Intelligent matching method and device for number source based on hybrid semantic model

Abstract

The present disclosure provides a number source intelligent matching method and device based on a hybrid semantic model, and relates to the technical field of artificial intelligence. In some embodiments of the present disclosure, structured prompt words are constructed according to obtained form structured metadata; the structured prompt words are input into a large language model for semantic reasoning to generate business semantic enhanced descriptions; the business semantic enhanced descriptions are input into a deep feature extraction network to obtain business semantic vectors; a heterogeneous attribute graph is constructed according to obtained structured metadata of multi-source heterogeneous data resources in a distributed environment; neighborhood topological information of nodes in the metadata heterogeneous graph is aggregated through a graph convolution coding network to obtain a structure-aware data source technology semantic vector; the business semantic vector and the data source technology semantic vector are respectively mapped to a shared latent space; the business semantic relationship between multi-candidate fields is determined in combination with reasoning of a domain ontology knowledge graph and a large language model; and an aggregation operator is generated according to the business semantic relationship.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a method and apparatus for intelligent matching of data sources based on a hybrid semantic model. Background Technology

[0002] With the deepening of digital transformation, business terminals, as the core carriers connecting users and data, urgently need to achieve automated association between business metrics and underlying multi-source heterogeneous data assets. Although a data resource pool covering the entire business domain has been built at the underlying level, the "semantic gap" remains significant: the business side is accustomed to using natural language to describe metrics, while the storage side uses technical specification naming, making it difficult for the two sides to be directly aligned.

[0003] Currently, the field matching accuracy between business metrics and underlying multi-source heterogeneous data is low, resulting in low data retrieval efficiency. Summary of the Invention

[0004] This disclosure provides a data source intelligent matching method and apparatus based on a hybrid semantic model, which at least solves the problems of low field matching accuracy and low data retrieval efficiency between existing business indicators and underlying multi-source heterogeneous data.

[0005] The technical solution disclosed herein is as follows: This disclosure provides a data source intelligent matching method based on a hybrid semantic model, including: Based on the obtained structured metadata of the form, structured prompt words are constructed; the structured prompt words are then input into a large language model for semantic reasoning to generate a business semantic enhancement description; the business semantic enhancement description is then input into a deep feature extraction network to obtain a business semantic vector. Based on the structured metadata of multi-source heterogeneous data resources obtained in a distributed environment, a heterogeneous attribute graph is constructed; the neighborhood topology information of the nodes in the heterogeneous metadata graph is aggregated through a graph convolutional coding network to obtain a structure-aware data source technology semantic vector. The business semantic vector and the data source technology semantic vector are respectively mapped to a shared latent space; candidate fields are recalled based on the vector similarity in the shared latent space. When the candidate fields meet the preset triggering conditions, the business semantic relationship between multiple candidate fields is determined by combining the reasoning of the domain ontology knowledge graph and the large language model; and an aggregation operator is generated based on the business semantic relationship to complete the data assembly.

[0006] Optionally, the method further includes: The manual correction operations of business personnel on the matching results are obtained, and the manual correction operations are divided into entity-level feedback, rule-level feedback or meta-knowledge feedback through a correction type identifier. For the entity-level feedback, a weight update strategy based on a time decay function is used to adjust the entity confidence score; For the rule-level feedback, a versioned management strategy is adopted, and when the number of accumulated feedback samples is greater than or equal to the first preset threshold, a low-rank incremental fine-tuning method is used to update the model parameters. For the aforementioned meta-knowledge feedback, if the number of accumulated feedback samples is greater than or equal to a second preset threshold, manual review is triggered.

[0007] Optionally, the form structured metadata includes: field identifiers, display labels, field hierarchy paths, filling instructions, statistical definition text, neighbor field sets, and historical fill example values.

[0008] Optionally, the enhanced business semantic description includes: the true meaning of the form field in the business context, statistical caliber, typical value range, and potential related concepts.

[0009] Optionally, the heterogeneous attribute graph includes: repository nodes, dataset nodes, data element nodes, and data service interface nodes.

[0010] Optionally, the step of aggregating the neighborhood topology information of nodes in the heterogeneous metadata graph through a graph convolutional coding network to obtain a structure-aware data source technology semantic vector includes: The graph convolutional coding network obtains the node embedding matrix by aggregating neighborhood topology information through multi-layer graph convolution; The output of the last layer graph convolution is concatenated with the field text embedding and then linearly transformed to generate the data source technology semantic vector of the same dimension as the business semantic vector.

[0011] Optionally, mapping the business semantic vector and the data source technology semantic vector to a shared latent space, respectively; and recalling candidate fields based on vector similarity in the shared latent space, includes: The business semantic vector is mapped using a nonlinear dimensionality reduction network to obtain a low-dimensional representation on the business side; The semantic vector of the data source technology is mapped using a single-layer linear projection to obtain a low-dimensional representation of the data source side; Calculate the cosine similarity based on the business-side low-dimensional representation and the data source-side low-dimensional representation; Based on the cosine similarity, the candidate fields are recalled.

[0012] Optionally, the step of determining the business semantic relationship between multiple candidate fields by combining the reasoning of the domain ontology knowledge graph and the large language model includes: The system queries a pre-built business domain ontology library to perform preliminary semantic relationship judgments on multiple candidate fields based on complementarity, cross-references, and authoritative coverage, and obtains the preliminary judgment results from the ontology library. The business meaning and statistical caliber of the multiple candidate fields are input into the large language model for verification and reasoning to obtain the reasoning result of the large language model. Based on the initial judgment results of the ontology and the reasoning results of the large language model, the business semantic relationships between multiple candidate fields are determined.

[0013] Optionally, determining the business semantic relationship between multiple candidate fields based on the initial judgment result of the ontology database and the inference result of the large language model includes: If the ontology library has clear annotations and the initial judgment result of the ontology library is consistent with the reasoning result of the large language model, then the initial judgment result of the ontology library shall be adopted. If the ontology library has clear annotations and the initial judgment result of the ontology library is consistent with the reasoning result of the large language model, then manual processing is triggered and it is marked as a potential ontology conflict. If the ontology is not explicitly labeled and the reasoning confidence of the large language model is greater than or equal to a preset threshold, then the reasoning result of the large language model is adopted and the candidate update pool of the ontology is updated. If the ontology is not explicitly labeled and the reasoning confidence of the large language model is less than the preset threshold, manual processing is triggered and the relationship is marked as unknown.

[0014] This disclosure also provides a data source intelligent matching device based on a hybrid semantic model, comprising: The construction module is used to construct structured prompt words based on the obtained structured metadata of the form; input the structured prompt words into a large language model for semantic reasoning to generate a business semantic enhancement description; input the business semantic enhancement description into a deep feature extraction network to obtain a business semantic vector. The aggregation module is used to construct a heterogeneous attribute graph based on the structured metadata of multi-source heterogeneous data resources acquired in a distributed environment; and to aggregate the neighborhood topology information of the nodes in the heterogeneous metadata graph through a graph convolutional coding network to obtain a structure-aware data source technology semantic vector. The mapping module is used to map the business semantic vector and the data source technology semantic vector to a shared latent space respectively; and to recall candidate fields based on the vector similarity in the shared latent space. The generation module is used to determine the business semantic relationship between multiple candidate fields by combining the reasoning of the domain ontology knowledge graph and the large language model when the candidate fields meet the preset trigger conditions; and to generate an aggregation operator based on the business semantic relationship to complete the data assembly.

[0015] The technical solutions provided by the embodiments of this disclosure bring at least the following beneficial effects: In some embodiments of this disclosure, structured prompt words are constructed based on the acquired structured metadata of the form; the structured prompt words are then input into a large language model for semantic reasoning to generate a business semantic enhancement description; the business semantic enhancement description is input into a deep feature extraction network to obtain a business semantic vector; a heterogeneous attribute graph is constructed based on the acquired structured metadata of multi-source heterogeneous data resources in a distributed environment; neighborhood topology information of nodes in the heterogeneous metadata graph is aggregated through a graph convolutional coding network to obtain a structure-aware data source technology semantic vector; the business semantic vector and the data source technology semantic vector are mapped to a shared latent space; candidate fields are recalled based on vector similarity in the shared latent space; when a candidate field meets a preset triggering condition, the business semantic relationship between multiple candidate fields is determined by combining the domain ontology knowledge graph and the reasoning of the large language model; and an aggregation operator is generated based on the business semantic relationship to complete data assembly; this disclosure combines business semantic reasoning using a large language model and data assembly based on the business semantic relationship between candidate fields to improve the field matching accuracy between business indicators and underlying multi-source heterogeneous data, thereby improving data retrieval efficiency.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0018] Figure 1 A flowchart illustrating an intelligent data source matching method based on a hybrid semantic model, provided as an exemplary embodiment of this disclosure; Figure 2 A flowchart of asymmetric bidirectional alignment and candidate recall provided for exemplary embodiments of this disclosure; Figure 3 A decision logic diagram for a dynamic semantic assembly engine provided as an exemplary embodiment of this disclosure; Figure 4 A differentiated adaptive feedback optimization logic diagram provided for an exemplary embodiment of this disclosure; Figure 5 A schematic diagram of the structure of a data source intelligent matching device based on a hybrid semantic model, provided as an exemplary embodiment of this disclosure; Figure 6 A schematic diagram of the structure of an electronic device provided for an exemplary embodiment of this disclosure. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0020] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.

[0021] It should be noted that the user information involved in this disclosure includes, but is not limited to, user device information and user personal information; the collection, storage, use, processing, transmission, provision and disclosure of user information in this disclosure all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0022] To address the aforementioned technical issues, in some embodiments of this disclosure, structured prompt words are constructed based on the acquired structured metadata of the form; these structured prompt words are then input into a large language model for semantic reasoning to generate a business semantic enhancement description; the business semantic enhancement description is input into a deep feature extraction network to obtain a business semantic vector; a heterogeneous attribute graph is constructed based on the acquired structured metadata of multi-source heterogeneous data resources in a distributed environment; neighborhood topology information of nodes in the heterogeneous metadata graph is aggregated through a graph convolutional coding network to obtain a structure-aware data source technology semantic vector; the business semantic vector and the data source technology semantic vector are mapped to a shared latent space; candidate fields are recalled based on vector similarity in the shared latent space; when a candidate field meets a preset triggering condition, the business semantic relationship between multiple candidate fields is determined by combining the domain ontology knowledge graph and the reasoning of the large language model; and an aggregation operator is generated based on the business semantic relationship to complete data assembly. This disclosure combines business semantic reasoning using a large language model and data assembly based on the business semantic relationship between candidate fields to improve the field matching accuracy between business indicators and underlying multi-source heterogeneous data, thereby improving data retrieval efficiency.

[0023] The technical solutions provided by the embodiments of this disclosure are described in detail below with reference to the accompanying drawings.

[0024] Figure 1 This is a flowchart illustrating an intelligent data source matching method based on a hybrid semantic model, provided as an exemplary embodiment of this disclosure. Figure 1 As shown, the method includes: S101: Based on the obtained structured metadata of the form, construct structured prompt words; input the structured prompt words into the large language model for semantic reasoning to generate a business semantic enhancement description; input the business semantic enhancement description into the deep feature extraction network to obtain the business semantic vector; S102: Based on the structured metadata of the multi-source heterogeneous data resources obtained in the distributed environment, construct a heterogeneous attribute graph; aggregate the neighborhood topology information of the nodes in the heterogeneous metadata graph through a graph convolutional coding network to obtain the structure-aware data source technology semantic vector. S103: Map the business semantic vector and the data source technology semantic vector to the shared latent space respectively; recall candidate fields based on the vector similarity in the shared latent space; S104: When a candidate field meets the preset triggering conditions, the business semantic relationship between multiple candidate fields is determined by combining the reasoning of the domain ontology knowledge graph and the large language model; and an aggregation operator is generated based on the business semantic relationship to complete the data assembly.

[0025] In this embodiment, the subject executing the above method is a terminal device or a server.

[0026] The terminal device includes, but is not limited to, mobile stations (MS), mobile terminals, mobile phones, handsets, and portable equipment. This terminal device can communicate with one or more core networks via a radio access network (RAN). For example, the terminal device can be a mobile phone (or "cellular" phone), a computer with wireless communication capabilities, a computer with wireless transceiver capabilities, a virtual reality (VR) terminal device, an AR terminal device, a wireless terminal in industrial control, a wireless terminal in self-driving, a wireless terminal in remote medical care, a wireless terminal in a smart grid, a wireless terminal in transportation safety, a wireless terminal in a smart city, a wireless terminal in a smart home, etc. The operating systems installed on the terminal device include, but are not limited to, iOS, Android, Windows, Linux, and Mac OS. In different networks, terminals may be called by different names, such as: user equipment, mobile station, user unit, station, cellular phone, personal digital assistant, wireless modem, wireless communication device, handheld device, laptop, cordless phone, wireless local loop station, television, etc. For ease of description, this embodiment will simply refer to it as terminal device.

[0027] In this embodiment, the implementation form of the server is not limited. For example, the server can be a conventional server, a cloud server, a cloud host, a virtual center, or other server devices. The server mainly consists of a processor, hard disk, memory, system bus, and other common computer architecture types.

[0028] In this embodiment, structured prompt words are constructed based on the acquired structured metadata of the form; these prompt words are then input into a large language model for semantic reasoning to generate a business semantic enhancement description; the business semantic enhancement description is input into a deep feature extraction network to obtain a business semantic vector; a heterogeneous attribute graph is constructed based on the acquired structured metadata of multi-source heterogeneous data resources in a distributed environment; neighborhood topology information of nodes in the heterogeneous metadata graph is aggregated through a graph convolutional coding network to obtain a structure-aware data source technology semantic vector; the business semantic vector and the data source technology semantic vector are mapped to a shared latent space; candidate fields are recalled based on vector similarity in the shared latent space; when a candidate field meets a preset triggering condition, the business semantic relationship between multiple candidate fields is determined by combining the domain ontology knowledge graph and the reasoning of the large language model; and an aggregation operator is generated based on the business semantic relationship to complete data assembly; this disclosure combines business semantic reasoning using a large language model and data assembly based on the business semantic relationship between candidate fields to improve the field matching accuracy between business indicators and underlying multi-source heterogeneous data, thereby improving data retrieval efficiency.

[0029] This disclosure addresses the challenges of high-precision, interpretable intelligent matching and aggregation of form requirement fields and multi-source data center fields in business form scenarios through business semantic reasoning driven by a large language model, an asymmetric bidirectional alignment architecture, and a dynamic semantic assembly engine. Specifically, it solves the problem of "one-to-N cross-source joint computation and aggregation." Furthermore, through a layered, differentiated feedback closed-loop mechanism, it achieves precise responses to different types of business corrections and adaptive system evolution.

[0030] It should be noted that the structured metadata of the form includes: field identifiers, display labels, field hierarchy paths, filling instructions, statistical caliber text, neighboring field sets, and historical fill example values. The enhanced business semantic description includes: the true meaning of the form fields in the business context, statistical caliber, typical value ranges, and potential related concepts.

[0031] In some embodiments of this disclosure, structured prompts are constructed based on the obtained form structured metadata. Specifically, the system obtains the form structured metadata from the business side and assembles the context information of the form fields into structured prompts. The context information of the form fields includes, but is not limited to: field name, the path of the form to which it belongs, filling instructions, neighboring fields, and historical filling examples.

[0032] The structured prompt templates are shown in the table below.

[0033]

[0034] In some embodiments of this disclosure, structured prompt words are input into a large language model for semantic reasoning to generate a business semantic enhancement description. Specifically, the language model, based on its pre-trained knowledge and contextual understanding capabilities, automatically infers the true meaning, statistical scope, typical value range, and potential related concepts of the field in the business context, thereby generating a structured business semantic enhancement description.

[0035] In some embodiments of this disclosure, the business semantic enhancement description is input into a deep feature extraction network to obtain a business semantic vector. Specifically, the business semantic enhancement description output by the large language model is input into the deep feature extraction network to extract a high-dimensional business semantic vector. Let the business semantic vector be V_biz, and its calculation is expressed as: V_biz = Encoder_biz(business_bridge_description) Where V_biz belongs to the d-dimensional real vector space R^d, where d is the semantic vector dimension, Encoder_biz is the pre-trained semantic encoder, and business_bridge_description is the business semantic enhancement description.

[0036] In some embodiments of this disclosure, a heterogeneous attribute graph is constructed based on the structured metadata of multi-source heterogeneous data resources acquired in a distributed environment. Specifically, the structured metadata of multi-source heterogeneous data resources in a distributed environment is integrated and modeled as a heterogeneous attribute graph G=(V, E); where, node V: encompasses multi-level data objects, including repository nodes, dataset (table) nodes, data element (field) nodes, and data service interface nodes. Edge E: defines the topological and logical relationships between entities, including structural affiliation edges, semantic equivalence edges, and logical association edges.

[0037] In some embodiments of this disclosure, a graph convolutional coding network aggregates neighborhood topology information of nodes in a heterogeneous graph of metadata to obtain a structure-aware data source technology semantic vector. One possible approach is that the graph convolutional coding network aggregates neighborhood topology information through multi-layer graph convolution to obtain a node embedding matrix; the output of the last layer of graph convolution is concatenated with the field text embedding and then linearly transformed to generate a data source technology semantic vector of the same dimension as the business semantic vector. Specifically, for each node, a multi-layer graph convolutional coding network aggregates neighborhood topology information to obtain a structure-aware field representation. Let the node embedding matrix of the k-th layer be H^(k), and its recursive formula is: H^(k) = σ( ^(-1 / 2) · ÷ ^(-1 / 2) · H^(k-1) · W^(k) ) Where H^(k) is the embedding matrix of all nodes in the k-th layer, H^(0) is the initial text embedding; à = A + I is the normalized adjacency matrix containing self-loops (A is the original adjacency matrix, I is the identity matrix). The degree matrix corresponding to à ( _ii =Σ_j Ã_ij); W^(k) is the learnable weight matrix of the k-th layer; σ is a non-linear activation function (such as ReLU).

[0038] The data source semantic vector V_src is obtained by linearly transforming the concatenation of the last layer graph convolution output and the field text embedding. V_src = Linear( Concat(H^(L), TextEmbed(field_desc)) ) Where L is the total number of layers in the graph convolutional network, and V_src belongs to the d-dimensional real vector space R^d, maintaining the same dimension as the business semantic vector V_biz.

[0039] In some embodiments of this disclosure, business semantic vectors and data source technology semantic vectors are mapped to a shared latent space; candidate fields are recalled based on vector similarity in the shared latent space. One possible approach is to map the business semantic vectors using a nonlinear dimensionality reduction network to obtain a low-dimensional representation on the business side; to map the data source technology semantic vectors using a single-layer linear projection to obtain a low-dimensional representation on the data source side; to calculate cosine similarity based on the low-dimensional representations on the business side and the data source side; and to recall candidate fields based on cosine similarity.

[0040] Figure 2 A flowchart illustrating an asymmetric bidirectional alignment and candidate recall process is provided for an exemplary embodiment of this disclosure. For example... Figure 2 As shown, specifically, this disclosure adopts the design principle of an asymmetric mapping layer. The number of fields on the business side is relatively small, but the semantic density of each field is high, requiring deep non-linear encoding; the number of fields on the data source side is large, the semantics are relatively sparse, and the structural relationships are strong, which can be effectively captured using lightweight linear transformations. This disclosure designs an asymmetric dimensionality reduction strategy, as shown in the table below:

[0041] This disclosure employs a low-dimensional shared latent space mapping. The low-dimensional representation H_biz on the business side is computed through a nonlinear dimensionality reduction network: H_biz = ReLU(W2·ReLU(W1·V_biz + b1) + b2) Where W1 and W2 are learnable weight matrices, b1 and b2 are bias vectors, and H_biz belongs to the m-dimensional real vector space R^m (m is the dimension of the low-dimensional space, typically 128).

[0042] The low-dimensional representation H_src on the data source side is calculated through a single-layer linear projection: H_src = W·V_src + b,H_src ∈ R^m The candidate field similarity score uses a cosine similarity function: Score(Q, S) = (H_biz^T·H_src) / (‖H_biz‖·‖H_src‖) This disclosure employs an asymmetric contrastive learning training strategy. The dual encoder is trained using this strategy, with manually labeled (form field, data source field) matching pairs as positive samples, and negative samples sampled from the business side to the technical side.

[0043] The loss function is defined as: L = -log [ exp(sim(H_biz, H_src + ) / τ) / Σ j exp(sim(H_biz, H_src j ) / τ) ] Where τ is the temperature hyperparameter (controlling the smoothness of the distribution, typically taking a value of 0.05~0.1); H_src + This is the positive sample technical vector (i.e., the data source field that actually matches the current business field); H_src j This is the negative sample technology vector.

[0044] This disclosure employs a vector approximate nearest neighbor retrieval engine. In the offline phase, the system pre-computes the low-dimensional representation vector H_src for all fields on the data source side, constructing an efficient vector index library. In the online phase, given a query vector H_biz, the system uses the vector approximate nearest neighbor retrieval engine to perform similarity searches of the data source fields, recalling Top-N candidate fields.

[0045] In some embodiments of this disclosure, when candidate fields meet preset triggering conditions, the business semantic relationship between multiple candidate fields is determined by combining the domain ontology knowledge graph and the reasoning of a large language model. One possible approach is to query a pre-built business domain ontology library to perform preliminary semantic relationship judgments on the multiple candidate fields based on complementarity, crossover, and authoritative coverage, obtaining the ontology library's preliminary judgment results; input the business meanings and statistical definitions of the multiple candidate fields into the large language model for verification and reasoning, obtaining the large language model's reasoning results; and determine the business semantic relationship between the multiple candidate fields based on the ontology library's preliminary judgment results and the large language model's reasoning results. Specifically, when multiple candidates with semantic scores exceeding a threshold θ appear among the recalled Top-N candidate fields, and these candidates belong to different business departments or different data tables, a dynamic semantic assembly engine is triggered. When a single candidate (only one field exceeds the threshold) is involved, a single-source matching result is directly output, skipping the assembly process. Based on the static ontology knowledge graph judgment, a large model is introduced for dynamic semantic relationship reasoning to handle novel relationships or ambiguous scenarios not covered by the ontology library.

[0046] Specifically, the system first queries the pre-built business domain ontology library and performs the following three types of semantic relationship determination on multiple candidate fields:

[0047] Furthermore, if the ontology has clearly labeled relationships, the system triggers the large model to perform verification reasoning, inputting the business meaning and statistical caliber of the two fields into the large model to verify whether the ontology's conclusions hold true; if the ontology does not have clearly labeled relationships, or if there are ambiguities (such as a relationship simultaneously satisfying complementary and overlapping features), then the large model is triggered to perform fallback reasoning:

[0048] In the embodiments described above, the business semantic relationships between multiple candidate fields are determined based on the initial judgment results of the ontology database and the inference results of the large language model. One possible approach is as follows: if the ontology database has explicit annotations and the initial judgment results of the ontology database are consistent with the inference results of the large language model, then the initial judgment results of the ontology database are adopted; if the ontology database has explicit annotations and the initial judgment results of the ontology database are consistent with the inference results of the large language model, then manual processing is triggered and the field is marked as a potential ontology conflict; if the ontology database has no explicit annotations and the confidence level of the large language model inference is greater than or equal to a preset threshold, then the inference results of the large language model are adopted and the candidate update pool of the ontology database is updated; if the ontology database has no explicit annotations and the confidence level of the large language model inference is less than a preset threshold, then manual processing is triggered and the field is marked as having an unknown relationship.

[0049] Figure 3 A decision logic diagram for a dynamic semantic assembly engine is provided as an exemplary embodiment of this disclosure. For example... Figure 3 As shown, specifically, the initial judgment results from the ontology database are fused with the inference results from the large model to make a decision:

[0050] In some embodiments of this disclosure, aggregation operators are generated based on business semantic relationships to complete data assembly. Specifically, taking the "Total Number of Employees in the Jurisdiction" field in a street's "Labor Force Employment Statistics Report" as an example, the processing procedure of the dynamic semantic assembly engine is as follows:

[0051] In some embodiments of this disclosure, manual corrections made by business personnel to the matching results are obtained. These manual corrections are categorized into entity-level feedback, rule-level feedback, or meta-knowledge feedback using a correction type identifier. For entity-level feedback, a weight update strategy based on a time decay function is used to adjust the entity confidence score. For rule-level feedback, a versioned management strategy is adopted, and when the number of accumulated feedback samples is greater than or equal to a first preset threshold, a low-rank incremental fine-tuning method is used to update the model parameters. For meta-knowledge feedback, manual review is triggered when the number of accumulated feedback samples is greater than or equal to a second preset threshold. It should be noted that this disclosure does not limit the first and second preset thresholds, and adjustments can be made according to actual circumstances.

[0052] Figure 4 A differentiated adaptive feedback optimization logic diagram is provided for an exemplary embodiment of this disclosure. For example... Figure 4 As shown, specifically, when business personnel manually correct the output matching results, the system records this correction operation, automatically determines the nature of the correction through the correction type identifier, and routes it to the corresponding feedback processing layer, thereby achieving accurate and reliable system adaptive evolution.

[0053] Based on the nature of the correction content, the system categorizes feedback into three types:

[0054] First layer: Entity-level feedback processing (time decay strategy).

[0055] For factual update feedback, the system records the corrected value, original value, timestamp, and information of the person making the correction. The historical weight is calculated using a time decay function. w(t) = w0× exp(-λ × t).

[0056] Where w(t) represents the historical feedback weights at a time interval t from the current time; w0 represents the initial weights (new feedback is set to 1.0); and λ represents the configurable decay coefficient. The time-decayed feedback weights are aggregated to update the entity's confidence score in the system, used to dynamically adjust the weights of each entity in the assembly result without triggering fine-tuning of the business semantic encoder model parameters.

[0057] Second layer: Rule-level feedback processing (version management strategy).

[0058] For feedback regarding changes in policy definitions, the system maintains a policy definition version chain, with each definition marked with its effective start and end dates. During matching, the system selects the corresponding policy definition version based on the period to which the report belongs. When a new rule takes effect, the old rule is automatically marked as invalid.

[0059] After a rule takes effect, it enters the "rule-level feedback pool". After accumulating enough feedback samples (recommended ≥20 different rules), the model is fine-tuned using a low-rank incremental fine-tuning method (injecting a low-rank matrix ΔW = A·B next to the original parameter matrix, where A∈R^(d×r), B∈R^(r×d), and r is much smaller than d).

[0060] The third layer is meta-knowledge feedback processing (cumulative consensus + manual review strategy).

[0061] For feedback of matching errors, the weight is constant at 1.0 and does not participate in decay. After accumulating sufficient feedback samples (recommended ≥10 feedbacks with the same meta-knowledge), a rapid manual review is triggered. After review and confirmation, the ontology is updated and the model's low-rank incremental fine-tuning is triggered.

[0062] Figure 5 This is a schematic diagram of the structure of a data source intelligent matching device 50 based on a hybrid semantic model, provided as an exemplary embodiment of this disclosure. Figure 5 As shown, the data source intelligent matching device 50 based on a hybrid semantic model includes: a construction module 51, an aggregation module 52, a mapping module 53, and a generation module 54.

[0063] Among them, the construction module 51 is used to construct structured prompt words based on the obtained structured metadata of the form; input the structured prompt words into the large language model for semantic reasoning to generate a business semantic enhancement description; input the business semantic enhancement description into the deep feature extraction network to obtain the business semantic vector; The aggregation module 52 is used to construct a heterogeneous attribute graph based on the structured metadata of the multi-source heterogeneous data resources obtained in the distributed environment; and to aggregate the neighborhood topology information of the nodes in the heterogeneous metadata graph through a graph convolutional coding network to obtain the structure-aware data source technology semantic vector. Mapping module 53 is used to map business semantic vectors and data source technology semantic vectors to a shared latent space respectively; and to recall candidate fields based on vector similarity in the shared latent space. The generation module 54 is used to determine the business semantic relationship between multiple candidate fields by combining the reasoning of the domain ontology knowledge graph and the large language model when the candidate fields meet the preset trigger conditions; and to generate aggregation operators based on the business semantic relationship to complete the data assembly.

[0064] Optionally, the generation module 54 can also be used for: The system obtains manual corrections made by business personnel to the matching results and classifies these manual corrections into entity-level feedback, rule-level feedback, or meta-knowledge feedback using a correction type identifier. For entity-level feedback, a weight update strategy based on a time decay function is used to adjust the entity confidence score; For rule-level feedback, a versioned management strategy is adopted, and when the number of accumulated feedback samples is greater than or equal to the first preset threshold, a low-rank incremental fine-tuning method is used to update the model parameters. For meta-knowledge feedback, manual review is triggered when the number of accumulated feedback samples is greater than or equal to the second preset threshold.

[0065] Optionally, the form's structured metadata includes: field identifiers, display labels, field hierarchy paths, fill-in instructions, statistical definitions, neighboring field sets, and historical fill-in example values.

[0066] Optionally, the enhanced business semantic description includes: the true meaning of the form field in the business context, statistical caliber, typical value range, and potential related concepts.

[0067] Optionally, the heterogeneous attribute graph includes: repository nodes, dataset nodes, data element nodes, and data service interface nodes.

[0068] Optionally, when the aggregation module 52 aggregates the neighborhood topology information of nodes in the heterogeneous graph of metadata through a graph convolutional coding network to obtain the structure-aware data source technology semantic vector, it is used for: Graph convolutional coding networks obtain node embedding matrices by aggregating neighborhood topology information through multi-layer graph convolution. The output of the last layer of graph convolution is concatenated with the field text embedding and then linearly transformed to generate a data source technology semantic vector with the same dimension as the business semantic vector.

[0069] Optionally, the mapping module 53, when mapping the business semantic vector and the data source technology semantic vector to the shared latent space respectively, and recalling candidate fields based on vector similarity in the shared latent space, is used for: A nonlinear dimensionality reduction network is used to map the business semantic vectors to obtain a low-dimensional representation of the business side. A single-layer linear projection is used to map the semantic vector of the data source technology to obtain a low-dimensional representation of the data source side. Calculate the cosine similarity based on the low-dimensional representations of the business side and the data source side. Candidate fields are recalled based on cosine similarity.

[0070] Optionally, when generating module 54 combines the reasoning of the domain ontology knowledge graph and the large language model to determine the business semantic relationship between multiple candidate fields, it is used for: The system queries a pre-built business domain ontology library to perform preliminary semantic relationship judgments on multiple candidate fields based on complementarity, cross-references, and authoritative coverage, and obtains the preliminary judgment results from the ontology library. The business meanings and statistical definitions of multiple candidate fields are input into the large language model for verification and reasoning to obtain the reasoning results of the large language model. Based on the initial judgment results from the ontology database and the inference results from the large language model, the business semantic relationships between multiple candidate fields are determined.

[0071] Optionally, when determining the business semantic relationships between multiple candidate fields based on the initial judgment results from the ontology library and the inference results from the large language model, the generation module 54 is used for: If the ontology is clearly labeled and the initial judgment result of the ontology is consistent with the reasoning result of the large language model, then the initial judgment result of the ontology shall be adopted. If the ontology is clearly labeled and the initial judgment result of the ontology is consistent with the reasoning result of the large language model, then manual processing is triggered and it is marked as a potential ontology conflict. If the ontology is not explicitly labeled and the reasoning confidence of the large language model is greater than or equal to the preset threshold, the reasoning result of the large language model is used and the candidate update pool of the ontology is updated. If the ontology is not explicitly labeled and the reasoning confidence of the large language model is less than a preset threshold, manual processing will be triggered and the relation will be marked as unknown.

[0072] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0073] Figure 6 This is a schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this disclosure. For example... Figure 6 As shown, the electronic device includes a memory 61 and a processor 62. Additionally, the electronic device also includes a power supply component 63 and a communication component 64.

[0074] Memory 61 is used to store computer programs and can be configured to store various other data to support operation on the electronic device. Examples of this data include instructions for any application or method used to operate on the electronic device.

[0075] The memory 61 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0076] Communication component 64 is used for data transmission with other devices.

[0077] The processor 62 executes computer instructions stored in the memory 61 to: construct structured prompt words based on the acquired form structured metadata; input the structured prompt words into a large language model for semantic reasoning to generate a business semantic enhancement description; input the business semantic enhancement description into a deep feature extraction network to obtain a business semantic vector; construct a heterogeneous attribute graph based on the acquired structured metadata of multi-source heterogeneous data resources in a distributed environment; aggregate the neighborhood topology information of nodes in the heterogeneous metadata graph through a graph convolutional coding network to obtain a structure-aware data source technology semantic vector; map the business semantic vector and the data source technology semantic vector to a shared latent space respectively; recall candidate fields based on vector similarity in the shared latent space; when the candidate fields meet preset triggering conditions, determine the business semantic relationship between multiple candidate fields by combining the reasoning of the domain ontology knowledge graph and the large language model; and generate an aggregation operator based on the business semantic relationship to complete data assembly.

[0078] Accordingly, embodiments of this disclosure also provide a computer-readable storage medium storing a computer program. When the computer-readable storage medium stores a computer program, and the computer program is executed by one or more processors, it causes one or more processors to perform... Figure 1 Each step in the method embodiment.

[0079] Accordingly, embodiments of this disclosure also provide a computer program product, which includes a computer program / instructions that are executed by a processor. Figure 1 Each step in the method embodiment.

[0080] The above Figure 6The communication component is configured to facilitate wired or wireless communication between the device containing the communication component and other devices. The device containing the communication component can access wireless networks based on communication standards, such as WiFi, 2G, 3G, 4G / LTE, 5G, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, the communication component also includes a Near Field Communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on Radio Frequency Identification (RFID), Infrared Data Association (IrDA) technology, Ultra-Wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0081] The above Figure 6 The power supply component provides power to the various components of the device in which it resides. The power supply component may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device in which it resides.

[0082] The aforementioned electronic devices also include a display screen and audio components.

[0083] The display includes a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can sense not only the boundaries of touch or swipe actions, but also the duration and pressure associated with the touch or swipe operation.

[0084] An audio component may be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals may be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0085] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0086] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0087] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0088] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0089] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0090] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0091] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0092] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0093] The above are merely specific embodiments of this disclosure, enabling those skilled in the art to understand or implement this disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to these embodiments, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data source intelligent matching method based on a hybrid semantic model, characterized in that, include: Based on the obtained structured metadata of the form, construct structured prompt words; The structured prompts are then input into a large language model for semantic reasoning to generate a business semantic enhancement description; the business semantic enhancement description is then input into a deep feature extraction network to obtain a business semantic vector. Based on the structured metadata of multi-source heterogeneous data resources obtained in a distributed environment, a heterogeneous attribute graph is constructed; the neighborhood topology information of the nodes in the heterogeneous metadata graph is aggregated through a graph convolutional coding network to obtain a structure-aware data source technology semantic vector. The business semantic vector and the data source technology semantic vector are respectively mapped to a shared latent space; Candidate fields are recalled based on vector similarity in the shared latent space; When the candidate fields meet the preset triggering conditions, the business semantic relationship between multiple candidate fields is determined by combining the reasoning of the domain ontology knowledge graph and the large language model; and an aggregation operator is generated based on the business semantic relationship to complete the data assembly.

2. The method according to claim 1, characterized in that, The method further includes: The manual correction operations of business personnel on the matching results are obtained, and the manual correction operations are divided into entity-level feedback, rule-level feedback or meta-knowledge feedback through a correction type identifier. For the entity-level feedback, a weight update strategy based on a time decay function is used to adjust the entity confidence score; For the rule-level feedback, a versioned management strategy is adopted, and when the number of accumulated feedback samples is greater than or equal to the first preset threshold, a low-rank incremental fine-tuning method is used to update the model parameters. For the aforementioned meta-knowledge feedback, if the number of accumulated feedback samples is greater than or equal to a second preset threshold, manual review is triggered.

3. The method according to claim 1, characterized in that, The structured metadata of the form includes: field identifiers, display labels, field hierarchy paths, filling instructions, statistical definition text, neighboring field sets, and historical filling example values.

4. The method according to claim 1, characterized in that, The enhanced business semantics description includes: the true meaning of the form fields in the business context, statistical caliber, typical value range, and potential related concepts.

5. The method according to claim 1, characterized in that, The heterogeneous attribute graph includes: repository nodes, dataset nodes, data element nodes, and data service interface nodes.

6. The method according to claim 1, characterized in that, The process of aggregating the neighborhood topology information of nodes in the heterogeneous metadata graph through a graph convolutional coding network to obtain a structure-aware data source technology semantic vector includes: The graph convolutional coding network obtains the node embedding matrix by aggregating neighborhood topology information through multi-layer graph convolution; The output of the last layer graph convolution is concatenated with the field text embedding and then linearly transformed to generate the data source technology semantic vector of the same dimension as the business semantic vector.

7. The method according to claim 1, characterized in that, The business semantic vector and the data source technology semantic vector are respectively mapped to a shared latent space; Candidate fields are recalled based on vector similarity in the shared latent space, including: The business semantic vector is mapped using a nonlinear dimensionality reduction network to obtain a low-dimensional representation on the business side; The semantic vector of the data source technology is mapped using a single-layer linear projection to obtain a low-dimensional representation of the data source side; Calculate the cosine similarity based on the business-side low-dimensional representation and the data source-side low-dimensional representation; Based on the cosine similarity, the candidate fields are recalled.

8. The method according to claim 1, characterized in that, The reasoning method that combines domain ontology knowledge graphs and large language models to determine the business semantic relationships between multiple candidate fields includes: The system queries a pre-built business domain ontology library to perform preliminary semantic relationship judgments on multiple candidate fields based on complementarity, cross-references, and authoritative coverage, and obtains the preliminary judgment results from the ontology library. The business meaning and statistical caliber of the multiple candidate fields are input into the large language model for verification and reasoning to obtain the reasoning result of the large language model. Based on the initial judgment results of the ontology and the reasoning results of the large language model, the business semantic relationships between multiple candidate fields are determined.

9. The method according to claim 8, characterized in that, The step of determining the business semantic relationships between multiple candidate fields based on the initial judgment results of the ontology database and the inference results of the large language model includes: If the ontology library has clear annotations and the initial judgment result of the ontology library is consistent with the reasoning result of the large language model, then the initial judgment result of the ontology library shall be adopted. If the ontology library has clear annotations and the initial judgment result of the ontology library is consistent with the reasoning result of the large language model, then manual processing is triggered and it is marked as a potential ontology conflict. If the ontology is not explicitly labeled and the reasoning confidence of the large language model is greater than or equal to a preset threshold, then the reasoning result of the large language model is adopted and the candidate update pool of the ontology is updated. If the ontology is not explicitly labeled and the reasoning confidence of the large language model is less than the preset threshold, manual processing is triggered and the relationship is marked as unknown.

10. A data source intelligent matching device based on a hybrid semantic model, characterized in that, include: The building module is used to construct structured prompts based on the obtained structured metadata of the form; The structured prompts are then input into a large language model for semantic reasoning to generate a business semantic enhancement description; the business semantic enhancement description is then input into a deep feature extraction network to obtain a business semantic vector. The aggregation module is used to construct a heterogeneous attribute graph based on the structured metadata of multi-source heterogeneous data resources acquired in a distributed environment; and to aggregate the neighborhood topology information of the nodes in the heterogeneous metadata graph through a graph convolutional coding network to obtain a structure-aware data source technology semantic vector. The mapping module is used to map the business semantic vector and the data source technology semantic vector to a shared latent space, respectively. Candidate fields are recalled based on vector similarity in the shared latent space; The generation module is used to determine the business semantic relationship between multiple candidate fields by combining the reasoning of the domain ontology knowledge graph and the large language model when the candidate fields meet the preset trigger conditions; and to generate an aggregation operator based on the business semantic relationship to complete the data assembly.

Images