An entity linking method, system, device and medium

Abstract

This invention discloses an entity linking method, system, device, and medium, belonging to the field of computer information processing technology. The method involves: performing multimodal feature integration and dynamic attribute perception on target entities in business text and preset entities in multiple business graphs, respectively, to obtain a first vector of the target entity and a second vector of each preset entity in a first semantic space; calculating the first semantic similarity between the first vector and each second vector, and filtering candidate entities by combining the ranking results of all first semantic similarities with dynamic business rules; inputting the first vector and each second vector into a target encoding network for transformation, respectively obtaining the target vector and candidate vectors in the second semantic space; calculating the second similarity between the target vector and each candidate vector, and obtaining the linking result by combining the ranking results of all second similarities with counterfactual verification and business rule constraints. Therefore, by implementing this invention, the accuracy of linking can be improved.

Description

Technical Field

[0001] This invention relates to the field of computer information processing technology, and in particular to a method, system, device and medium for linking entities. Background Technology

[0002] Entity linking is a key technology in the field of natural language processing and knowledge graph fusion. It aims to accurately associate entities mentioned in unstructured or semi-structured text with predefined standardized entities in structured knowledge bases or knowledge graphs. This technology plays an important supporting role in information integration, intelligent question answering, content recommendation, and business decisions in vertical fields (such as power operation).

[0003] Currently, the main approach is to use a contextual semantic matching scheme based on pre-trained language models. This method first identifies entity mentions in business texts and then uses pre-trained language models (such as the BERT series) to deeply encode the text context of the mentions to obtain semantically rich context vectors. At the same time, the text descriptions of candidate entities in the knowledge base are encoded into vectors. By calculating the semantic similarity between the context vector and the description vector, the candidate with the highest similarity is selected as the link result.

[0004] However, the optimization path of existing entity linking methods focuses on deepening text context modeling. Therefore, in complex business scenarios that require decision-making by combining multiple dynamically changing business graphs, the existing technology does not make full use of information and the expressive power of feature vectors is insufficient, resulting in low accuracy of entity linking. Summary of the Invention

[0005] This invention provides an entity linking method, system, device, and medium that can solve the problem of low accuracy in entity linking caused by insufficient expressive power of feature vectors.

[0006] This invention provides an entity linking method, comprising: Multimodal feature integration and dynamic attribute perception are performed on the target entity in the business text and the preset entities of multiple preset business graphs respectively to obtain the first representation vector corresponding to the target entity in the first vector semantic space, and the second representation vector corresponding to each preset entity. In the first vector semantic space, the first semantic similarity between the first representation vector and each of the second representation vectors is calculated. The first ranking result among all the first semantic similarities is combined with the preset dynamic business rules for filtering, and a preset number of candidate entities are determined in the preset entities. The first representation vector and the second representation vector corresponding to each candidate entity are input into a preset target encoding network for transformation to obtain the target vector and candidate vector in the second vector semantic space, respectively. The target encoding network is trained by contrastive learning, and the semantic discriminative power of the second vector semantic space is greater than that of the first vector semantic space. Calculate the second semantic similarity between the target vector and each of the candidate vectors. Then, combine the second ranking results among all the second semantic similarities with counterfactual verification and preset business rule constraints to obtain the entity linking results.

[0007] This invention integrates multimodal features and dynamic attribute perception to achieve multimodal vector representation of target entities and preset entities, thereby increasing semantic information and providing a high-discrimination foundation for subsequent similarity calculations. First semantic similarity is calculated in a first vector semantic space, enabling rapid quantification of semantic distance between entities during the initial screening stage. Further, combining ranking results with dynamic business rules, the candidate size is accurately compressed, reducing subsequent computational load. Based on the target encoding network trained through contrastive learning, vectors are mapped to a second vector semantic space with higher discriminative power, increasing inter-class distance and decreasing intra-class distance, thus improving similarity resolution. Combining counterfactual verification and business rule constraints, causal consistency verification is performed on the ranking results, quantifying and eliminating link conflicts, ultimately increasing the accuracy and stability of the final entity linking results.

[0008] Furthermore, the step of performing multimodal feature integration and dynamic attribute perception on the target entities in the business text and the preset entities in multiple preset business graphs specifically involves: Obtain multimodal data corresponding to each of the preset business maps, wherein the multimodal data includes data corresponding to text modality, structured modality and semi-structured modality; Modal features corresponding to each modality are extracted using a preset feature extraction model corresponding to each modality. Obtain a preset cross-graph alignment feature set, wherein the cross-graph alignment feature set includes a first feature of the aligned target entity and a second feature of each preset entity; The first feature and each modal feature are fused to obtain the first static attribute vector corresponding to the target entity. For each preset entity, the second feature and each modal feature are fused to obtain the second static attribute vector corresponding to each preset entity. The first static attribute vector and each of the second static attribute vectors are respectively input into a preset temporal feature extraction model to extract temporal information associated with the business text, thereby obtaining the first dynamic attribute vector and the second dynamic attribute vector. The first static attribute vector is concatenated with the corresponding first dynamic attribute vector to obtain the first representation vector, and the second static attribute vector is concatenated with the corresponding second dynamic attribute vector to obtain the second representation vector.

[0009] By integrating textual, structured, and semi-structured trimodal data, the entity representation dimensions are comprehensively covered, and the information abundance of static attribute vectors is improved. The introduction of cross-graph alignment feature sets injects prior alignment information, making the initial distribution of the vector space closer to the true semantic relationships. Dynamic attribute vectors obtained from the temporal feature extraction model can capture entity state drift in real time, allowing representations to be refreshed with business updates, enhancing timeliness. Concatenating static and dynamic vectors allows the first / second representation vectors to simultaneously carry long-term attributes and short-term changes, which helps improve the accuracy of subsequent similarity calculations.

[0010] Furthermore, the entity linking method further includes: Obtain business text, and perform entity recognition on the business text to obtain multiple target entities; The target entity and the entity association identifiers of multiple business graphs are fused to obtain a fusion sequence; The fusion sequence is input into a bidirectional Transformer network to extract the contextual semantics and graph association semantics of the fusion sequence through a self-attention mechanism. The fusion vector is determined by combining the fusion sequence, the contextual semantics, and the graph association semantics. At the same time, the fusion vector is subjected to span analysis, relation classification, and attribute extraction to obtain multiple sets of triples. Based on each of the triples, the target entity and the preset entity are used as nodes, and the relationships between entities are used as edges, to obtain a graph structure across the graph spectrum; The graph structure is updated using a graph attention network to obtain the cross-graph alignment feature set.

[0011] This approach, based on the bidirectional Transformer, generates a fusion sequence that simultaneously encodes contextual semantics and graph association semantics in a single forward pass, improving feature extraction efficiency. Parallel span analysis, relation classification, and attribute extraction yield entity-relation-attribute triples, enhancing extraction consistency. By updating the graph attention network, node features are weighted and aggregated through attention, further improving the discriminative power of feature vectors after cross-graph alignment.

[0012] Furthermore, updating the graph structure based on the graph attention network specifically involves: In the graph attention network, the semantic association weights corresponding to each business graph are dynamically determined based on the business semantic priority of the business text, the attention coefficients between nodes are calculated based on the semantic association weights, and the graph structure is updated according to the attention coefficients.

[0013] This approach dynamically determines the attention coefficient based on business semantic priority, giving higher weight to important graphs, ensuring that node update directions align with business focus, and improving feature alignment accuracy. The attention coefficient is adjusted in real time, allowing graph weights to adaptively refresh as the business context changes, ensuring that the alignment results always closely match the current marketing scenario.

[0014] Furthermore, the second ranking results among all the second semantic similarities, combined with counterfactual verification and preset business rule constraints, yield the entity linking results, specifically as follows: The candidate entities are sorted according to the second semantic similarity to obtain the second sorting result; Based on the second sorting result, at least one initial link relationship between the target entity and the candidate entity is determined; For each set of initial link relationships, the target attribute value of the target vector or the candidate vector is replaced or perturbed to obtain counterfactual samples; The counterfactual sample is input into a preset link model to obtain the corresponding initial counterfactual link relationship. The initial counterfactual link relationship is then checked for consistency and corrected according to business rule constraints to obtain the target counterfactual link relationship. The initial link relationship and the target counterfactual link relationship are compared to calculate the difference degree, and the existence of a link conflict in the initial link relationship is determined based on the difference degree. The entity linking result is obtained by combining all initial link relationships that do not have the aforementioned link conflict.

[0015] By replacing / perturbing the target attributes, counterfactual samples are generated, quickly obtaining a large number of control groups for hypotheses and results, which is beneficial for evaluating the robustness of links; consistency checks are performed through business rule constraints to improve the consistency of link results; the degree of difference is calculated to quantify the degree of conflict, and only links with low degree of difference are retained, ultimately increasing the accuracy of entity link results.

[0016] Furthermore, the entity linking method is specifically as follows: Acquire training data, wherein the training data includes target entity sample data and corresponding candidate entity sample data, the candidate entity sample data includes positive sample data and negative sample data, the positive sample data is entity data that has a link relationship with the target entity sample data, and the negative sample data is entity data that does not have a link relationship with the target entity sample data; The representation vectors corresponding to the target entity sample data and the candidate entity sample data are respectively input into a preset initial encoding network for transformation. By minimizing the distance between the target entity sample data and the positive sample data in the second vector semantic space and maximizing the distance between the target entity sample data and the negative sample data in the second vector semantic space, the initial encoding network is optimized to obtain the target encoding network.

[0017] By employing a contrastive learning training strategy, the distance between positive samples is minimized and the distance between negative samples is maximized, thereby improving the semantic discriminative power of the target encoding network. Based on the joint loss function, the network parameters are updated in the direction of high discriminative power, the semantic space boundary of the second vector is clear, and the similarity calculation error is reduced.

[0018] Furthermore, the entity linking method further includes: Based on the entity linking results, business rules, solutions, risk schemes, and problem priorities associated with the target entity are extracted from the multiple preset business graphs. Combined with the work order requirements of the business text, a structured work order solution is obtained.

[0019] This automatically extracts key information from four types of data: business rules, solutions, risk management plans, and issue priorities, bringing them together in one go and improving information completeness. It also generates structured work order solutions based on work order requirements, resulting in highly executable solutions. Through structured output, the solutions are stored in association with the original work orders, enhancing continuous optimization capabilities.

[0020] Another embodiment of the present invention provides an entity linking system, including: a first vector module, a first filtering module, a second vector module, and a second filtering module; The first vector module is used to perform multimodal feature integration and dynamic attribute perception on the target entity in the business text and the preset entities of multiple preset business graphs respectively, to obtain the first representation vector corresponding to the target entity in the first vector semantic space, and the second representation vector corresponding to each preset entity. The first filtering module is used to calculate the first semantic similarity between the first representation vector and each of the second representation vectors in the first vector semantic space, and to filter the first ranking results among all the first semantic similarities in combination with preset dynamic business rules, and to determine a preset number of candidate entities in the preset entities. The second vector module is used to input the first representation vector and the second representation vector corresponding to each candidate entity into a preset target encoding network for transformation, so as to obtain the target vector and candidate vector in the second vector semantic space respectively. The target encoding network is trained by contrastive learning, and the semantic discriminative power of the second vector semantic space is greater than that of the first vector semantic space. The second filtering module is used to calculate the second semantic similarity between the target vector and each of the candidate vectors, and to obtain the entity linking result by combining the second ranking result among all the second semantic similarities with counterfactual verification and preset business rule constraints.

[0021] Another embodiment of the present invention provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the steps of the entity linking method of the present invention.

[0022] Another embodiment of the present invention provides a computer-readable storage medium item, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to perform steps as described in the entity linking method of the present invention. Attached Figure Description

[0023] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating an entity linking method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of an alignment process for a first vector semantic space provided by an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an entity linking system provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0027] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0028] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0029] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0030] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0031] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0032] See Figure 1To address the problem of low accuracy in entity linking due to insufficient expressive power of feature vectors in existing technologies, an embodiment of the present invention provides an entity linking method, comprising: Step 101: Perform multimodal feature integration and dynamic attribute perception on the target entity in the business text and the preset entities of multiple preset business graphs respectively, to obtain the first representation vector corresponding to the target entity in the first vector semantic space, and the second representation vector corresponding to each preset entity.

[0033] In the above steps, by integrating information from multi-source heterogeneous data and perceiving the temporal changes of entity attributes, a unified vector representation is constructed for each entity. Specifically, various types of entity-related data, such as text descriptions, structured records, and semi-structured forms, are accessed and processed. Using general feature extraction techniques adapted to various data modalities, these heterogeneous data are converted into a unified vector form and fused to form basic features representing the static characteristics of the entity. Time-related dimensions in entity attributes are identified and extracted. Through temporal analysis techniques, patterns or trends in the changes of key attribute values ​​over time are captured, generating dynamic attribute vectors reflecting the current state or dynamic evolution of the entity. The static attribute vectors obtained in the above steps are combined with the dynamic attribute vectors (e.g., vector concatenation or weighted fusion) to obtain a comprehensive representation vector for each entity in a first vector semantic space (i.e., the first representation vector and the second representation vector). This space is used to initially accommodate and compare the semantic information of all entities.

[0034] Step 102: In the first vector semantic space, calculate the first semantic similarity between the first representation vector and each of the second representation vectors, and filter the first ranking results among all the first semantic similarities in combination with preset dynamic business rules to determine a preset number of candidate entities in the preset entities.

[0035] In the above steps, by calculating semantic similarity and incorporating business logic, a small number of relevant candidates are efficiently and accurately selected from a large number of pre-defined entities. Specifically, in the first vector semantic space, a common vector similarity measurement method (such as cosine similarity, inner product, etc.) is used to calculate the similarity score between the target entity representation vector and each pre-defined entity representation vector. Subsequently, all pre-defined entities are ranked based on this score. Configurable and updatable business rules are introduced to intervene and filter the ranking results. These rules can be updated or their weights adjusted in real time according to business needs to achieve flexible business adaptation. For example, the business graph weight dynamic adjustment rule: in the electricity marketing scenario, if the business text involves "overdue payment risk", the screening priority of entities in the risk graph and problem graph is dynamically increased; if it involves "billing rule consultation", relevant entities in the rule graph are matched first. Combining similarity ranking and dynamic business rules, the TOP-K candidate entities are obtained.

[0036] Step 103: Input the first representation vector and the second representation vector corresponding to each candidate entity into a preset target encoding network for transformation to obtain the target vector and candidate vector in the second vector semantic space, wherein the target encoding network is trained by contrastive learning and the semantic discriminative power of the second vector semantic space is greater than that of the first vector semantic space.

[0037] In the above steps, a deep transformation is performed on the initially screened entity representations using an encoding network trained through contrastive learning. This aims to improve the model's ability to discriminate subtle semantic differences, especially attribute distribution differences arising across different knowledge graphs. Specifically, a neural network model is used as the encoder, trained using a contrastive learning paradigm. The training objective is to enable the model to learn to reduce the distance between entity pairs (positive sample pairs) that should be linked in business logic, while increasing the distance between unrelated entity pairs (negative sample pairs) in the transformed space. After this training, the encoding network can map the representation vectors in the input first semantic space to a second vector semantic space. In this space, the semantic relationships and differences between entities are more clearly and accurately characterized, and its discriminative power is higher than that of the initial first vector semantic space. The vectors of the target entity and candidate entities in this space are called the target vector and candidate vector, respectively.

[0038] Step 104: Calculate the second semantic similarity between the target vector and each of the candidate vectors. For the second ranking results among all the second semantic similarities, combine counterfactual verification and preset business rule constraints to obtain the entity linking results.

[0039] In the above steps, based on more accurate semantic similarity calculation, an automated verification and constraint mechanism is introduced to obtain a final reliable and conflict-free entity linking result. Specifically, in the second vector semantic space, the semantic similarity between the target vector and each candidate vector (second semantic similarity) is recalculated, and a more accurate ranking is performed accordingly. To avoid excessive dependence of the linking result on certain entity attributes or the existence of hidden conflicts, counterfactual reasoning is adopted. The core idea is to simulate changing a key attribute of the entity (target entity or candidate entity) for a linking relationship that is initially considered correct, observe whether the linking conclusion undergoes a disruptive change, and evaluate the robustness of the link by quantifying this change, identifying and eliminating potential and unstable links. Business rule constraints transform domain expert knowledge or business specifications into machine-executable logical rules. These rules are used to perform consistency checks on the links after counterfactual verification, automatically correcting links that violate basic business logic or hard constraints. This ensures that the final output is valid both semantically and in terms of rules. For example, entity type matching constraints: target entities of the "user problem" class can only link to entities of the "user service rule" class in the rule graph, and cannot link to "equipment operation and maintenance rules"; attribute value range constraints: if the "applicable voltage level" attribute value of a candidate entity is "10kV", while the target entity's scenario is "low-voltage residential electricity", then the link is determined not to meet the business range constraint and should be excluded or downgraded. The entity link result set is obtained by combining the links that have undergone fine-grained sorting, robustness verification, and rule consistency checks.

[0040] As an example of an embodiment of the present invention, the multimodal feature integration and dynamic attribute perception of the target entity in the business text and the preset entities of multiple preset business graphs specifically involves: acquiring multimodal data corresponding to each preset business graph, wherein the multimodal data includes data corresponding to text modality, structured modality, and semi-structured modality; extracting modal features corresponding to each modality through a preset feature extraction model corresponding to each modality; acquiring a preset cross-graph alignment feature set, wherein the cross-graph alignment feature set includes a first feature of the aligned target entity and a second feature of each preset entity; and fusing the first feature and each modal feature. A first static attribute vector corresponding to the target entity is obtained. For each preset entity, the second feature and each modal feature are fused to obtain a second static attribute vector corresponding to each preset entity. The first static attribute vector and each second static attribute vector are respectively input into a preset temporal feature extraction model to extract temporal information associated with the business text, thereby obtaining a first dynamic attribute vector and a second dynamic attribute vector. The first static attribute vector is concatenated with the corresponding first dynamic attribute vector to obtain a first representation vector, and the second static attribute vector is concatenated with the corresponding second dynamic attribute vector to obtain a second representation vector.

[0041] In this embodiment, as Figure 2As shown, raw data corresponding to entities related to the current business text are collected from multiple preset business graphs (such as knowledge graphs, rule graphs, risk graphs, and problem graphs) and their associated business systems. In the raw multimodal data, text modalities can be entity names, description fields, or associated document content, such as rule clauses or problem reports; structured modalities can be entity attribute value pairs stored in database tables, such as a user's "credit rating = AAA" or a device's "rated power = 1000kW"; semi-structured modalities can be work order records or log entries in JSON or XML format, or report fragments in HTML format, containing non-fully organized key-value pairs and text. For textual modalities, pre-trained language models (e.g., a variant of RoBERTa) can be used to encode the text. For example, for a piece of business text, the vector of the [CLS] tag output by the model or the sequence can be pooled to obtain a fixed-dimensional textual semantic feature vector. For structured modalities, translation-based embedding models (e.g., TransE or its variants) can be used. Each "entity-relationship-attribute value" triple is input into the model to learn the structured embedding vector of the entity, which contains its position and relation information in the graph structure. For semi-structured modalities, a lightweight table convolutional network (TCN) or an attention-based sequence model can be designed. First, the semi-structured data is parsed into a table-like row and column structure or a key-value sequence. Then, the network is used to extract its implicit pattern and semantic features and output a feature vector.

[0042] The preset cross-graph alignment feature set is obtained by taking the received business text and the association identifiers of multiple graphs as input, simultaneously extracting entities, relations, and attributes through a bidirectional Transformer architecture, and using a Graph Attention Network (GAT) to map entity features from different graphs to a unified semantic space for alignment. For target entities identified in the business text, the alignment feature set provides their vector representation in the unified space (i.e., the first feature); for preset entities in each preset business graph that are related to the current task, it also provides their vector representation in the aligned space (i.e., the second feature).

[0043] The first feature (cross-graph alignment feature) of the target entity is fused with its corresponding textual, structured, and semi-structured modal features. The fusion method can employ feature concatenation, followed by a fully connected layer for dimensionality reduction and integration, ultimately generating the first static attribute vector of the target entity. This vector integrates the semantics of the target entity after cross-graph alignment, the original features from multiple modalities, and preliminary dimensionality reduction and fusion information. For each preset entity, its second feature (cross-graph alignment feature) is fused with its corresponding multimodal features (also from the textual, structured, and semi-structured data of its respective graph) using the same operation, generating the second static attribute vector for each preset entity.

[0044] For each entity (including the target entity and preset entities), time-related attribute sequence data is collected. For example, for the "overdue payment" entity, the sequence of changes in its "overdue amount" over time is collected; for the "user" entity, the sequence of its "monthly electricity consumption" is collected. A temporal convolutional network (TCN) is used as the temporal feature extraction model. The temporal attribute sequence of each entity is input into the TCN. The TCN captures the long-term dependencies in the sequence through dilated causal convolution and finally outputs a dynamic attribute vector representing the temporal change pattern of the entity (the first dynamic attribute vector for the target entity and the second dynamic attribute vector for the preset entities).

[0045] The static attribute vector of each entity (containing multimodal fusion and cross-graph alignment information) is concatenated with its dynamic attribute vector (containing temporal variation information) to form the final representation vector of that entity. This yields the first representation vector of the target entity and the second representation vectors of each predefined entity. These vectors simultaneously contain rich information on static and dynamic aspects, as well as multimodal and cross-graph alignment, providing a comprehensive and high-quality feature foundation for subsequent similarity calculations and link decisions.

[0046] As an example of an embodiment of the present invention, the entity linking method further includes: acquiring business text; performing entity recognition on the business text to obtain multiple target entities; fusing the target entities and the entity association identifiers of multiple business graphs to obtain a fusion sequence; inputting the fusion sequence into a bidirectional Transformer network to extract the contextual semantics and graph association semantics of the fusion sequence through a self-attention mechanism; combining the fusion sequence, the contextual semantics, and the graph association semantics to determine a fusion vector; and simultaneously performing span analysis, relationship classification, and attribute extraction on the fusion vector to obtain multiple sets of triples; based on each triple, using the target entity and the preset entity as nodes, and the relationships between entities as edges, to obtain a cross-graph structure; and updating the graph structure according to a graph attention network to obtain the cross-graph alignment feature set.

[0047] In this embodiment, as Figure 2 As shown, firstly, the input marketing business text is preprocessed and segmented. A pre-trained Named Entity Recognition (NER) model or a rule / dictionary-based method is used to identify key entity mentions in the text. Entity indexes are pre-built for each preset business graph (knowledge graph K, rule graph R, risk graph V, and problem graph P). Based on the target entity names or keywords identified in the previous step, a fast parallel search is performed across all graph indices to find potentially related preset entities. Each target entity mention is then labeled with its possible associated graph type identifier. The fused sequence is then input into an encoder based on a bidirectional Transformer (such as the BERT architecture). This encoder, through its multi-layered self-attention mechanism, captures the dependencies between any two tokens in the entire sequence in parallel. During this process, the contextual semantic information of the text and the prior graph association information introduced through graph type embedding are dynamically interacted and fused. Finally, the Transformer encoder outputs a fusion vector for each token in the sequence, which simultaneously encodes the token's lexical semantics, its positional context in the sentence, and its potential association information with a specific business graph.

[0048] Above the Transformer encoder, three parallel task-specific layers (task heads) are connected. They share the underlying fusion vector representation but perform different subtasks. Specifically, for the span prediction head (entity recognition), a fully connected layer with Softmax is used to predict the probability of each token being the start (B) or interior (I) position of an entity. By decoding consecutive BI tags, the precise boundary (span) of the entity in the sequence is determined, and the entity span coordinates are obtained. For the relationship classification head, based on a convolutional neural network (CNN), the fusion vector of the corresponding token is extracted from the identified entity pair, pooled to obtain the joint representation of the entity pair, and this representation is input into a relationship classifier (such as a multilayer perceptron MLP) to predict the predefined relationship type between the entity pair (such as "user-overdue" corresponding to "problem exists", "overdue-billing rule" corresponding to "association rule"), and the relationship label is obtained. For the attribute extraction head, based on the gated recurrent unit (GRU), the fusion vector of each identified entity pair is time-series modeled to extract the core attributes and attribute values ​​of the entity (such as the attribute "address = a certain unit in a certain community" for "user" and the attribute "status = paid but not cancelled" for "overdue payment"), resulting in entity-attribute-value triples.

[0049] Based on all the triples extracted in the previous step, each entity in the triple (regardless of whether it originates from business text or a pre-defined graph) is treated as a node in the graph, and the relations defined in the triples are used as edges connecting the corresponding entity nodes. Each node (entity) is assigned an initial feature vector. This feature can be directly obtained from the pooling result of the fusion vector of the corresponding span in the Transformer, or it can be embedded in conjunction with its entity type. The heterogeneous graph (cross-graph hybrid graph) constructed above, containing target entity nodes and pre-defined entity nodes, is input into a multi-layer graph attention network. In each layer, for each node in the graph, GAT calculates its association strength (attention coefficient) with all neighboring nodes through an attention mechanism. The calculation of the attention coefficient considers the node's own features, the features of neighboring nodes, and the type of connecting edges (relationships). Each node, based on its attention coefficient with its neighbors, performs weighted aggregation of the features of its neighboring nodes, thereby updating its own feature representation. Through this multi-layer "aggregation-update" operation, feature information from nodes in different graphs is propagated and exchanged on the graph. For example, the attention coefficient... ,in 、 Let W be the adjacent node vector, W be the weight matrix, and a be the attention vector. Furthermore, the initial weights or biases of nodes of different graph types in the attention calculation can be dynamically adjusted according to the business semantics (e.g., the core of the current work order is "billing issues"), guiding the model to pay more attention to the graphs related to the current problem (such as rule graphs and problem graphs).

[0050] After multiple rounds of iteration and updates by GAT, the feature vector of each node in the graph incorporates the semantic information of its multi-hop neighbors across the graph. At this point, the originally heterogeneous entity features in different graphs are mapped to a unified, semantically aligned vector space, outputting the final feature vectors of all nodes, which together constitute the cross-graph aligned feature set.

[0051] For example, a customer complaint work order: "After a user in Unit 3, Building 10, XX Community paid 500 yuan in May 2024, the system still shows an outstanding balance, suspected of violating Article 3 of the Electric Power Marketing Billing Rules". After word segmentation, it is ["XX Community" (K), "Building 10" (K), "Unit 3" (K), "user" (K), "May 2024" (O), "paid" (R), "500 yuan" (R), "after" (O), ",", "system" (K), "still" (O), "shows" (O), "outstanding balance" (V / P), ",", "suspected" (O), "violating" (R), "<", "Article 3 of the Electric Power Marketing Billing Rules" (R), ">"]. The Transformer encoder outputs a 768-dimensional fused encoding vector for each token. The vector of "outstanding balance" contains features of both text semantics (outstanding balance status) and graph type (V / P). The span prediction head outputs the recognized entities "user in Unit 3, Building 10, XX Community" (positions 0-3), "outstanding balance" (position 12), "Article 3 of the Electric Power Marketing Billing Rules" (position 17); the relation classification head outputs the relations "user in Unit 3, Building 10, XX Community - has problem - outstanding balance", "outstanding balance - associated rule - Article 3 of the Electric Power Marketing Billing Rules"; the attribute extraction head outputs the extracted triples "<user in Unit 3, Building 10, XX Community, payment amount, 500 yuan>", "<outstanding balance, occurrence time, May 2024>", and "<Article 3 of the Electric Power Marketing Billing Rules, rule type, billing execution>". Map the above entities, relations, and attributes to a unified vector space to establish a cross-graph association of "knowledge graph (user, system) - rule graph (billing rules) - risk graph (outstanding balance risk) - problem graph (outstanding balance problem)".

[0052] As an example of an embodiment of the present invention, the updating of the graph structure according to the graph attention network is specifically as follows: In the graph attention network, the semantic association weights corresponding to each business graph are dynamically determined based on the business semantic priorities of the business text, the attention coefficients between nodes are calculated based on the semantic association weights, and the graph structure is updated according to the attention coefficients.

[0053] In this embodiment, a multi-head self-attention mechanism optimization is carried out in the graph attention network, a cross-graph attention weight allocation strategy is designed, different weights are assigned to the entity features of different graphs based on the marketing business semantic priorities, and the attention scores are calculated: (where Q is the query vector, K is the key vector, V is the value vector, is the dimension of the key vector) to dynamically adjust the semantic association strength, reduce the dependence on a single graph, and improve the cross-graph feature synergy.

[0054] As an example of an embodiment of the present invention, the step of obtaining entity linking results by combining the second ranking results among all the second semantic similarities with counterfactual verification and preset business rule constraints specifically involves: ranking the candidate entities according to the second semantic similarity to obtain the second ranking results; determining at least one initial linking relationship between the target entity and the candidate entity based on the second ranking results; replacing or perturbing the target attribute value of the target vector or the candidate vector for each set of initial linking relationships to obtain counterfactual samples; inputting the counterfactual samples into a preset linking model to obtain the corresponding initial counterfactual linking relationship; performing a consistency check and correction on the initial counterfactual linking relationship according to business rule constraints to obtain the target counterfactual linking relationship; comparing the initial linking relationship and the target counterfactual linking relationship to calculate the difference degree; determining whether the initial linking relationship has a link conflict based on the difference degree; and combining all initial linking relationships without link conflicts to obtain the entity linking results.

[0055] In this embodiment, for an "initial link relationship" (e.g., T (target entity) <--> C (candidate entity)), it is determined whether changing a key attribute of T or C would lead to a reversal of the link relationship. Specifically, attributes in T and C that may be crucial to the link decision are identified. For example, for an "overdue payment" entity, the key attribute might be "overdue amount"; for a "billing rule" entity, the key attribute might be "rule application area". The selected key attribute value is modified, such as replacing "rule application area" from "area A" to "area B", or increasing or decreasing "overdue amount" by a significant percentage (e.g., ±30%). Based on the modified attribute value, the "complete representation vector" (i.e., the first representation vector or the second representation vector) corresponding to the entity is locally adjusted. This can be achieved through a small attribute value embedding network, which learns the impact of attribute value changes on the overall representation vector, generating a modified entity vector, i.e., a counterfactual sample (T' or C'). Input counterfactual samples (e.g., T' and the original C, or T and C') into a trained linking model (which can be the entire network or a key part responsible for calculating and ranking the second semantic similarity). The model outputs a new round of similarity scores and rankings based on the counterfactual input, resulting in a new, hypothetical linking outcome, i.e., the initial counterfactual linking relationship. This may indicate that T' no longer highly matches C, but rather matches another candidate entity C2 more closely. Business rule constraints are a predefined set of logical rules, expressed in language such as first-order logic, for example: "IF entity type (problem) = 'billing dispute' AND associated entity type (rule) = 'tiered electricity pricing rule' THEN must NOT link status = 'resolved'", or more simply: "Customer problems with a risk level of 'high' cannot be directly linked to processing rules that only apply to ordinary customers". Design a rule parsing and execution engine to transform the initial counterfactual linking relationship (which may contain multiple entities and relationships) obtained in the previous step into a series of logical assertions. The engine matches logical assertions against a predefined business rule constraint rule library. If a rule violation is found (for example, a counterfactual link links a high-risk issue to a basic rule), the link relationship is automatically corrected according to the correction strategy defined in the rule (such as "disable this link" or "replace with the default security rule") to obtain the target counterfactual link relationship that meets all hard constraints.

[0056] The comparison between the initial link relationship and the target counterfactual link relationship primarily calculates two types of dissimilarity. For example, the difference between the ranking of C in the initial relationship and the ranking of C (or C2 in the new link) in the counterfactual relationship; or the absolute value of the difference between the matching confidence score of T and C in the initial relationship and the corresponding matching confidence score in the counterfactual relationship. A dissimilarity threshold is set. If the calculated dissimilarity exceeds this threshold, the current initial link relationship is considered overly sensitive to changes in this key attribute, indicating potential instability or implicit conflict, and is thus judged as having a link conflict. All initial link relationships judged to have link conflicts are removed or marked as low-confidence pending review. All initial link relationships without link conflicts are considered reliable, robustly validated, and business rule compliance-checked final results, combined and output to obtain the final entity link result.

[0057] As an example of an embodiment of the present invention, the entity linking method specifically includes: acquiring training data, wherein the training data includes target entity sample data and corresponding candidate entity sample data, the candidate entity sample data includes positive sample data and negative sample data, the positive sample data is entity data that has a link relationship with the target entity sample data, and the negative sample data is entity data that does not have a link relationship with the target entity sample data; inputting the representation vector corresponding to the target entity sample data and the representation vector corresponding to the candidate entity sample data into a preset initial encoding network for transformation, so as to optimize the initial encoding network by minimizing the distance between the target entity sample data and the positive sample data in the second vector semantic space and maximizing the distance between the target entity sample data and the negative sample data in the second vector semantic space, thereby obtaining the target encoding network.

[0058] In this embodiment, the training data comes from historical marketing business processing records, such as closed-loop customer work orders, reviewed rule association records, and confirmed risk event reports. These records explicitly contain business text (including target entities) and standard entities (preset entities) in the various business graphs that are ultimately confirmed to be linked. The target entity sample data is the first representation vector extracted from the historical business text after undergoing the aforementioned multimodal feature integration and dynamic attribute perception, which serves as the anchor of the training sample. The positive sample data is the second representation vector corresponding to the preset entity that was actually linked in the historical record. This vector has also undergone a complete feature processing flow. A training data set typically contains one or more positive samples, representing the correct link target. In the same training batch, the positive samples corresponding to other target entity samples are used as negative samples of the current target entity to obtain negative sample data.

[0059] A two-branch Siamese network architecture or a symmetric encoder architecture is used as the initial encoding network, with both branches sharing the same network structure and parameters. One branch takes the representation vector of the target entity sample as input, and the other branch takes the representation vector of the candidate entities (positive or negative samples). The target entity vector and a batch of candidate entity vectors are respectively input into the two branches of the shared encoding network. The encoding network performs a nonlinear transformation on the input vectors, mapping them to the second vector semantic space to be learned, and outputs the corresponding target vector and candidate vectors. The contrastive loss is used as the optimization objective, and the loss function is expressed as: ; in, For the target entity vector, These are positive sample candidate vectors. Here, N represents the temperature coefficient, and N represents the number of target entity vectors in a training batch.

[0060] By minimizing the contrastive loss using the backpropagation algorithm and gradient descent optimizer (such as Adam), the initial encoding network is iteratively trained. Training stops when the model's loss on the validation set no longer decreases significantly or when a predetermined number of iterations are reached, thus obtaining the target encoding network.

[0061] It should be noted that, through the aforementioned contrastive learning training, the distance between positive sample pairs in the second vector semantic space is reduced, while the distance between negative sample pairs is increased, thereby effectively improving the inter-class separation and intra-class aggregation of the vector space, that is, enhancing the semantic discriminativeness of the second vector semantic space.

[0062] As an example of an embodiment of the present invention, the entity linking method further includes: extracting business rules, solutions, risk schemes and problem priorities associated with the target entity from the plurality of preset business graphs based on the entity linking results, and obtaining a structured work order solution by combining the work order request of the business text.

[0063] This embodiment provides a work order solution generation strategy suitable for automated processing scenarios in power marketing business (such as closed-loop processing of customer complaint work orders, consultation work orders, and fault repair work orders). The input is the accurate entity linking result output by the semantic-aware link resolution module. The core is to integrate cross-graphite association information to generate a standardized and executable work order processing solution, realizing the transformation of technical linking results into business implementation results. The specific implementation steps of this strategy are as follows: ① First, extract and integrate related information, analyze the entity linking results, and extract key business information from four types of graphs: 1) Extract the execution requirements of related rules from the rule graph; 2) Extract related solution knowledge from the knowledge graph; 3) Extract potential risk prevention and control points from the risk graph; 4) Extract problem classification and priority information from the problem graph, forming a comprehensive information pool of rule requirements + knowledge solutions + risk prevention and control + priority.

[0064] ② Next, construct the solution logic. Based on the comprehensive information pool and the core requirements of the work order, construct the solution framework according to the logic of priority sorting, process decomposition, responsibility allocation and clear time limit: 1) Determine the handling order according to the priority of the problem; 2) Decompose the solution into an operable step-by-step process; 3) Based on the responsibility division requirements of the rule graph, clarify the executing department / position of each step; 4) Combine the time limit requirements of the rules and set the completion time node of each step.

[0065] ③ Finally, the structured solution output transforms the constructed solution framework into a standardized format, containing eight core fields: "work order number, problem description, handling priority, step-by-step process, responsible department, time limit requirement, risk control measures, and acceptance criteria," forming a structured work order solution that can be directly used for business execution. At the same time, the solution is associated with the original work order for storage, supporting subsequent business traceability and model optimization.

[0066] like Figure 3 As shown, based on the above method embodiments, an embodiment of the present invention provides an entity linking system 300, including: a first vector module 301, a first filtering module 302, a second vector module 303, and a second filtering module 304; The first vector module 301 is used to perform multimodal feature integration and dynamic attribute perception on the target entity in the business text and the preset entities of multiple preset business graphs respectively, to obtain the first representation vector corresponding to the target entity in the first vector semantic space, and the second representation vector corresponding to each preset entity. The first filtering module 302 is used to calculate the first semantic similarity between the first representation vector and each of the second representation vectors in the first vector semantic space, and to filter the first ranking results among all the first semantic similarities in combination with preset dynamic business rules, and to determine a preset number of candidate entities in the preset entities. The second vector module 303 is used to input the first representation vector and the second representation vector corresponding to each candidate entity into a preset target encoding network for transformation, so as to obtain the target vector and candidate vector in the second vector semantic space respectively. The target encoding network is trained by contrastive learning, and the semantic discriminative power of the second vector semantic space is greater than that of the first vector semantic space. The second filtering module 304 is used to calculate the second semantic similarity between the target vector and each of the candidate vectors, and to obtain the entity linking result by combining the second ranking result among all the second semantic similarities with counterfactual verification and preset business rule constraints.

[0067] It is understood that the above system item embodiments correspond to the method item embodiments of the present invention, and can implement the entity linking method provided by any of the above method item embodiments of the present invention.

[0068] Furthermore, the step of performing multimodal feature integration and dynamic attribute perception on the target entities in the business text and the preset entities in multiple preset business graphs specifically involves: Obtain multimodal data corresponding to each of the preset business maps, wherein the multimodal data includes data corresponding to text modality, structured modality and semi-structured modality; Modal features corresponding to each modality are extracted using a preset feature extraction model corresponding to each modality. Obtain a preset cross-graph alignment feature set, wherein the cross-graph alignment feature set includes a first feature of the aligned target entity and a second feature of each preset entity; The first feature and each modal feature are fused to obtain the first static attribute vector corresponding to the target entity. For each preset entity, the second feature and each modal feature are fused to obtain the second static attribute vector corresponding to each preset entity. The first static attribute vector and each of the second static attribute vectors are respectively input into a preset temporal feature extraction model to extract temporal information associated with the business text, thereby obtaining the first dynamic attribute vector and the second dynamic attribute vector. The first static attribute vector is concatenated with the corresponding first dynamic attribute vector to obtain the first representation vector, and the second static attribute vector is concatenated with the corresponding second dynamic attribute vector to obtain the second representation vector.

[0069] By integrating textual, structured, and semi-structured trimodal data, the entity representation dimensions are comprehensively covered, and the information abundance of static attribute vectors is improved. The introduction of cross-graph alignment feature sets injects prior alignment information, making the initial distribution of the vector space closer to the true semantic relationships. Dynamic attribute vectors obtained from the temporal feature extraction model can capture entity state drift in real time, allowing representations to be refreshed with business updates, enhancing timeliness. Concatenating static and dynamic vectors allows the first / second representation vectors to simultaneously carry long-term attributes and short-term changes, which helps improve the accuracy of subsequent similarity calculations.

[0070] Furthermore, the entity linking method further includes: Obtain business text, and perform entity recognition on the business text to obtain multiple target entities; The target entity and the entity association identifiers of multiple business graphs are fused to obtain a fusion sequence; The fusion sequence is input into a bidirectional Transformer network to extract the contextual semantics and graph association semantics of the fusion sequence through a self-attention mechanism. The fusion vector is determined by combining the fusion sequence, the contextual semantics, and the graph association semantics. At the same time, the fusion vector is subjected to span analysis, relation classification, and attribute extraction to obtain multiple sets of triples. Based on each of the triples, the target entity and the preset entity are used as nodes, and the relationships between entities are used as edges, to obtain a graph structure across the graph spectrum; The graph structure is updated using a graph attention network to obtain the cross-graph alignment feature set.

[0071] This approach, based on the bidirectional Transformer, generates a fusion sequence that simultaneously encodes contextual semantics and graph association semantics in a single forward pass, improving feature extraction efficiency. Parallel span analysis, relation classification, and attribute extraction yield entity-relation-attribute triples, enhancing extraction consistency. By updating the graph attention network, node features are weighted and aggregated through attention, further improving the discriminative power of feature vectors after cross-graph alignment.

[0072] Furthermore, updating the graph structure based on the graph attention network specifically involves: In the graph attention network, the semantic association weights corresponding to each business graph are dynamically determined based on the business semantic priority of the business text, the attention coefficients between nodes are calculated based on the semantic association weights, and the graph structure is updated according to the attention coefficients.

[0073] This approach dynamically determines the attention coefficient based on business semantic priority, giving higher weight to important graphs, ensuring that node update directions align with business focus, and improving feature alignment accuracy. The attention coefficient is adjusted in real time, allowing graph weights to adaptively refresh as the business context changes, ensuring that the alignment results always closely match the current marketing scenario.

[0074] Furthermore, the second ranking results among all the second semantic similarities, combined with counterfactual verification and preset business rule constraints, yield the entity linking results, specifically as follows: The candidate entities are sorted according to the second semantic similarity to obtain the second sorting result; Based on the second sorting result, at least one initial link relationship between the target entity and the candidate entity is determined; For each set of initial link relationships, the target attribute value of the target vector or the candidate vector is replaced or perturbed to obtain counterfactual samples; The counterfactual sample is input into a preset link model to obtain the corresponding initial counterfactual link relationship. The initial counterfactual link relationship is then checked for consistency and corrected according to business rule constraints to obtain the target counterfactual link relationship. The initial link relationship and the target counterfactual link relationship are compared to calculate the difference degree, and the existence of a link conflict in the initial link relationship is determined based on the difference degree. The entity linking result is obtained by combining all initial link relationships that do not have the aforementioned link conflict.

[0075] By replacing / perturbing the target attributes, counterfactual samples are generated, quickly obtaining a large number of control groups for hypotheses and results, which is beneficial for evaluating the robustness of links; consistency checks are performed through business rule constraints to improve the consistency of link results; the degree of difference is calculated to quantify the degree of conflict, and only links with low degree of difference are retained, ultimately increasing the accuracy of entity link results.

[0076] Furthermore, the entity linking method is specifically as follows: Acquire training data, wherein the training data includes target entity sample data and corresponding candidate entity sample data, the candidate entity sample data includes positive sample data and negative sample data, the positive sample data is entity data that has a link relationship with the target entity sample data, and the negative sample data is entity data that does not have a link relationship with the target entity sample data; The representation vectors corresponding to the target entity sample data and the candidate entity sample data are respectively input into a preset initial encoding network for transformation. By minimizing the distance between the target entity sample data and the positive sample data in the second vector semantic space and maximizing the distance between the target entity sample data and the negative sample data in the second vector semantic space, the initial encoding network is optimized to obtain the target encoding network.

[0077] By employing a contrastive learning training strategy, the distance between positive samples is minimized and the distance between negative samples is maximized, thereby improving the semantic discriminative power of the target encoding network. Based on the joint loss function, the network parameters are updated in the direction of high discriminative power, the semantic space boundary of the second vector is clear, and the similarity calculation error is reduced.

[0078] Furthermore, the entity linking method further includes: Based on the entity linking results, business rules, solutions, risk schemes, and problem priorities associated with the target entity are extracted from the multiple preset business graphs. Combined with the work order requirements of the business text, a structured work order solution is obtained.

[0079] This automatically extracts key information from four types of data: business rules, solutions, risk management plans, and issue priorities, bringing them together in one go and improving information completeness. It also generates structured work order solutions based on work order requirements, resulting in highly executable solutions. Through structured output, the solutions are stored in association with the original work orders, enhancing continuous optimization capabilities.

[0080] It should be noted that the system embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0081] For ease of description and brevity, the system embodiments of the present invention include all the implementation methods described in the above entity linking method embodiments, and will not be repeated here.

[0082] Based on the above-described entity linking method embodiments, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it performs the following steps: Multimodal feature integration and dynamic attribute perception are performed on the target entity in the business text and the preset entities of multiple preset business graphs respectively to obtain the first representation vector corresponding to the target entity in the first vector semantic space, and the second representation vector corresponding to each preset entity. In the first vector semantic space, the first semantic similarity between the first representation vector and each of the second representation vectors is calculated. The first ranking result among all the first semantic similarities is combined with the preset dynamic business rules for filtering, and a preset number of candidate entities are determined in the preset entities. The first representation vector and the second representation vector corresponding to each candidate entity are input into a preset target encoding network for transformation to obtain the target vector and candidate vector in the second vector semantic space, respectively. The target encoding network is trained by contrastive learning, and the semantic discriminative power of the second vector semantic space is greater than that of the first vector semantic space. Calculate the second semantic similarity between the target vector and each of the candidate vectors. Then, combine the second ranking results among all the second semantic similarities with counterfactual verification and preset business rule constraints to obtain the entity linking results.

[0083] This invention integrates multimodal features and dynamic attribute perception to achieve multimodal vector representation of target entities and preset entities, thereby increasing semantic information and providing a high-discrimination foundation for subsequent similarity calculations. First semantic similarity is calculated in a first vector semantic space, enabling rapid quantification of semantic distance between entities during the initial screening stage. Further, combining ranking results with dynamic business rules, the candidate size is accurately compressed, reducing subsequent computational load. Based on the target encoding network trained through contrastive learning, vectors are mapped to a second vector semantic space with higher discriminative power, increasing inter-class distance and decreasing intra-class distance, thus improving similarity resolution. Combining counterfactual verification and business rule constraints, causal consistency verification is performed on the ranking results, quantifying and eliminating link conflicts, ultimately increasing the accuracy and stability of the final entity linking results.

[0084] Furthermore, the step of performing multimodal feature integration and dynamic attribute perception on the target entities in the business text and the preset entities in multiple preset business graphs specifically involves: Obtain multimodal data corresponding to each of the preset business maps, wherein the multimodal data includes data corresponding to text modality, structured modality and semi-structured modality; Modal features corresponding to each modality are extracted using a preset feature extraction model corresponding to each modality. Obtain a preset cross-graph alignment feature set, wherein the cross-graph alignment feature set includes a first feature of the aligned target entity and a second feature of each preset entity; The first feature and each modal feature are fused to obtain the first static attribute vector corresponding to the target entity. For each preset entity, the second feature and each modal feature are fused to obtain the second static attribute vector corresponding to each preset entity. The first static attribute vector and each of the second static attribute vectors are respectively input into a preset temporal feature extraction model to extract temporal information associated with the business text, thereby obtaining the first dynamic attribute vector and the second dynamic attribute vector. The first static attribute vector is concatenated with the corresponding first dynamic attribute vector to obtain the first representation vector, and the second static attribute vector is concatenated with the corresponding second dynamic attribute vector to obtain the second representation vector.

[0085] By integrating textual, structured, and semi-structured trimodal data, the entity representation dimensions are comprehensively covered, and the information abundance of static attribute vectors is improved. The introduction of cross-graph alignment feature sets injects prior alignment information, making the initial distribution of the vector space closer to the true semantic relationships. Dynamic attribute vectors obtained from the temporal feature extraction model can capture entity state drift in real time, allowing representations to be refreshed with business updates, enhancing timeliness. Concatenating static and dynamic vectors allows the first / second representation vectors to simultaneously carry long-term attributes and short-term changes, which helps improve the accuracy of subsequent similarity calculations.

[0086] Furthermore, the entity linking method further includes: Obtain business text, and perform entity recognition on the business text to obtain multiple target entities; The target entity and the entity association identifiers of multiple business graphs are fused to obtain a fusion sequence; The fusion sequence is input into a bidirectional Transformer network to extract the contextual semantics and graph association semantics of the fusion sequence through a self-attention mechanism. The fusion vector is determined by combining the fusion sequence, the contextual semantics, and the graph association semantics. At the same time, the fusion vector is subjected to span analysis, relation classification, and attribute extraction to obtain multiple sets of triples. Based on each of the triples, the target entity and the preset entity are used as nodes, and the relationships between entities are used as edges, to obtain a graph structure across the graph spectrum; The graph structure is updated using a graph attention network to obtain the cross-graph alignment feature set.

[0087] This approach, based on the bidirectional Transformer, generates a fusion sequence that simultaneously encodes contextual semantics and graph association semantics in a single forward pass, improving feature extraction efficiency. Parallel span analysis, relation classification, and attribute extraction yield entity-relation-attribute triples, enhancing extraction consistency. By updating the graph attention network, node features are weighted and aggregated through attention, further improving the discriminative power of feature vectors after cross-graph alignment.

[0088] Furthermore, updating the graph structure based on the graph attention network specifically involves: In the graph attention network, the semantic association weights corresponding to each business graph are dynamically determined based on the business semantic priority of the business text, the attention coefficients between nodes are calculated based on the semantic association weights, and the graph structure is updated according to the attention coefficients.

[0089] This approach dynamically determines the attention coefficient based on business semantic priority, giving higher weight to important graphs, ensuring that node update directions align with business focus, and improving feature alignment accuracy. The attention coefficient is adjusted in real time, allowing graph weights to adaptively refresh as the business context changes, ensuring that the alignment results always closely match the current marketing scenario.

[0090] Furthermore, the second ranking results among all the second semantic similarities, combined with counterfactual verification and preset business rule constraints, yield the entity linking results, specifically as follows: The candidate entities are sorted according to the second semantic similarity to obtain the second sorting result; Based on the second sorting result, at least one initial link relationship between the target entity and the candidate entity is determined; For each set of initial link relationships, the target attribute value of the target vector or the candidate vector is replaced or perturbed to obtain counterfactual samples; The counterfactual sample is input into a preset link model to obtain the corresponding initial counterfactual link relationship. The initial counterfactual link relationship is then checked for consistency and corrected according to business rule constraints to obtain the target counterfactual link relationship. The initial link relationship and the target counterfactual link relationship are compared to calculate the difference degree, and the existence of a link conflict in the initial link relationship is determined based on the difference degree. The entity linking result is obtained by combining all initial link relationships that do not have the aforementioned link conflict.

[0091] By replacing / perturbing the target attributes, counterfactual samples are generated, quickly obtaining a large number of control groups for hypotheses and results, which is beneficial for evaluating the robustness of links; consistency checks are performed through business rule constraints to improve the consistency of link results; the degree of difference is calculated to quantify the degree of conflict, and only links with low degree of difference are retained, ultimately increasing the accuracy of entity link results.

[0092] Furthermore, the entity linking method is specifically as follows: Acquire training data, wherein the training data includes target entity sample data and corresponding candidate entity sample data, the candidate entity sample data includes positive sample data and negative sample data, the positive sample data is entity data that has a link relationship with the target entity sample data, and the negative sample data is entity data that does not have a link relationship with the target entity sample data; The representation vectors corresponding to the target entity sample data and the candidate entity sample data are respectively input into a preset initial encoding network for transformation. By minimizing the distance between the target entity sample data and the positive sample data in the second vector semantic space and maximizing the distance between the target entity sample data and the negative sample data in the second vector semantic space, the initial encoding network is optimized to obtain the target encoding network.

[0093] By employing a contrastive learning training strategy, the distance between positive samples is minimized and the distance between negative samples is maximized, thereby improving the semantic discriminative power of the target encoding network. Based on the joint loss function, the network parameters are updated in the direction of high discriminative power, the semantic space boundary of the second vector is clear, and the similarity calculation error is reduced.

[0094] Furthermore, the entity linking method further includes: Based on the entity linking results, business rules, solutions, risk schemes, and problem priorities associated with the target entity are extracted from the multiple preset business graphs. Combined with the work order requirements of the business text, a structured work order solution is obtained.

[0095] This automatically extracts key information from four types of data: business rules, solutions, risk management plans, and issue priorities, bringing them together in one go and improving information completeness. It also generates structured work order solutions based on work order requirements, resulting in highly executable solutions. Through structured output, the solutions are stored in association with the original work orders, enhancing continuous optimization capabilities.

[0096] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0097] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0098] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0099] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the entity linking method described in any of the above-described method embodiments of the present invention.

[0100] Based on the above-described method embodiments, this invention also provides a computer program / program product, which is stored in a storage medium and executed by at least one processor to implement the various processes of any of the above-described method embodiments, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0101] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0102] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for linking entities, characterized in that, include: Multimodal feature integration and dynamic attribute perception are performed on the target entity in the business text and the preset entities of multiple preset business graphs respectively to obtain the first representation vector corresponding to the target entity in the first vector semantic space, and the second representation vector corresponding to each preset entity. In the first vector semantic space, the first semantic similarity between the first representation vector and each of the second representation vectors is calculated. The first ranking result among all the first semantic similarities is combined with the preset dynamic business rules for filtering, and a preset number of candidate entities are determined in the preset entities. The first representation vector and the second representation vector corresponding to each candidate entity are input into a preset target encoding network for transformation to obtain the target vector and candidate vector in the second vector semantic space, respectively. The target encoding network is trained by contrastive learning, and the semantic discriminative power of the second vector semantic space is greater than that of the first vector semantic space. Calculate the second semantic similarity between the target vector and each of the candidate vectors. Then, combine the second ranking results among all the second semantic similarities with counterfactual verification and preset business rule constraints to obtain the entity linking results.

2. The entity linking method as described in claim 1, characterized in that, The process of integrating multimodal features and dynamically perceiving target entities in business texts and preset entities in multiple preset business graphs specifically involves: Obtain multimodal data corresponding to each of the preset business maps, wherein the multimodal data includes data corresponding to text modality, structured modality and semi-structured modality; Modal features corresponding to each modality are extracted using a preset feature extraction model corresponding to each modality. Obtain a preset cross-graph alignment feature set, wherein the cross-graph alignment feature set includes a first feature of the aligned target entity and a second feature of each preset entity; The first feature and each modal feature are fused to obtain the first static attribute vector corresponding to the target entity. For each preset entity, the second feature and each modal feature are fused to obtain the second static attribute vector corresponding to each preset entity. The first static attribute vector and each of the second static attribute vectors are respectively input into a preset temporal feature extraction model to extract temporal information associated with the business text, thereby obtaining the first dynamic attribute vector and the second dynamic attribute vector. The first static attribute vector is concatenated with the corresponding first dynamic attribute vector to obtain the first representation vector, and the second static attribute vector is concatenated with the corresponding second dynamic attribute vector to obtain the second representation vector.

3. The entity linking method as described in claim 2, characterized in that, The entity linking method further includes: Obtain business text, and perform entity recognition on the business text to obtain multiple target entities; The target entity and the entity association identifiers of multiple business graphs are fused to obtain a fusion sequence; The fusion sequence is input into a bidirectional Transformer network to extract the contextual semantics and graph association semantics of the fusion sequence through a self-attention mechanism. The fusion vector is determined by combining the fusion sequence, the contextual semantics, and the graph association semantics. At the same time, the fusion vector is subjected to span analysis, relation classification, and attribute extraction to obtain multiple sets of triples. Based on each of the triples, the target entity and the preset entity are used as nodes, and the relationships between entities are used as edges, to obtain a graph structure across the graph spectrum; The graph structure is updated using a graph attention network to obtain the cross-graph alignment feature set.

4. The entity linking method as described in claim 3, characterized in that, The step of updating the graph structure according to the graph attention network specifically involves: In the graph attention network, the semantic association weights corresponding to each business graph are dynamically determined based on the business semantic priority of the business text, the attention coefficients between nodes are calculated based on the semantic association weights, and the graph structure is updated according to the attention coefficients.

5. The entity linking method as described in any one of claims 1-4, characterized in that, The second ranking result among all the second semantic similarities, combined with counterfactual verification and preset business rule constraints, yields the entity linking result, specifically as follows: The candidate entities are sorted according to the second semantic similarity to obtain the second sorting result; Based on the second sorting result, at least one initial link relationship between the target entity and the candidate entity is determined; For each set of initial link relationships, the target attribute value of the target vector or the candidate vector is replaced or perturbed to obtain counterfactual samples; The counterfactual sample is input into a preset link model to obtain the corresponding initial counterfactual link relationship. The initial counterfactual link relationship is then checked for consistency and corrected according to business rule constraints to obtain the target counterfactual link relationship. The initial link relationship and the target counterfactual link relationship are compared to calculate the difference degree, and the existence of a link conflict in the initial link relationship is determined based on the difference degree. The entity linking result is obtained by combining all initial link relationships that do not have the aforementioned link conflict.

6. The entity linking method as described in claim 1, characterized in that, The entity linking method is specifically as follows: Acquire training data, wherein the training data includes target entity sample data and corresponding candidate entity sample data, the candidate entity sample data includes positive sample data and negative sample data, the positive sample data is entity data that has a link relationship with the target entity sample data, and the negative sample data is entity data that does not have a link relationship with the target entity sample data; The representation vectors corresponding to the target entity sample data and the candidate entity sample data are respectively input into a preset initial encoding network for transformation. By minimizing the distance between the target entity sample data and the positive sample data in the second vector semantic space and maximizing the distance between the target entity sample data and the negative sample data in the second vector semantic space, the initial encoding network is optimized to obtain the target encoding network.

7. The entity linking method as described in claim 1, characterized in that, The entity linking method further includes: Based on the entity linking results, business rules, solutions, risk schemes, and problem priorities associated with the target entity are extracted from the multiple preset business graphs. Combined with the work order requirements of the business text, a structured work order solution is obtained.

8. A physical linking system, characterized in that, include: First vector module, first filtering module, second vector module, and second filtering module; The first vector module is used to perform multimodal feature integration and dynamic attribute perception on the target entity in the business text and the preset entities of multiple preset business graphs respectively, to obtain the first representation vector corresponding to the target entity in the first vector semantic space, and the second representation vector corresponding to each preset entity. The first filtering module is used to calculate the first semantic similarity between the first representation vector and each of the second representation vectors in the first vector semantic space, and to filter the first ranking results among all the first semantic similarities in combination with preset dynamic business rules, and to determine a preset number of candidate entities in the preset entities. The second vector module is used to input the first representation vector and the second representation vector corresponding to each candidate entity into a preset target encoding network for transformation, so as to obtain the target vector and candidate vector in the second vector semantic space respectively. The target encoding network is trained by contrastive learning, and the semantic discriminative power of the second vector semantic space is greater than that of the first vector semantic space. The second filtering module is used to calculate the second semantic similarity between the target vector and each of the candidate vectors, and to obtain the entity linking result by combining the second ranking result among all the second semantic similarities with counterfactual verification and preset business rule constraints.

9. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the entity linking method as described in claim 1.

10. A computer-readable storage medium, characterized in that, include: A stored computer program, wherein, when the computer program is executed, the device containing the computer-readable storage medium is controlled to perform the entity linking method as described in claim 1.

Images