improves retrieval accuracy, enhances semantic understanding capability, optimizes retrieval efficiency, and improves the relevance and accuracy of the result.

Description

Technical Field

[0001] This invention relates to information retrieval technology, and more particularly to an intelligent retrieval and precise matching method and system based on multi-level semantic decomposition. Background Technology

[0002] With the explosive growth of the internet and digital information, information retrieval technology is increasingly widely used in various fields. Traditional information retrieval mainly relies on keyword matching, determining search results by calculating the similarity between query terms and keywords in documents. With technological advancements, semantic retrieval has gradually become a research hotspot. Its core is understanding the deep semantics of text and matching based on semantic similarity to improve retrieval accuracy. Existing semantic retrieval technologies mainly include word vector-based methods, deep learning-based methods, and knowledge graph-based methods. Word vector-based methods, such as Word2Vec and GloVe, map words to a vector space and measure semantic similarity by calculating the distance between vectors. Deep learning-based methods, such as BERT and RoBERTa pre-trained models, can capture contextual information and generate richer semantic representations. Knowledge graph-based methods achieve structured semantic representation and reasoning by constructing a network of relationships between concepts.

[0003] Existing technologies perform poorly when dealing with complex semantic structures, especially for texts with multi-layered semantic connotations. They struggle to perform accurate semantic decomposition and representation, resulting in insufficient deep semantic understanding of the text during the retrieval process and affecting retrieval accuracy.

[0004] Traditional retrieval systems typically ignore contextual dependencies during semantic feature extraction, focusing only on local semantic information. This makes it difficult to effectively construct a re-ranking mechanism that comprehensively considers semantic similarity and node importance. Consequently, retrieval results often contain a large amount of semantically related but not precisely matched content, reducing retrieval efficiency and user experience. Summary of the Invention

[0005] The embodiments of the present invention provide an intelligent retrieval and precise matching method and system based on multi-level semantic decomposition, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides an intelligent retrieval and precise matching method based on multi-level semantic decomposition, comprising:

[0007] The text to be retrieved is obtained, and the text to be retrieved is semantically decomposed to generate a hierarchical semantic representation tree in a multidimensional semantic feature space.

[0008] The hierarchical semantic representation tree is traversed based on an adaptive probability sampling strategy. The semantic association strength between nodes is calculated, a semantic propagation graph is constructed, and iterative optimization is performed on the semantic propagation graph to obtain the optimal semantic path. The semantic nodes are reorganized according to the optimal semantic path to form an optimized hierarchical semantic representation tree.

[0009] Semantic feature vectors that incorporate contextual dependencies are extracted from the optimized hierarchical semantic representation tree to establish a semantic feature index;

[0010] The system receives retrieval requests and performs semantic parsing. It then matches the parsing results with the node information in the semantic propagation graph to generate an initial candidate result set. The initial candidate result set is then reordered according to the node hierarchy, and filtered based on semantic similarity and node importance to obtain the final accurate matching result.

[0011] Semantic decomposition of the text to be retrieved, and generation of a hierarchical semantic representation tree in a multidimensional semantic feature space, includes:

[0012] The text to be retrieved is input into the deep semantic processor, which extracts the semantic features from the text to obtain the initial semantic feature sequence;

[0013] Distributed Monte Carlo tree sampling is performed on the initial semantic feature sequence to calculate the semantic correlation between feature nodes. A semantic propagation network is constructed based on the semantic correlation, and the node weights in the semantic propagation network are updated through iterative optimization to identify the hierarchical semantic structure of the text.

[0014] Based on the hierarchical semantic structure, semantic levels are divided, and a multidimensional semantic feature space is constructed.

[0015] An adaptive bundle search is performed in the multidimensional semantic feature space to dynamically adjust the search width, map semantic features to different levels, and perform feature fusion according to the importance of features to generate a hierarchical semantic representation tree with conceptual features, relational features, and attribute features.

[0016] Distributed Monte Carlo tree sampling is performed on the initial semantic feature sequence to calculate the semantic correlation between feature nodes. A semantic propagation network is then constructed based on the semantic correlation, including:

[0017] Feature node information is extracted from the initial semantic feature sequence. Multiple rounds of random walk sampling are performed on the feature node information based on a distributed Monte Carlo tree. The cumulative reward value of the sampling path in each round is calculated. The node transition probability is determined based on the cumulative reward value. The node transition probability is optimized by a temperature annealing strategy to obtain the semantic correlation between feature nodes.

[0018] The feature node information is grouped according to semantic relevance to form multiple semantic clusters;

[0019] Based on the semantic relevance, an initial topology of the semantic propagation network is constructed. Adaptive weight propagation is performed on the initial topology, and the network connection strength is iteratively updated through local neighborhood expansion to finally generate a semantic propagation network with a hierarchical structure.

[0020] Based on an adaptive probability sampling strategy, the hierarchical semantic representation tree is traversed node by node, the semantic association strength between nodes is calculated, and a semantic propagation graph is constructed, including:

[0021] The node traversal order is determined based on a distributed Monte Carlo search strategy, and adaptive probability sampling is performed on each traversed node. The sampling probability is dynamically adjusted according to the distribution of nodes in different semantic levels to generate a node sampling sequence.

[0022] Semantic feature vectors are extracted from the node sampling sequence to construct a semantic space mapping;

[0023] The semantic association strength between nodes is calculated in the semantic space mapping. Temporal dependency analysis of the semantic association strength is performed through a gated recurrent network to establish node propagation paths. A semantic propagation graph is constructed based on local neighborhood aggregation.

[0024] The semantic association strength between nodes is calculated in the semantic space mapping. Temporal dependency analysis of this semantic association strength is performed using a gated recurrent network to establish node propagation paths. A semantic propagation graph is then constructed based on local neighborhood aggregation.

[0025] Construct a node co-occurrence matrix, calculate the semantic association strength between nodes, perform multi-round iterative propagation analysis on the node co-occurrence matrix, update the node state and adjust the propagation weight in each round of iteration, determine the temporal dependency relationship based on the node state change, generate a node association sequence, and optimize the structure of the node association sequence using the node co-occurrence matrix.

[0026] Hierarchical clustering is performed on the node association sequences, and node semantic clusters are divided based on the clustering results. The degree of association between semantic clusters is then calculated.

[0027] An initial propagation path is established based on the semantic association strength in the node co-occurrence matrix. A node propagation subgraph is constructed by combining local neighborhood information. The node propagation subgraph is iteratively optimized by an adaptive weight adjustment method. Multi-scale node association information is integrated to finally construct a semantic propagation graph with a hierarchical structure.

[0028] The process involves receiving a retrieval request, performing semantic parsing, matching the parsing results with node information in the semantic propagation graph, and generating an initial candidate result set, including:

[0029] The system receives a retrieval request, extracts feature information from the request, performs distributed Monte Carlo sampling on the feature information, dynamically adjusts the sampling weights during the sampling process, maps the weighted sampling path to a semantic feature vector, optimizes the semantic feature vector through multiple rounds of iteration to obtain the first semantic vector of the retrieval request, and generates a multi-scale feature representation of the first semantic vector.

[0030] Node information is extracted from the semantic propagation graph, the node information is converted into a second semantic vector, and a multi-level node feature index is constructed based on the second semantic vector;

[0031] An adaptive beam search is performed in the multi-level node feature index. The multi-scale feature representation is used to perform feature matching at different levels. The matching scores of each level are accumulated to generate a path score. The matching path with the highest score is selected as the optimal matching path. Node information is collected along the optimal matching path. The collected node information is sorted according to the connection strength of the nodes in the semantic propagation graph to generate an initial candidate result set.

[0032] The initial candidate result set is reordered according to the node hierarchy, and then filtered based on semantic similarity and node importance to obtain the final matching results, including:

[0033] A directed reordering graph is constructed based on the node hierarchy. A distributed Monte Carlo tree search is performed on the directed reordering graph. The node importance score is determined by the node access frequency. At the same time, the semantic similarity between adjacent nodes is calculated. The node importance score and the semantic similarity are combined to generate the node ranking weight.

[0034] The nodes in the initial candidate result set are grouped according to the node sorting weight to form multiple candidate result clusters;

[0035] Within each candidate result cluster, a reordering path is constructed. The optimal reordering sequence is determined based on the cumulative sorting weight of the nodes on the reordering path. The candidate results are then rearranged along the optimal reordering sequence. A hierarchical screening is performed on the rearranged candidate results in combination with a preset sorting weight threshold to obtain the final matching result.

[0036] A second aspect of the present invention provides an intelligent retrieval and precise matching system based on multi-level semantic decomposition, comprising:

[0037] The acquisition unit is used to acquire the text to be retrieved, perform semantic decomposition on the text to be retrieved, and generate a hierarchical semantic representation tree in a multidimensional semantic feature space.

[0038] The optimization unit is used to perform node traversal on the hierarchical semantic representation tree based on an adaptive probability sampling strategy, calculate the semantic association strength between nodes, construct a semantic propagation graph, perform iterative optimization on the semantic propagation graph to obtain the optimal semantic path, and reorganize the semantic nodes according to the optimal semantic path to form an optimized hierarchical semantic representation tree.

[0039] The extraction unit is used to extract semantic feature vectors that incorporate contextual dependencies from the optimized hierarchical semantic representation tree and establish a semantic feature index.

[0040] The filtering unit is used to receive retrieval requests and perform semantic parsing, match the parsing results with the node information in the semantic propagation graph, generate an initial candidate result set, reorder the initial candidate result set according to the node hierarchy, and filter the results by combining semantic similarity and node importance to obtain the final accurate matching result.

[0041] A third aspect of the present invention provides an electronic device, comprising:

[0042] processor;

[0043] Memory used to store processor-executable instructions;

[0044] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0045] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0046] The beneficial effects of this application are as follows:

[0047] By semantically decomposing the text to be retrieved and generating a hierarchical semantic representation tree in a multidimensional semantic feature space, the retrieval system is able to capture the deep semantic structure of the text, rather than being limited to surface text matching, thus significantly improving the semantic understanding capability of the retrieval.

[0048] An adaptive probability sampling strategy is adopted to traverse the nodes of the hierarchical semantic representation tree, and a semantic propagation graph is constructed by calculating the semantic association strength between nodes. The optimal semantic path is obtained through iterative optimization, thereby reorganizing the semantic nodes to form a more reasonable hierarchical semantic representation structure, which greatly improves the accuracy and efficiency of semantic representation.

[0049] By extracting semantic feature vectors that incorporate contextual dependencies from the optimized hierarchical semantic representation tree, the retrieval system can understand the contextual semantics of words in the text, effectively solving the problems of polysemy and synonyms in traditional retrieval and improving the semantic understanding capability of retrieval.

[0050] The retrieval process matches nodes based on information in the semantic propagation graph and reorders nodes using hierarchical relationships. It also filters nodes based on semantic similarity and importance, thus realizing an intelligent retrieval mechanism that evaluates matching from multiple dimensions. This significantly improves the accuracy and relevance of retrieval results. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the intelligent retrieval and precise matching method based on multi-level semantic decomposition according to an embodiment of the present invention.

[0052] Figure 2 This is a flowchart illustrating the semantic propagation network construction and node weight optimization iteration of an embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0054] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0055] Figure 1 This is a flowchart illustrating the intelligent retrieval and precise matching method based on multi-level semantic decomposition according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0056] The text to be retrieved is obtained, and the text to be retrieved is semantically decomposed to generate a hierarchical semantic representation tree in a multidimensional semantic feature space.

[0057] The hierarchical semantic representation tree is traversed based on an adaptive probability sampling strategy. The semantic association strength between nodes is calculated, a semantic propagation graph is constructed, and iterative optimization is performed on the semantic propagation graph to obtain the optimal semantic path. The semantic nodes are reorganized according to the optimal semantic path to form an optimized hierarchical semantic representation tree.

[0058] Semantic feature vectors that incorporate contextual dependencies are extracted from the optimized hierarchical semantic representation tree to establish a semantic feature index;

[0059] The system receives retrieval requests and performs semantic parsing. It then matches the parsing results with the node information in the semantic propagation graph to generate an initial candidate result set. The initial candidate result set is then reordered according to the node hierarchy, and filtered based on semantic similarity and node importance to obtain the final accurate matching result.

[0060] In one optional implementation, semantically decomposing the text to be retrieved and generating a hierarchical semantic representation tree in a multidimensional semantic feature space includes:

[0061] The text to be retrieved is input into the deep semantic processor, which extracts the semantic features from the text to obtain the initial semantic feature sequence;

[0062] Distributed Monte Carlo tree sampling is performed on the initial semantic feature sequence to calculate the semantic correlation between feature nodes. A semantic propagation network is constructed based on the semantic correlation, and the node weights in the semantic propagation network are updated through iterative optimization to identify the hierarchical semantic structure of the text.

[0063] Based on the hierarchical semantic structure, semantic levels are divided, and a multidimensional semantic feature space is constructed.

[0064] An adaptive bundle search is performed in the multidimensional semantic feature space to dynamically adjust the search width, map semantic features to different levels, and perform feature fusion according to the importance of features to generate a hierarchical semantic representation tree with conceptual features, relational features, and attribute features.

[0065] The text to be retrieved is first input into a deep semantic processor for processing. The deep semantic processor employs a multi-layer transformer architecture, containing 12 encoding layers, each with 12 attention heads. The hidden layer dimension is 768, and the intermediate layer dimension is 3072. The deep semantic processor receives text input with a maximum length of 4096 tokens, and performs word segmentation using byte-based encoding. The processor performs context-sensitive vectorization transformation on the input text, extracting semantic features from the text. Each token is mapped to a 768-dimensional vector representation. The extracted semantic features include three categories: lexical features, syntactic features, and discourse features. Lexical features include part-of-speech and word morphology information; syntactic features include dependency relations and syntactic tree structure; and discourse features include referential relations and topic coherence. Through feature extraction, the text to be retrieved is converted into an initial semantic feature sequence, where each element is a 768-dimensional vector, preserving the temporal relationships of the original text.

[0066] When performing distributed Monte Carlo tree sampling on the initial semantic feature sequence, 16 parallel computing nodes are used, each responsible for sampling different feature subspaces. Each node randomly selects a starting point from the feature sequence, performs 1000 simulations to explore the feature space, and the sampling depth is set to 15 layers. During the sampling process, the node selection strategy follows the upper confidence bound algorithm, and the exploration coefficient is set to 1.414 to balance exploration and utilization. The semantic correlation between feature nodes is calculated using a weighted combination of cosine similarity and semantic distance. In the calculation formula, the cosine similarity weight is 0.7, and the semantic distance weight is 0.3. The semantic distance is obtained by normalizing the Euclidean distance of the feature vectors. The correlation threshold is set to 0.65, and node pairs above this threshold are considered to have semantic connections. A semantic propagation network is constructed based on the calculated semantic correlation. The weight of each node is set to 1.0 during network initialization. The node weights in the semantic propagation network are updated iteratively using stochastic gradient descent. The initial learning rate is set to 0.01 and decreases to 0.95 times the original value after every 500 iterations according to the exponential decay rule. The iteration terminates when the node weight change rate is less than 0.001 or when the maximum number of iterations (5000) is reached. The optimization process is achieved by minimizing a global semantic loss function, which comprehensively considers node cohesion and inter-node separation. After iteration, the hierarchical semantic structure of the text is identified based on node weights and connection patterns. High-weight nodes are considered higher-level concepts, and interconnected node groups form semantic clusters.

[0067] When dividing the semantic hierarchy based on the identified hierarchical semantic structure, a combination of bottom-up clustering and top-down segmentation is used. The semantic hierarchy is divided into three main levels: the concept level, the relation level, and the attribute level. The concept level is at the top, the relation level is in the middle, and the attribute level is at the bottom. Each level is further subdivided according to semantic importance: the concept level is subdivided into core concept sub-layers and auxiliary concept sub-layers; the relation level is subdivided into explicit relation sub-layers and implicit relation sub-layers; and the attribute level is subdivided into feature attribute sub-layers and constraint attribute sub-layers. The hierarchical division uses an improved spectral clustering algorithm, with the eigenvalue decomposition threshold set to 0.1. The number of clusters is adaptively determined, with a minimum of 3 and a maximum of the square root of the number of sentences in the text to be retrieved, rounded up. When constructing the multidimensional semantic feature space, the number of spatial dimensions is the same as the initial semantic feature dimensions, which is 768 dimensions. However, dimensionality reduction is achieved through principal component analysis, retaining principal components that explain 95% of the variance, typically reducing the dimension to between 100 and 150. Each dimension represents a type of semantic feature, and the importance of a dimension is determined by the magnitude of its eigenvalues, with the importance threshold set to 1% of the sum of all eigenvalues.

[0068] When performing adaptive beam search in the multidimensional semantic feature space, the initial beam width is set to 16, and dynamically adjusted according to the search level and semantic complexity. The adjustment rule is as follows: when the variance of the scores of candidate nodes in three consecutive layers of search is less than 0.05, the beam width is increased by 50%; when the variance is greater than 0.2, the beam width is decreased by 25%, with an upper limit of 64 and a lower limit of 8. During the search process, a heuristic scoring function is used to evaluate the quality of candidate paths. The scoring function comprehensively considers three factors: node weight, similarity to the target semantics, and path length, with weights of 0.5, 0.3, and 0.2, respectively. When mapping semantic features to different levels, a hierarchical attention mechanism is adopted. The attention calculation uses a scaled dot product, and the temperature parameter is set to 2.0. The mapping threshold increases with the level: 0.3 for the bottom layer, 0.5 for the middle layer, and 0.7 for the top layer. The information gain ratio method is used to evaluate the importance of features, calculating the contribution of each feature to the semantic discriminative ability. Features with a gain ratio higher than 0.1 are considered important features. During feature fusion, the weight of important features is set to three times that of ordinary features, and a weighted average method is used to generate node representation vectors. The final hierarchical semantic representation tree contains three types of features: concept features, relation features, and attribute features. The concept feature vector has a dimension of 256, describing the text's topic and core concepts; the relation feature vector has a dimension of 128, describing the dependencies and logical relationships between concepts; and the attribute feature vector has a dimension of 64, describing the modifying and limiting information of concepts and relationships. The maximum depth of the tree is 7 levels, and the maximum number of nodes in each level is four times the current depth. The root node represents the global semantics of the text, and the leaf nodes represent fine-grained semantic units.

[0069] In one optional implementation, distributed Monte Carlo tree sampling is performed on the initial semantic feature sequence to calculate the semantic correlation between feature nodes, and a semantic propagation network is constructed based on the semantic correlation, including:

[0070] Feature node information is extracted from the initial semantic feature sequence. Multiple rounds of random walk sampling are performed on the feature node information based on a distributed Monte Carlo tree. The cumulative reward value of the sampling path in each round is calculated. The node transition probability is determined based on the cumulative reward value. The node transition probability is optimized by a temperature annealing strategy to obtain the semantic correlation between feature nodes.

[0071] The feature node information is grouped according to semantic relevance to form multiple semantic clusters;

[0072] Based on the semantic relevance, an initial topology of the semantic propagation network is constructed. Adaptive weight propagation is performed on the initial topology, and the network connection strength is iteratively updated through local neighborhood expansion to finally generate a semantic propagation network with a hierarchical structure.

[0073] like Figure 2 As shown, the method includes:

[0074] When extracting feature node information from the initial semantic feature sequence, the semantic feature sequence is represented as a set of N feature nodes, each containing a 768-dimensional vector representation. The feature node information extraction module receives the initial semantic feature sequence and extracts the position information, semantic vector, and contextual information for each node. The position information records the relative position of the node in the original text, with values ​​ranging from 0 to 1 (normalized values). The semantic vector is the 768-dimensional vector corresponding to the position in the initial feature sequence. The contextual information is obtained using a sliding window method, with a default window size of 5, extracting the indices of the five nodes before and after the current node. For each feature node, a node descriptor data structure is constructed, containing four fields: node ID, semantic vector, position information, and neighborhood information. The node ID is an integer, numbered sequentially starting from 0; the semantic vector is a 768-dimensional floating-point array; the position information is a floating-point number between 0 and 1; and the neighborhood information is an integer array storing the IDs of neighboring nodes. After extraction, a set of feature node information is formed, serving as input for subsequent distributed Monte Carlo tree sampling.

[0075] When performing multi-round random walk sampling of feature node information based on a distributed Monte Carlo tree, an 8-node parallel computing architecture is adopted, with each computing node configured with 16GB of memory and an 8-core processor. The total number of sampling rounds is set to 10,000, with a maximum step size of 50 steps per round. Before sampling begins, the root node set is initialized, selecting the top 10 nodes with the largest semantic vector magnitudes from the feature node information set as root nodes. Each computing node randomly selects a node from the root node set as its starting point and performs a random walk. During the walk, the current node determines its next transition target based on its semantic similarity to other nodes. Semantic similarity is calculated using cosine distance, ranging from -1 to 1; a larger value indicates greater semantic similarity. Based on the semantic similarity between the current node i and candidate node j, the initial transition probability is the normalized similarity value. The cumulative reward value is calculated using a discount factor γ and a node value function, with γ set to 0.85 by default. The node value function combines two metrics: semantic importance and structural centrality. Semantic importance is represented by the L2 norm of the semantic vector, while structural centrality is calculated from the node's degree. During multiple rounds of sampling, the cumulative reward value of each sampling path is recorded and used to update the node transition probability.

[0076] When optimizing node transition probabilities using a temperature annealing strategy, the initial temperature T0 is set to 1.0, the temperature decay rate is 0.95, and the minimum temperature threshold is 0.01. After each round of sampling, the node transition probability is adjusted according to the current temperature T. The higher the temperature, the more uniform the transition probability distribution; the lower the temperature, the more prominent the probability of high-reward paths. The temperature is updated according to an exponential decay rule, and after every 100 rounds of sampling, the temperature is multiplied by the decay rate. The annealing process ends when the temperature falls below the minimum temperature threshold or the number of sampling rounds reaches the upper limit. Finally, the semantic correlation between feature nodes is calculated by statistically analyzing the co-occurrence frequency of node pairs in all sampled paths and combining it with the cumulative reward value. The semantic correlation matrix is ​​an N×N square matrix, where each element represents the correlation strength between the corresponding node pairs, with values ​​ranging from 0 to 1. The correlation threshold is set to 0.3; correlations below this threshold are considered noise and filtered out.

[0077] When grouping feature node information into multiple semantic clusters based on semantic relevance, an improved spectral clustering algorithm is employed. A similarity graph is constructed based on the semantic relevance matrix, with a similarity threshold of 0.25; edges below this threshold are pruned. A Laplacian matrix transformation is performed on the similarity graph, and the eigenvalues ​​and eigenvectors of the Laplacian matrix are calculated. The eigenvalues ​​are sorted in descending order, and the optimal number of clusters k is determined using feature interval analysis, with a default range of 3 to 15. The first k eigenvectors are selected to construct a low-dimensional representation space, where the K-means algorithm is used to cluster the nodes. The maximum number of iterations is set to 500, and the convergence threshold is 1e-4. After clustering, k semantic clusters are formed, each containing a set of semantically closely related feature nodes. For boundary nodes (nodes with similar relevance among multiple clusters), a soft clustering strategy is adopted, allowing nodes to be assigned to multiple clusters with different weights. The weights are allocated according to normalized relevance, and the threshold for assigning a primary cluster is set to 0.6.

[0078] When constructing the initial topology of the semantic propagation network based on semantic relevance, each semantic cluster is considered as a community within the network. Connection weights between nodes within a community are directly based on semantic relevance, while connections between communities are established through boundary nodes. The initial connection weights are the geometric mean of the boundary node's weights in both communities. To ensure network connectivity, a global connection sparsity constraint is added, retaining the top 20% of connections with the highest relevance for each node, with at least three connections maintained. Each node in the network maintains three types of information: a node feature vector (768 dimensions), a set of node connections, and a current node activity value. After the initial topology is constructed, node activity is uniformly initialized to 1.0.

[0079] When performing adaptive weight propagation on the initial topology, an information propagation model is used, with a maximum of 200 iterations and a convergence threshold of 0.005. In each iteration, nodes propagate their activity values ​​to their neighbors through connections, with the propagation strength proportional to the connection weight. The activity update formula considers the node's own retention factor α and the information gain from its neighbors, with α having a default value of 0.15. To prevent activity values ​​from diverging, the activity values ​​of all nodes are normalized after each iteration, ensuring the sum equals the number of nodes N. When updating network connection strength through local neighborhood expansion, each node examines its second-order neighbors (neighbors of neighbors) and calculates the potential association strength with these nodes. If the potential association strength exceeds a threshold (defaulting to 1.5 times the strength of the weakest existing connection), a new connection is established. Simultaneously, the strength of existing connections is updated based on the amount of activity propagation; connections with propagation amounts less than a threshold (defaulting to 0.1 times the average propagation amount) are weakened or removed. Network updates are performed in batches, once every 10 iterations to reduce computational overhead. When the connection change rate is less than 1% after three consecutive network updates or when the maximum number of iterations is reached, the iteration terminates, ultimately generating a semantic propagation network with a hierarchical structure.

[0080] In one optional implementation, the hierarchical semantic representation tree is traversed based on an adaptive probability sampling strategy to calculate the semantic association strength between nodes and construct a semantic propagation graph, including:

[0081] The node traversal order is determined based on a distributed Monte Carlo search strategy, and adaptive probability sampling is performed on each traversed node. The sampling probability is dynamically adjusted according to the distribution of nodes in different semantic levels to generate a node sampling sequence.

[0082] Semantic feature vectors are extracted from the node sampling sequence to construct a semantic space mapping;

[0083] The semantic association strength between nodes is calculated in the semantic space mapping. Temporal dependency analysis of the semantic association strength is performed through a gated recurrent network to establish node propagation paths. A semantic propagation graph is constructed based on local neighborhood aggregation.

[0084] A hierarchical semantic representation tree, as the input data structure, consists of multiple levels of nodes. Each node contains a node identifier, a semantic vector, hierarchical information, and pointers to child nodes. The node identifier is a unique string consisting of 32 alphanumeric characters; the semantic vector is a 768-dimensional floating-point array representing the semantic features of the node; the hierarchical information is an integer value representing the depth of the node in the tree, with the root node having a hierarchical level of 0; and the child node pointers are arrays of node identifiers pointing to all child nodes of the current node. Representation trees typically have a 3- to 7-level structure, with a total of 500 to 10,000 nodes, occupying approximately 50MB to 1GB of memory.

[0085] When determining the node traversal order based on the distributed Monte Carlo search strategy, a multi-agent collaborative computing framework is adopted. The computing cluster is configured with 8 computing nodes, each equipped with 16GB of memory and an 8-core processor. Nodes communicate through a message queue mechanism, and the message format includes five fields: sending node identifier, receiving node identifier, message type, message body, and timestamp. The traversal process is divided into two phases: global planning and local execution. In the global planning phase, the master node starts from the root node and calculates the access value of each node based on the structural characteristics of the hierarchical semantic representation tree. The node access value consists of three parts: semantic importance, structural complexity, and historical access frequency. Semantic importance is calculated using the L2 norm of the node's semantic vector, with a value ranging from 0 to 10; structural complexity is the ratio of the number of nodes in the node's subtree to the total number of nodes, with a value ranging from 0 to 1; historical access frequency is initially 0 and subsequently obtained by dividing the access count by the total number of accesses, with a value ranging from 0 to 1. The three parts are weighted and summed with weights of 0.5, 0.3, and 0.2 to obtain the comprehensive access value of the node, which ranges from 0 to 8.2.

[0086] After the node access value is calculated, the master node constructs a priority queue, sorting all nodes in descending order of access value. The top 20% of nodes with the highest access value are selected from the priority queue as seed nodes and assigned to each compute node for processing. Each compute node, upon receiving its assigned seed node, performs a Monte Carlo tree search based on an improved Upper Confidence Bound (UCB) strategy. The UCB strategy combines node access value with exploration uncertainty to control the balance between search depth and breadth. The exploration constant C in the UCB calculation formula is set to 1.414 by default and can be adjusted between 0.5 and 3. A larger C value increases the tendency to explore unvisited nodes, while a smaller C value relies more on known node values. Each compute node performs 10,000 iterations, continuously updating node access statistics and synchronizing with the master node via a message queue during the iteration process. To prevent search overlap between compute nodes, each node maintains a set of explored regions to avoid duplicate calculations. When a region is selected by multiple nodes simultaneously, a timestamp-based distributed lock mechanism determines the processing right, avoiding conflicts. The intermediate results generated during the traversal process are stored in a distributed key-value store. The key is the node identifier, and the value is the node state object, which contains the number of visits, cumulative value, and child node statistics.

[0087] When performing adaptive probability sampling for each traversed node, the sampling probability is dynamically adjusted based on the node's distribution across different semantic levels. The semantic level distribution is determined by the node's own level information and its distribution density in the global semantic space. The distribution density is calculated using the K-nearest neighbor algorithm, with K=15. For each node, the 15 closest nodes in the semantic space are found, and the average semantic distance is calculated as the density index; a smaller distance indicates higher density. The sampling probability is inversely proportional to the node level and directly proportional to the distribution density. Specifically, the inverse level factor α in the sampling probability calculation formula is set to 0.7, and the direct density factor β is set to 1.2. For top-level nodes (levels 0 and 1), a minimum sampling probability of 0.05 is set to ensure that higher-level nodes are not completely ignored; for bottom-level nodes (the two deepest levels), a maximum sampling probability upper limit of 0.4 is set to prevent oversampling of specific regions. After calculation, the sampling probabilities are normalized so that the sum of the sampling probabilities of all nodes is 1.

[0088] Based on the adjusted sampling probabilities, a weighted random sampling method is used to select nodes from the traversal sequence, generating a node sampling sequence. The sampling sequence length is set to 30% of the original total number of nodes by default, with a minimum of 300 nodes and a maximum of 3000 nodes. To enhance sampling diversity, a tiered sampling strategy is adopted to ensure that nodes at each tier are sampled in a reasonable proportion. In tiered sampling, the target sampling proportion for each tier is determined by the tier importance configuration, with a default configuration of {0:0.15, 1:0.25, 2:0.3, 3:0.2, 4 and above:0.1}, meaning that 15% of the nodes are sampled from tier 0, 25% from tier 1, and so on. When the number of nodes at a certain tier is insufficient to meet the target sampling quantity, its sampling quota will be allocated to other tiers proportionally. The sampling process is implemented using a reservoir sampling algorithm, ensuring that sampling can be completed in a single traversal, with a time complexity of O(n), where n is the total number of nodes.

[0089] When extracting semantic feature vectors from the node sampling sequence, for each sampled node, its original 768-dimensional semantic vector is extracted, while the node's contextual information is fused. The contextual information includes the semantic features of the node's parent, sibling, and child nodes, which are integrated through an attention mechanism. Attention weights are calculated based on the semantic similarity between nodes, measured using cosine distance, with values ​​ranging from -1 to 1. Weight normalization uses the softmax function, with the temperature parameter τ defaulting to 0.1 and adjustable from 0.05 to 0.5. Smaller τ values ​​concentrate the weight distribution more on nodes with high similarity, while larger τ values ​​result in a more even weight distribution. The fused feature vector remains 768-dimensional, with each dimension taking a floating-point number between -1 and 1.

[0090] After semantic feature extraction, a semantic space mapping is constructed through dimensionality reduction. Dimensionality reduction is achieved using a variational autoencoder, with both the encoder and decoder being three-layer fully connected networks. The encoder maps the 768-dimensional feature vector to a 128-dimensional latent representation, with intermediate layer dimensions of 512 and 256; the decoder restores the 128-dimensional latent representation to a 768-dimensional feature vector, with intermediate layer dimensions of 256 and 512. The activation function used is LeakyReLU with a negative slope of 0.1. Training employs mini-batch gradient descent with a batch size of 64, an initial learning rate of 0.001, and a learning rate decay strategy, decreasing to 0.9 times the original value every 50 iterations. The training process iterates for 200 epochs, using a weighted sum of reconstruction error and KL divergence as the loss function, with a weight ratio of 10:1. After training, the encoder maps the feature vectors of all sampled nodes to the 128-dimensional latent space, forming the semantic space mapping.

[0091] An improved attention network mechanism is used to calculate the semantic association strength between nodes in the semantic space mapping. For any two nodes i and j, their Euclidean distance and cosine similarity in the semantic space are calculated. The Euclidean distance reflects the absolute semantic difference between nodes, while the cosine similarity reflects the directional consistency between nodes. The two are combined to form the basic association strength. Specifically, the Euclidean distance is normalized to the interval between 0 and 1, inverted, and then weighted and averaged with the cosine similarity at a weight ratio of 1:2. The basic association strength is mapped to the interval between 0 and 1 using the sigmoid function to form the initial semantic association strength. The association strength matrix is ​​an N×N square matrix, where N is the number of sampled nodes, and each element in the matrix represents the semantic association strength of the corresponding node pair. To reduce computational complexity, an association threshold of 0.2 is set. Association strengths below this threshold are considered noise and set to zero, forming a sparse association matrix with a sparsity rate typically between 85% and 95%.

[0092] When performing temporal dependency analysis on semantic association strength using a gated recurrent network (GRU), the sampled sequence is treated as time-series data. A bidirectional gated recurrent unit (GRU) network structure is adopted, with a hidden layer dimension of 64 and a layer count of 2. The network input is a sequence of semantic feature vectors of nodes, with each time step corresponding to a sampled node. To handle variable-length sequences, a dynamic time warping strategy is used to segment the original sequence into subsequences of fixed length 50, with adjacent subsequences overlapping by 10 nodes. The network training uses a truncated backpropagation algorithm with a truncation length of 20 and a learning rate of 0.001. The mean squared error loss function is used, and the training epochs are 100. The early stopping strategy is set to stop after 10 consecutive epochs of validation loss without improvement. The ratio of training data to validation data is 8:2. After the network training is completed, for each node pair (i,j), the temporal dependency strength from i to j is calculated. Combined with the previous semantic association strength, the final association strength value is obtained, ranging from 0 to 1.

[0093] When establishing node propagation paths, a directed weighted graph is constructed based on the pruned association strength matrix. For each node, the K outgoing edges with the highest association strength are retained. The value of K is dynamically determined and proportional to the semantic importance of the node, with a minimum of 3 and a maximum of 15. The propagation path is calculated using an improved Dijkstra algorithm, considering the cumulative decay of association strength along the path. The decay factor γ defaults to 0.8, meaning that for each step the path extends, the association strength decreases by 0.8 times. Path strength is defined as the weighted geometric mean of the association strengths of all edges on the path. The weight is related to the edge's position in the path; the earlier the position, the greater the weight. For any pair of nodes (i,j), the path with the highest strength is calculated as its main propagation path. Simultaneously, the second-best paths with a strength not less than 90% of the main path are recorded as alternative propagation paths.

[0094] When constructing the semantic propagation graph based on local neighborhood aggregation, a graph attention aggregation mechanism from graph neural networks is employed. Each node aggregates information from its neighboring nodes, with the aggregation weights determined by the association strength. The aggregation process iterates for three rounds, updating the node representation vector in each round. The aggregation function uses a weighted sum form, and the aggregated vector is processed through linear transformation and a nonlinear activation function, using ELU. To prevent oversmoothing, a residual connection mechanism is introduced to preserve the original feature information of the nodes. After aggregation, the association strength between nodes is recalculated based on the updated node representations, constructing an optimized semantic propagation graph. The final propagation graph is stored using an adjacency list, with each node recording its outgoing target node and the corresponding association strength. The graph density is controlled between 5% and 15%, ensuring that the graph is neither too sparse nor too dense. The graph structure supports incremental updates; when a new node is added, only the association strength between the new node and existing nodes needs to be calculated, and the affected local region is updated.

[0095] In one optional implementation, the semantic association strength between nodes is calculated in the semantic space mapping, and the semantic association strength is subjected to temporal dependency analysis through a gated recurrent network to establish node propagation paths. A semantic propagation graph is then constructed based on a local neighborhood aggregation method, including:

[0096] Construct a node co-occurrence matrix, calculate the semantic association strength between nodes, perform multi-round iterative propagation analysis on the node co-occurrence matrix, update the node state and adjust the propagation weight in each round of iteration, determine the temporal dependency relationship based on the node state change, generate a node association sequence, and optimize the structure of the node association sequence using the node co-occurrence matrix.

[0097] Hierarchical clustering is performed on the node association sequences, and node semantic clusters are divided based on the clustering results. The degree of association between semantic clusters is then calculated.

[0098] An initial propagation path is established based on the semantic association strength in the node co-occurrence matrix. A node propagation subgraph is constructed by combining local neighborhood information. The node propagation subgraph is iteratively optimized by an adaptive weight adjustment method. Multi-scale node association information is integrated to finally construct a semantic propagation graph with a hierarchical structure.

[0099] When constructing the node co-occurrence matrix, an N×N co-occurrence matrix is ​​established based on the node set in the semantic space mapping, where N represents the total number of nodes. Node co-occurrence statistics employ a sliding window strategy, with a default window size of 5, configurable from 3 to 10. For each window position, the frequency of node pairs occurring within the window is counted and accumulated to the corresponding position in the co-occurrence matrix. The distance between nodes within the window affects the contribution weight, which decreases as distance increases, with a default decay factor of 0.85. After co-occurrence statistics are completed, the original frequencies are normalized, converting them into probability values ​​between 0 and 1. Semantic association strength calculation combines node co-occurrence information and node semantic feature similarity; similarity is derived by calculating the cosine distance between node semantic vectors. The semantic association strength is a comprehensive association strength matrix formed by weighting co-occurrence probability and semantic similarity in a 7:3 ratio. Elements with an association strength below the threshold of 0.05 are considered noise and set to 0 to improve matrix sparsity and reduce subsequent computational overhead.

[0100] When performing multi-round iterative propagation analysis on the node co-occurrence matrix, a gated recurrent network is used to capture the temporal dependencies between nodes. The gated recurrent network architecture contains a 64-dimensional hidden state vector, and the initial state vector of each node is obtained by linear transformation of its semantic features. The iterative propagation is set to 20 rounds. In each round, the node state is updated based on the current state and the information flow from related nodes. The information flow calculation comprehensively considers the source node state and the semantic association strength between the two nodes, and is implemented as a gated recurrent unit, which includes two components: an update gate and a reset gate. The update gate controls the degree of influence of new information on the current node state, and the reset gate adjusts the retention ratio of historical information. The propagation weight is initially equal to the semantic association strength and is dynamically adjusted during the iteration process. The weight adjustment is based on the magnitude and direction of the node state change. Node pairs with significant and consistent state changes have their propagation weight increased, with an enhancement factor of 1.2; node pairs with inconsistent or weak state changes have their weight decreased, with a decrease factor of 0.8. During the propagation process, the node state changes are recorded as a difference sequence. By analyzing the correlation of the difference sequence, the temporal dependencies between nodes are identified. Node pairs with a correlation coefficient greater than 0.6 are considered to have significant temporal dependencies and are recorded in the dependency set.

[0101] The process of determining temporal dependencies based on node state changes and generating node association sequences employs a topological sorting algorithm. A directed graph is constructed based on these temporal dependencies, where nodes are entities in the semantic space and edges represent dependencies. Topological sorting starts with nodes of zero in-degree and progressively removes processed nodes and their associated edges to generate a node sequence that satisfies dependency constraints. To handle cyclic dependencies, a cycle detection and breaking mechanism is introduced, resolving the cycle problem by deleting the weakest dependency edge. After the node association sequence is generated, optimization is performed using a node co-occurrence matrix, adjusting the arrangement of adjacent nodes to ensure that nodes with strong semantic relationships are adjacent. Optimization employs a local exchange strategy, evaluating the benefit of node exchanges within a sliding window (default size 7), with the benefit calculated based on the overall semantic coherence improvement before and after the exchange. Exchanges are only performed when the exchange benefit exceeds a threshold of 0.05 to prevent falling into local optimization traps. The optimization iteration is set to 10 rounds, with each round scanning the entire sequence from beginning to end.

[0102] When performing hierarchical clustering on node association sequences, an improved DBSCAN clustering algorithm is used. Key parameters include: neighborhood radius ε (default 0.25, indicating that nodes with semantic association strengths higher than this value are considered neighbors); and minimum number of nodes (MinPts, default 4, representing the minimum number of neighbors required to form a core node). The clustering process starts with unvisited nodes, exploring their ε-neighborhoods. If the number of neighbors is greater than or equal to MinPts, the node is marked as a core node, and the cluster is recursively expanded; otherwise, it is marked as a boundary point or noise point. To adapt to the hierarchical characteristics of the semantic space, an adaptive ε mechanism is introduced, dynamically adjusting the ε value based on the semantic density region of the node: decreasing the ε value by 10% in dense regions and increasing it by 15% in sparse regions. Hierarchical clustering is implemented recursively, first using a larger ε value (e.g., 0.4) for coarse-grained clustering, then using a smaller ε value (e.g., 0.2) for fine-grained partitioning of each generated cluster, forming a two-level hierarchical structure.

[0103] When dividing nodes into semantic clusters based on clustering results, each cluster is assigned a unique identifier and descriptive label. Simultaneously, the centroid of the cluster is calculated, represented as the weighted average of the semantic vectors of all nodes within the cluster. The weight is proportional to the node's centrality within the cluster. Centrality is calculated using the average association strength between a node and other nodes within the cluster, ranging from 0 to 1. The association degree between semantic clusters is calculated based on a combination of the semantic similarity between cluster centroids and the connection strength between cluster boundary nodes. Cluster boundary nodes are defined as nodes with a semantic association higher than a threshold of 0.15 with other cluster nodes. The association degree between two clusters A and B is calculated as the sum of two parts: the semantic similarity of the cluster centroids (weight 0.4) and the maximum association strength between the cluster boundary nodes (weight 0.6). The association degree ranges from 0 to 1. Generally, cluster pairs with an association degree greater than 0.5 are considered strongly associated, those between 0.2 and 0.5 are considered moderately associated, and those below 0.2 are considered weakly associated.

[0104] When establishing the initial propagation path based on the semantic association strength in the node co-occurrence matrix, an improved Dijkstra's shortest path algorithm is used. For any node pair (i,j), the reciprocal of the semantic association strength is used as the edge weight to calculate the shortest path from i to j. Path search prioritizes edges with high semantic association strength, while introducing a path length limit to avoid generating excessively long paths. The default maximum path length is 6, which can be adjusted according to the total number of nodes; the maximum length increases by 1 for every 1000 increase in the total number of nodes. The initial propagation path set contains all node pair paths that meet the conditions, forming a weighted directed graph structure. To reduce redundancy, the path set is pruned, retaining the top 60% of paths by association strength. The path weight is defined as the geometric mean of the semantic association strength of all edges on the path, ensuring that long paths do not become excessively weighted due to cumulative effects.

[0105] When constructing a node propagation subgraph by combining local neighborhood information, for each node, the top K nodes with the highest semantic relevance are selected to form its local neighborhood. The value of K is dynamically set according to the importance of the node in the semantic space; nodes with higher importance have larger K values, ranging from 5 to 15. Importance assessment is based on a comprehensive calculation of the node's degree centrality and proximity centrality. Within the local neighborhood, the edge weights retain the original semantic relevance strength, forming the node's propagation subgraph. Each propagation subgraph is a subset of the original semantic network, maintaining the integrity of local semantic relevance.

[0106] When iteratively optimizing the node propagation subgraph using adaptive weight adjustment, a node influence propagation model is introduced. Initially, the influence value of each node is set to 1.0, and the weight of the edges is the original semantic association strength. The iterative optimization is performed for 15 rounds, each round including two steps: influence propagation and weight adjustment. Influence propagation follows the principle of energy conservation; a node distributes a portion of its influence (default 80%) to connected nodes according to the weight of its outgoing edges, retaining the remaining portion (20%). Weight adjustment is based on feedback from changes in node influence. When the influence transfer between two nodes is significantly higher than expected, the weight of the corresponding edge is increased; when it is significantly lower than expected, the weight of the corresponding edge is decreased. The iteration is terminated early when the rate of change is less than 1% for three consecutive rounds, indicating that convergence has been achieved.

[0107] The process of constructing a hierarchical semantic propagation graph by integrating multi-scale node association information consists of three stages. The first stage merges all optimized node propagation subgraphs to form an initial fused graph, with edge weights set to the maximum value of the corresponding edge weights across all subgraphs. The second stage introduces a semantic cluster hierarchy, constructing an inter-cluster relationship layer and an intra-cluster relationship layer. The inter-cluster relationship layer connects the central nodes of different semantic clusters, with edge weights equal to the inter-cluster association degree. The intra-cluster relationship layer retains high-weight edges (top 40%) from the original node propagation subgraphs to ensure semantic flow within the cluster. The third stage integrates the two layers through cross-layer connections. Cross-layer connections are established from the cluster center node to important intra-cluster nodes (top 20% in influence), with weights equal to the product of the semantic association strength between the two nodes and the importance of the intra-cluster node. The final generated semantic propagation graph contains two layers: the original node layer and the semantic cluster layer, as well as two types of edges: intra-layer connections and cross-layer connections, forming a complete hierarchical semantic propagation structure.

[0108] In one optional implementation, receiving a retrieval request and performing semantic parsing, matching the parsing results with node information in the semantic propagation graph, and generating an initial candidate result set includes:

[0109] The system receives a retrieval request, extracts feature information from the request, performs distributed Monte Carlo sampling on the feature information, dynamically adjusts the sampling weights during the sampling process, maps the weighted sampling path to a semantic feature vector, optimizes the semantic feature vector through multiple rounds of iteration to obtain the first semantic vector of the retrieval request, and generates a multi-scale feature representation of the first semantic vector.

[0110] Node information is extracted from the semantic propagation graph, the node information is converted into a second semantic vector, and a multi-level node feature index is constructed based on the second semantic vector;

[0111] An adaptive beam search is performed in the multi-level node feature index. The multi-scale feature representation is used to perform feature matching at different levels. The matching scores of each level are accumulated to generate a path score. The matching path with the highest score is selected as the optimal matching path. Node information is collected along the optimal matching path. The collected node information is sorted according to the connection strength of the nodes in the semantic propagation graph to generate an initial candidate result set.

[0112] The retrieval request receiving module receives retrieval requests sent by clients via the HTTP protocol. The request format supports both JSON and XML standard formats. JSON format requests include three main fields: query, filter, and limit. The query is a string storing the user-input retrieval text; the filter is an object type containing optional filtering conditions such as time range and type; and the limit is an integer with a default value of 20, representing the maximum number of results to return. Upon receiving a retrieval request, a request verification step checks the completeness of necessary fields and the validity of the format. The request length is limited to 5 to 200 characters. For requests exceeding this length, the first 200 characters are truncated; for requests that are too short, an error message is returned. After successful request verification, the feature information extraction stage begins. Feature information extraction employs a combined strategy of word segmentation and tagging, splitting the retrieval text into a sequence of tokens while simultaneously tagging linguistic features such as part-of-speech and named entity types. The word segmentation precision is set to bidirectional maximum matching with a recall-first mode, focusing on identifying technical terms and entity names to form a preliminary feature sequence. Each token in the feature sequence contains five attribute fields: the word itself, location information, part-of-speech tagging, entity type, and importance score. Importance scores are calculated based on the term frequency-inverse document frequency algorithm, with values ​​ranging from 0 to 5. Higher scores indicate a greater contribution of the term to the semantics of the retrieval.

[0113] When performing distributed Monte Carlo sampling on the extracted feature information, a multi-node parallel computing architecture is adopted, with the computing cluster configured with 8 computing nodes. A message queue mechanism coordinates task allocation and result merging. Each node is responsible for performing 1000 rounds of sampling, starting with the top 3 words with the highest importance scores in the feature sequence. The sampling process uses an improved non-uniform random walk algorithm, where each step starts from the current word and moves to the next word based on semantic association probability. The initial sampling weights are proportional to the word importance scores, and the weight vector is normalized using softmax to convert it into a probability distribution. During the sampling process, the mechanism for dynamically adjusting the sampling weights is based on a Markov decision process, introducing a state value evaluation function. The state value function combines two indicators: semantic coherence and diversity of the current sampling path. Coherence is calculated using the semantic similarity of adjacent words in the path, and diversity is measured by Shannon entropy to measure the uniformity of word distribution in the path. When coherence falls below a threshold of 0.3 or diversity falls below a threshold of 0.5, the weight adjustment mechanism is triggered. The weight adjustment follows a temperature annealing strategy, with an initial temperature T set to 1.0. After every 100 rounds of sampling, the temperature decays to 0.9 times its original value, with a minimum temperature limit of 0.1. Higher temperatures lead to more diversified sampling; lower temperatures favor higher-value paths.

[0114] The process of mapping weighted sampling paths to semantic feature vectors utilizes an attention mechanism. Each sampling path is represented as a sequence of words, with a length ranging from 5 to 15. Paths exceeding this range are adjusted to the standard length through truncation or padding. Each word in the path is converted into a 768-dimensional vector representation using a pre-trained word embedding model. The attention layer calculates the semantic correlation between words, which determines the contribution weight of each word in the semantic feature vector. The attention calculation uses a scaled dot product mode, with the scaling factor set to the square root of the vector dimension, i.e., 27.7. After softmax normalization, the attention weights are used to weightedly combine word vectors to generate a semantic feature vector representing the entire path. The initial semantic feature vector has a dimension of 768, which is further refined through multiple rounds of iterative optimization. The iterative optimization process uses a contrastive learning framework to construct positive and negative sample pairs. Positive samples are path vector pairs with similar semantics, while negative samples are path vector pairs with significant semantic differences. The optimization objective is to maximize the similarity of positive sample pairs while minimizing the similarity of negative sample pairs. The optimization process was performed for 10 rounds, with an initial learning rate of 0.001, which decreased to 0.9 times the original value in each round. The convergence criterion was that the similarity change rate was less than 0.5% for three consecutive rounds. After optimization, the semantic feature vectors of all paths were merged, and the first semantic vector of the retrieval request was obtained by weighted averaging, maintaining a dimension of 768.

[0115] The multi-scale feature representation generation employs a feature pyramid network structure. Based on the first semantic vector, it constructs feature representations at different resolutions through convolution and pooling operations, forming four scale levels. The highest resolution layer retains the original 768-dimensional features; the second layer reduces the dimensions to 384 through max pooling; the third layer further reduces it to 192 dimensions; and the lowest resolution layer is a simplified 96-dimensional representation. Feature vectors from each layer are fused with contextual information through residual connections to ensure information flow between different scales. The design of the multi-scale feature representation aims to adapt to the hierarchical structure of node feature indexes. Higher-level index matching uses low-dimensional features to improve retrieval speed; lower-level fine-grained matching uses high-dimensional features to ensure matching accuracy. Feature vectors at each scale are L2 normalized to ensure a vector magnitude of 1, facilitating subsequent similarity calculations.

[0116] When extracting node information from the semantic propagation graph, attribute data for all nodes is obtained through a graph traversal algorithm. The semantic propagation graph uses an adjacency list storage structure, and each node contains four basic attributes: a unique identifier, original content, a semantic feature vector, and connection relationships. Node traversal uses a breadth-first strategy to ensure that related nodes are processed consecutively, improving cache hit rate. The extracted node information undergoes data cleaning and standardization, including steps such as removing redundant attributes, standardizing numerical ranges, and filling in missing values. The semantic feature vector of the original node content has a dimension of 768, consistent with the first semantic vector of the retrieval request, facilitating direct comparison. The process of converting node information into a second semantic vector is similar to the feature processing flow of the retrieval request, the difference being that the node information already possesses structured features, eliminating the need for additional word segmentation and annotation. After conversion, each node corresponds to a standardized second semantic vector, used for subsequent construction of a multi-level node feature index.

[0117] The multi-level node feature index is constructed based on hierarchical clustering results, corresponding to the multi-scale feature representation of the retrieval request. The index structure contains four levels, with increasing precision and complexity from the top to the bottom. The top-level index uses a 96-dimensional feature vector, dividing the semantic space into 200 buckets using Locality Sensitive Hash (LSH) technology; the second level uses 192-dimensional features, subdivided into 800 sub-regions; the third level uses 384-dimensional features, further subdivided into 3200 fine-grained regions; the bottom level uses complete 768-dimensional features, directly storing the original node information. Index construction adopts a batch processing mode, building all nodes at once when the number is less than 100,000, and using an incremental construction strategy when the number is greater than 100,000, processing 50,000 nodes per batch. The index update strategy combines scheduled full updates and real-time incremental updates, with a full update cycle of once a day and an incremental update latency of no more than 30 seconds. Index storage adopts a hybrid memory and disk scheme, with hot data stored in memory and cold data stored on solid-state drives, with memory usage capped at 70% of available physical memory.

[0118] When performing adaptive beam search in a multi-level node feature index, the multi-scale feature representation of the retrieval request is used as the query condition. Starting from the top-level index, the matching is refined layer by layer downwards. The beam width parameter k is initially set to 20 and dynamically adjusted as the level deepens. The top level maintains a wider beam width to cover more candidate regions, while the bottom level narrows the beam width to improve accuracy. Specifically, the beam width is k for the first layer, 0.8k for the second, 0.5k for the third, and 0.3k for the fourth. The value of k can be dynamically adjusted according to server load, ranging from 10 to 50. Feature matching uses cosine similarity to calculate the similarity between the retrieval feature and the index node feature, with a value ranging from -1 to 1. The closer to 1, the more semantically similar the features. Each matching layer retains the top k candidate nodes with the highest similarity to proceed to the next layer. The matching scores from different levels are accumulated into a path score through a weighted average, with the weight distribution being {top level: 0.1, second level: 0.2, third level: 0.3, bottom level: 0.4}, reflecting the importance of the bottom-level matching results. The score is accumulated using logarithmic space addition to prevent numerical underflow, and the final score is restored to the original range through exponential transformation.

[0119] After selecting the highest-scoring matching path as the optimal matching path, node information is collected along the path. This collection process includes not only direct nodes along the path but also their one-hop neighbors in the semantic propagation graph, enriching the candidate result set. For each collected node, its basic attributes (identifier, content, feature vector) and relational attributes (semantic similarity to the query node, in-degree and out-degree in the graph, and connection strength) are extracted. Connection strength is defined as the weight of the edges between nodes, ranging from 0 to 1; a larger value indicates a stronger connection. When sorting nodes based on their connection strength in the semantic propagation graph, a comprehensive scoring mechanism is used. The scoring formula combines semantic similarity (weight 0.6) and connection strength (weight 0.4), with a score range of 0 to 1. The sorted list of nodes constitutes the initial candidate result set. The maximum number of results is specified by the request parameters, with a default value of 20. Each entry in the result set contains four basic fields: node identifier, original content summary, matching score, and type label, and is returned to the client in JSON format.

[0120] In one optional implementation, the initial candidate result set is reordered according to the node hierarchy, and filtered based on semantic similarity and node importance to obtain the final matching results, including:

[0121] A directed reordering graph is constructed based on the node hierarchy. A distributed Monte Carlo tree search is performed on the directed reordering graph. The node importance score is determined by the node access frequency. At the same time, the semantic similarity between adjacent nodes is calculated. The node importance score and the semantic similarity are combined to generate the node ranking weight.

[0122] The nodes in the initial candidate result set are grouped according to the node sorting weight to form multiple candidate result clusters;

[0123] Within each candidate result cluster, a reordering path is constructed. The optimal reordering sequence is determined based on the cumulative sorting weight of the nodes on the reordering path. The candidate results are then rearranged along the optimal reordering sequence. A hierarchical screening is performed on the rearranged candidate results in combination with a preset sorting weight threshold to obtain the final matching result.

[0124] A directed reordering graph is constructed based on the hierarchical relationships between nodes. Each node in the initial candidate result set is treated as a vertex in the graph, and directed connections are established according to the hierarchical relationships between nodes. These hierarchical relationships can be determined through node attribute labels or predefined domain knowledge. For example, in a knowledge graph matching scenario, concept nodes point to instance nodes, and higher-level classes point to lower-level classes, forming a clear hierarchical structure. After construction, a directed graph structure reflecting the hierarchical dependencies between nodes is obtained.

[0125] A distributed Monte Carlo tree search is performed on the constructed directed reordered graph. The search process starts from the root node and iterates through four phases: selection, expansion, simulation, and backpropagation. In the selection phase, the UCT (Upper Confidence Bound for Trees) formula is used to select the next node to be explored, balancing the node's current value and number of visits. In the expansion phase, the unvisited child nodes of the selected node are added to the search tree. The simulation phase starts from the newly added node and executes randomly according to a predetermined strategy until the termination state. The backpropagation phase propagates the simulation results back along the search path, updating the node's visit count and value. This process is repeated until the preset number of iterations or computational resource limits are reached.

[0126] Distributed Monte Carlo tree search is used to count the frequency of visits to each node. After normalizing the visit frequency, a node importance score is obtained. Nodes with higher visit frequency occupy more critical positions in the graph structure and have higher importance scores. Simultaneously, semantic similarity between adjacent nodes in the directed graph is calculated. Semantic similarity is calculated using the cosine similarity of node representation vectors. These node representation vectors are obtained through a pre-trained semantic encoder trained on domain-specific corpora, capable of capturing subtle semantic relationships between nodes.

[0127] The node importance score and semantic similarity are combined to generate the node ranking weight. The combination method uses a weighted summation, i.e., ranking weight = α × node importance score + (1-α) × semantic similarity, where α is a weight adjustment parameter, with a value between 0 and 1, which can be adjusted according to the specific application scenario. In practical applications, the optimal value of α can be determined through validation set tuning.

[0128] The nodes in the initial candidate result set are grouped according to their ranking weights to form multiple candidate result clusters. Grouping employs a density-based clustering algorithm, grouping nodes with similar ranking weights and close graph distances together. Specifically, all nodes are first sorted in descending order of their ranking weights. Then, a ranking weight difference threshold β and a graph distance threshold γ are set. If the ranking weight difference between two nodes does not exceed β and their graph distance does not exceed γ, they are grouped into the same cluster.

[0129] Within each candidate result cluster, a reordering path is constructed. Starting from the highest-weighted node in each cluster, a greedy strategy is used to select the next optimal connection node, prioritizing the adjacent node with the largest product of its sorting weight with the current path's endpoint to form a reordering path. If multiple candidate nodes have similar weights (the difference is less than a preset threshold δ), the node with higher semantic similarity to the path's starting point is selected. The path construction process continues until all nodes within the cluster are included in the path or the maximum path length limit is reached.

[0130] The optimal reordering sequence is determined based on the cumulative ranking weights of nodes along the reordering path. The cumulative ranking weights are calculated as a weighted sum of the ranking weights of all nodes on the path, with nodes closer to the path's starting point receiving a higher decay coefficient. This weighting method ensures that the early nodes of the path have a greater impact on the final reordering result. The path with the highest cumulative ranking weight is selected as the optimal reordering sequence.

[0131] The candidate results are rearranged along the optimal reordering sequence. The initial candidate results are adjusted according to the order of appearance of nodes in the optimal reordering sequence, so that important nodes and nodes with high semantic relevance get higher ranking positions.

[0132] A hierarchical screening process is applied to the rearranged candidate results using preset ranking weight thresholds. Three ranking weight thresholds (λ1>λ2>λ3) are set, corresponding to different levels of screening rigor. First, all nodes with ranking weights higher than λ1 are retained. Then, nodes with ranking weights between λ1 and λ2 are checked. If their average semantic similarity with the retained nodes exceeds a preset threshold ε1, they are retained. For nodes with ranking weights between λ2 and λ3, if their maximum semantic similarity with the retained nodes exceeds a preset threshold ε2, they are retained. Nodes with ranking weights lower than λ3 are discarded. This hierarchical screening mechanism ensures the retention of important nodes while avoiding semantic redundancy, ultimately achieving a matching result that balances accuracy and diversity.

[0133] In practical applications, such as medical knowledge graph matching tasks, disease nodes can be used as query nodes, while related symptoms and treatment methods can be used as candidate nodes. Through the aforementioned reordering and filtering mechanisms, the medical knowledge points most relevant to the queried disease can be efficiently identified, while maintaining the semantic diversity of the results, providing more comprehensive reference information for medical decision-making.

[0134] A second aspect of the present invention provides an intelligent retrieval and precise matching system based on multi-level semantic decomposition, comprising:

[0135] The acquisition unit is used to acquire the text to be retrieved, perform semantic decomposition on the text to be retrieved, and generate a hierarchical semantic representation tree in a multidimensional semantic feature space.

[0136] The optimization unit is used to perform node traversal on the hierarchical semantic representation tree based on an adaptive probability sampling strategy, calculate the semantic association strength between nodes, construct a semantic propagation graph, perform iterative optimization on the semantic propagation graph to obtain the optimal semantic path, and reorganize the semantic nodes according to the optimal semantic path to form an optimized hierarchical semantic representation tree.

[0137] The extraction unit is used to extract semantic feature vectors that incorporate contextual dependencies from the optimized hierarchical semantic representation tree and establish a semantic feature index.

[0138] The filtering unit is used to receive retrieval requests and perform semantic parsing, match the parsing results with the node information in the semantic propagation graph, generate an initial candidate result set, reorder the initial candidate result set according to the node hierarchy, and filter the results by combining semantic similarity and node importance to obtain the final accurate matching result.

[0139] A third aspect of the present invention provides an electronic device, comprising:

[0140] processor;

[0141] Memory used to store processor-executable instructions;

[0142] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0143] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0144] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent retrieval and precise matching method based on multi-level semantic decomposition, characterized in that, include: The text to be retrieved is obtained, and the text to be retrieved is semantically decomposed to generate a hierarchical semantic representation tree in a multidimensional semantic feature space. The hierarchical semantic representation tree is traversed based on an adaptive probability sampling strategy. The semantic association strength between nodes is calculated, and a semantic propagation graph is constructed. In this graph, a node co-occurrence matrix is ​​constructed, the semantic association strength between nodes is calculated, and a multi-round iterative propagation analysis is performed on the node co-occurrence matrix. In each round of iteration, the node state is updated and the propagation weight is adjusted. The temporal dependency relationship is determined based on the node state change, and a node association sequence is generated. At the same time, the structure of the node association sequence is optimized using the node co-occurrence matrix. Hierarchical clustering is performed on the node association sequences, and node semantic clusters are divided based on the clustering results. The degree of association between semantic clusters is then calculated. An initial propagation path is established based on the semantic association strength in the node co-occurrence matrix. A node propagation subgraph is constructed by combining local neighborhood information. The node propagation subgraph is iteratively optimized by an adaptive weight adjustment method. Multi-scale node association information is integrated to finally construct a semantic propagation graph with a hierarchical structure. Iterative optimization is performed on the semantic propagation graph to obtain the optimal semantic path. The semantic nodes are then reorganized according to the optimal semantic path to form an optimized hierarchical semantic representation tree. Semantic feature vectors that incorporate contextual dependencies are extracted from the optimized hierarchical semantic representation tree to establish a semantic feature index; The system receives retrieval requests and performs semantic parsing. It then matches the parsing results with the node information in the semantic propagation graph to generate an initial candidate result set. The initial candidate result set is then reordered according to the node hierarchy, and filtered based on semantic similarity and node importance to obtain the final accurate matching result.

2. The method according to claim 1, characterized in that, Semantic decomposition of the text to be retrieved, and generation of a hierarchical semantic representation tree in a multidimensional semantic feature space, includes: The text to be retrieved is input into the deep semantic processor, which extracts the semantic features from the text to obtain the initial semantic feature sequence; Distributed Monte Carlo tree sampling is performed on the initial semantic feature sequence to calculate the semantic correlation between feature nodes. A semantic propagation network is constructed based on the semantic correlation, and the node weights in the semantic propagation network are updated through iterative optimization to identify the hierarchical semantic structure of the text. Based on the hierarchical semantic structure, semantic levels are divided, and a multidimensional semantic feature space is constructed. An adaptive bundle search is performed in the multidimensional semantic feature space to dynamically adjust the search width, map semantic features to different levels, and perform feature fusion according to the importance of features to generate a hierarchical semantic representation tree with conceptual features, relational features, and attribute features.

3. The method according to claim 2, characterized in that, Distributed Monte Carlo tree sampling is performed on the initial semantic feature sequence. Calculating the semantic association degree between feature nodes and constructing a semantic propagation network based on the semantic association degree includes: Feature node information is extracted from the initial semantic feature sequence. Multiple rounds of random walk sampling are performed on the feature node information based on a distributed Monte Carlo tree. The cumulative reward value of the sampling path in each round is calculated. The node transition probability is determined based on the cumulative reward value. The node transition probability is optimized by a temperature annealing strategy to obtain the semantic correlation between feature nodes. The feature node information is grouped according to semantic relevance to form multiple semantic clusters; Based on the semantic relevance, an initial topology of the semantic propagation network is constructed. Adaptive weight propagation is performed on the initial topology, and the network connection strength is iteratively updated through local neighborhood expansion to finally generate a semantic propagation network with a hierarchical structure.

4. The method according to claim 1, characterized in that, Based on an adaptive probability sampling strategy, the hierarchical semantic representation tree is traversed node by node, the semantic association strength between nodes is calculated, and a semantic propagation graph is constructed, including: The node traversal order is determined based on a distributed Monte Carlo search strategy, and adaptive probability sampling is performed on each traversed node. The sampling probability is dynamically adjusted according to the distribution of nodes in different semantic levels to generate a node sampling sequence. Semantic feature vectors are extracted from the node sampling sequence to construct a semantic space mapping; The semantic association strength between nodes is calculated in the semantic space mapping. Temporal dependency analysis of the semantic association strength is performed through a gated recurrent network to establish node propagation paths. A semantic propagation graph is constructed based on local neighborhood aggregation.

5. The method according to claim 1, characterized in that, The process involves receiving a retrieval request, performing semantic parsing, matching the parsing results with node information in the semantic propagation graph, and generating an initial candidate result set, including: The system receives a retrieval request, extracts feature information from the request, performs distributed Monte Carlo sampling on the feature information, dynamically adjusts the sampling weights during the sampling process, maps the weighted sampling path to a semantic feature vector, optimizes the semantic feature vector through multiple rounds of iteration to obtain the first semantic vector of the retrieval request, and generates a multi-scale feature representation of the first semantic vector. Node information is extracted from the semantic propagation graph, the node information is converted into a second semantic vector, and a multi-level node feature index is constructed based on the second semantic vector; An adaptive beam search is performed in the multi-level node feature index. The multi-scale feature representation is used to perform feature matching at different levels. The matching scores of each level are accumulated to generate a path score. The matching path with the highest score is selected as the optimal matching path. Node information is collected along the optimal matching path. The collected node information is sorted according to the connection strength of the nodes in the semantic propagation graph to generate an initial candidate result set.

6. The method according to claim 1, characterized in that, The initial candidate result set is reordered according to the node hierarchy, and then filtered based on semantic similarity and node importance to obtain the final matching results, including: A directed reordering graph is constructed based on the node hierarchy. A distributed Monte Carlo tree search is performed on the directed reordering graph. The node importance score is determined by the node access frequency. At the same time, the semantic similarity between adjacent nodes is calculated. The node importance score and the semantic similarity are combined to generate the node ranking weight. The nodes in the initial candidate result set are grouped according to the node sorting weight to form multiple candidate result clusters; Within each candidate result cluster, a reordering path is constructed. The optimal reordering sequence is determined based on the cumulative sorting weight of the nodes on the reordering path. The candidate results are then rearranged along the optimal reordering sequence. A hierarchical screening is performed on the rearranged candidate results in combination with a preset sorting weight threshold to obtain the final matching result.

7. An intelligent retrieval and precise matching system based on multi-level semantic decomposition, used to implement the method of any one of claims 1-6, characterized in that, include: The acquisition unit is used to acquire the text to be retrieved, perform semantic decomposition on the text to be retrieved, and generate a hierarchical semantic representation tree in a multidimensional semantic feature space. The optimization unit is used to perform node traversal on the hierarchical semantic representation tree based on an adaptive probability sampling strategy, calculate the semantic association strength between nodes, construct a semantic propagation graph, perform iterative optimization on the semantic propagation graph to obtain the optimal semantic path, and reorganize the semantic nodes according to the optimal semantic path to form an optimized hierarchical semantic representation tree. The extraction unit is used to extract semantic feature vectors that incorporate contextual dependencies from the optimized hierarchical semantic representation tree and establish a semantic feature index. The filtering unit is used to receive retrieval requests and perform semantic parsing, match the parsing results with the node information in the semantic propagation graph, generate an initial candidate result set, reorder the initial candidate result set according to the node hierarchy, and filter the results by combining semantic similarity and node importance to obtain the final accurate matching result.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images