Entity alignment method and device and electronic equipment

Abstract

The invention discloses an entity alignment method and device and electronic equipment. The method comprises the following steps: acquiring knowledge graph data respectively corresponding to a plurality of heterogeneous communication networks; the knowledge graph data is converted into target data in a preset format, and the preset format comprises a code format which can be understood by the generative model; determining the similarity between entities in the multiple pieces of knowledge graph data, and determining candidate entity pairs of which the similarity is greater than a preset similarity from the entities; and taking the target data and the candidate entity pair as input, and reasoning by adopting a generative model to obtain an aligned target entity pair. According to the entity alignment method and device, the technical problem that the entity alignment accuracy is low due to the fact that related technologies depend on entity name similarity or static rule matching for entity alignment of a cross-domain knowledge graph is solved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a method, apparatus, and electronic device for entity alignment. Background Technology

[0002] For the integration of multi-source heterogeneous data, the relevant technologies mainly rely on the Knowledge Representation Learning (KRL) framework for entity alignment. However, traditional KRL frameworks are difficult to effectively integrate external knowledge, especially in cross-domain data alignment. The lack of deep understanding of specific domains or entities may lead to mismatches during entity alignment. That is, for entity alignment of cross-domain knowledge graphs, the relevant technologies rely on entity name similarity or static rule matching, resulting in low entity alignment accuracy.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This application provides a method, apparatus, and electronic device for entity alignment, which at least solves the technical problem that related technologies for entity alignment of cross-domain knowledge graphs rely on entity name similarity or static rule matching, resulting in low entity alignment accuracy.

[0005] According to one aspect of the embodiments of this application, a method for entity alignment is provided, comprising: acquiring knowledge graph data corresponding to multiple heterogeneous communication networks respectively, wherein entities in the knowledge graph data are used to represent network devices in the communication networks; converting the knowledge graph data into target data in a preset format, wherein the preset format includes a code format that can be understood by a generative model; determining the similarity between entities in the multiple knowledge graph data, and determining candidate entity pairs from the entities whose similarity is greater than a preset similarity, wherein each candidate entity pair consists of entities from different knowledge graph data; using the target data and candidate entity pairs as input, and employing a generative model to infer and obtain aligned target entity pairs, wherein the target entity pairs include entity pairs among the candidate entity pairs whose confidence is greater than a preset confidence.

[0006] In some embodiments of this application, converting knowledge graph data into target data in a preset format includes: extracting network topology information from the knowledge graph data, wherein the network topology information includes entity information, relationship information describing the connections between network devices, and attribute information describing the characteristics of network devices; and reorganizing the network topology information according to a preset format to obtain target data, wherein the preset format encapsulates the network topology information according to the hierarchical relationship between entity information, relationship information, and attribute information.

[0007] In some embodiments of this application, determining the similarity between entities in multiple knowledge graph data includes: mapping the multiple knowledge graph data using a knowledge representation learning framework to obtain multiple low-dimensional dense vectors, wherein the knowledge representation learning framework is used to transform entities and topological relationships in the knowledge graph data into mathematical representations, and each entity corresponds to a low-dimensional dense vector; determining the cosine similarity between a first low-dimensional dense vector and a second low-dimensional dense vector among the multiple low-dimensional dense vectors, wherein the first entity corresponding to the first low-dimensional dense vector and the second entity corresponding to the second low-dimensional dense vector are entities in different knowledge graph data; and determining the cosine similarity as the similarity between the first entity and the second entity.

[0008] In some embodiments of this application, target data and candidate entity pairs are used as input, and a generative model is used to infer aligned target entity pairs. This includes: determining a query instruction based on the target data and candidate entity pairs, wherein the query instruction is used to guide the generative model to perform logical verification on the candidate entity pairs based on at least one of the following: entity name similarity, topological structure association, and attribute consistency; for each candidate entity pair, the generative model is used to parse the query instruction corresponding to the candidate entity pair to obtain a verification result, wherein the verification result is used to determine the matching degree of the candidate entity pairs; and determining the target entity pair based on the verification result.

[0009] In some embodiments of this application, determining the target entity pair based on the verification result includes: for each candidate entity pair, determining the confidence level corresponding to the verification result; if the confidence level is greater than a preset confidence level, determining the candidate entity pair as the target entity pair.

[0010] In some embodiments of this application, the method further includes: determining the alignment criteria of the target entity pair based on the verification results, wherein the alignment criteria are used to describe the matching criteria and reasoning process of the target entities in the target entity pair.

[0011] In some embodiments of this application, the method further includes: determining an update query instruction based on the verification result when the confidence level is less than or equal to a preset confidence level, wherein the update query instruction is used to guide the generative model to expand the search range of the target data; and using the generative model to parse the update query instruction to obtain the update verification result.

[0012] In some embodiments of this application, the knowledge graph data includes the raw data of the Product Service Resource (PSR) graph, wherein resource objects in the PSR graph represent network devices of a communication network, network service objects in the PSR graph represent network services established based on multiple network devices, and business service objects in the PSR graph represent business services established based on at least one network service.

[0013] According to another aspect of the embodiments of this application, an entity alignment apparatus is also provided, comprising: an acquisition module, configured to acquire knowledge graph data corresponding to multiple heterogeneous communication networks respectively, wherein entities in the knowledge graph data are used to represent network devices in the communication networks; a conversion module, configured to convert the knowledge graph data into target data in a preset format, wherein the preset format includes a code format that can be understood by a generative model; a determination module, configured to determine the similarity between entities in the multiple knowledge graph data, and determine candidate entity pairs from the entities whose similarity is greater than a preset similarity, wherein each candidate entity pair consists of entities from different knowledge graph data; and a reasoning module, configured to take the target data and candidate entity pairs as input, and use a generative model to reason to obtain aligned target entity pairs, wherein the target entity pairs include entity pairs among the candidate entity pairs whose confidence is greater than a preset confidence.

[0014] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory and a processor, wherein the memory is used to store program instructions; and the processor is connected to the memory and used to execute the method for implementing the above-described entity alignment.

[0015] According to another aspect of the embodiments of this application, a non-volatile storage medium is also provided, the non-volatile storage medium including a stored computer program, wherein the device on which the non-volatile storage medium is located executes the above-described entity alignment method by running the computer program.

[0016] According to another aspect of the embodiments of this application, a computer program product is also provided, including computer instructions that, when executed by a processor, implement the above-described entity alignment method.

[0017] In this embodiment, knowledge graph data from multiple heterogeneous communication networks is acquired and integrated, and converted into a preset code format that can be efficiently parsed by a generative model. Similarity analysis technology is used to initially screen out candidate entity pairs with high similarity. Then, through deep conversational reasoning, a generative model is used to screen out target entity pairs with high confidence for the second time. This achieves the goal of enhancing the accuracy of entity alignment, thereby avoiding matching errors caused by differences in naming rules. This solves the technical problem that entity alignment for cross-domain knowledge graphs relies on entity name similarity or static rule matching, resulting in low entity alignment accuracy. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1This is a hardware structure block diagram of a computer terminal for an entity alignment method according to an embodiment of this application;

[0020] Figure 2 This is a flowchart of an entity alignment method according to an embodiment of this application;

[0021] Figure 3 This is a schematic flowchart of an entity alignment method according to an embodiment of this application;

[0022] Figure 4 This is a schematic diagram of a physical alignment device according to an embodiment of this application. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] To better understand the embodiments of this application, the technical terms involved in the embodiments of this application are explained below:

[0026] PSR Model (Product, Service, Resource Model): The PSR model is an architectural design methodology for the telecommunications industry and cloud computing. It decouples and abstracts communication products, services, and underlying resources according to certain logical layers. The PSR model emphasizes the flexible combination and rapid iteration of the underlying hardware resources to user-facing services and products by defining clear resource (RES), service (RFS / CFS) and product (Product) layers.

[0027] Resource (RES): In the telecommunications and cloud computing industries, RES refers to the basic units that constitute network services, including hardware devices, software components, and data resources. In the embodiments of this application, RES is a fundamental component of the PSR model, used to describe and define the attributes, relationships, and functions of various network devices in the communication network, providing core data support for subsequent entity alignment.

[0028] Resource Facing Service (RFS): Refers to an abstract layer built on top of RES that provides services to internal network resources. It aims to achieve specific network functions through the combination and encapsulation of RES. In this embodiment, RFS integrates multiple RES to form a higher-level service unit. During entity alignment, it serves as an important carrier of logical relationships and attribute information between entities, helping to improve the accuracy and relevance of entity alignment.

[0029] Customer Facing Service (CFS): A combination of services customized for end-user needs and directly facing customers. It consists of one or more RFS. In this embodiment, CFS represents business functions and service requirements from the customer's perspective. In the entity alignment strategy, it is used to verify the consistency of business logic in entity matching and ensure the applicability of the alignment result to the customer's business scenario.

[0030] It should be noted that multiple resources (RES) constitute a network service (RFS), one or more network services constitute a service (CFS), and one or more services constitute a communication service / product (Product).

[0031] Knowledge Representation Learning (KRL) is a machine learning technique that aims to learn low-dimensional vector representations of entities and relations from knowledge graphs to capture their inherent semantics and structure. In this application embodiment, KRL is used for the initial representation and screening of entities. By calculating the similarity of entity embeddings, potential matching candidates are quickly located, providing an efficient basis for subsequent reasoning.

[0032] Large Language Models (LLMs) are deep learning models trained on large-scale corpora that possess powerful text understanding and generation capabilities. In this embodiment, LLMs serve as advanced inference engines in the entity alignment process, enabling multi-turn conversational verification based on structured entity descriptions and contextual information, thereby improving the accuracy and interpretability of entity alignment.

[0033] Entity alignment is a core technology in knowledge graph integration, aiming to identify and match identical entities from different graphs to establish cross-graph entity mapping relationships. In this embodiment, entity alignment is achieved through a two-stage strategy. The first stage uses KRL to coarsely screen candidate entity pairs with high similarity, and the second stage uses LLMs for deep conversational verification, ultimately outputting high-confidence target entity pairs. This solves the naming differences and rule matching problems in cross-domain PSR graph integration, improving the accuracy and efficiency of entity alignment.

[0034] The PSR model design philosophy is to layer and decouple resources, services, and products, and assemble them based on underlying cloud and network capabilities to form services, providing communication services / products externally. The Product Service Design Center, as a visual design tool for the next-generation cloud and network operation business system, is a crucial component of the development framework. As a business design-oriented system, it provides product design loading capabilities based on reusable and decoupled cloud and network capabilities. By constructing a PSR product loading system, it achieves rapid product loading and realizes the goal of cloud and network self-intelligence. In the PSR model, PSR objects include Business Service Objects (CFS), Network Service Objects (RFS), and Resource Objects (RES).

[0035] In the process of integrating various types of PSR knowledge graphs, the relevant technologies adopt entity alignment methods based on knowledge representation learning (KRL), but the results are unsatisfactory. These methods have two major limitations: first, they lack external knowledge fusion; second, they rely on black-box similarity calculation and lack interpretable reasoning, which leads to limited performance in heterogeneous KG alignment tasks.

[0036] To address the aforementioned technical problems, this application provides corresponding solutions, which are detailed below.

[0037] The entity alignment method embodiments provided in this application can be executed on mobile terminals, computer terminals, or similar computing devices. Figure 1 A hardware block diagram of a computer terminal for implementing a method of entity alignment is shown. Figure 1As shown, the computer terminal 10 may include one or more processors (shown as 102a, 102b, ..., 102n in the figure) (the processor may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission module 106 for communication functions connected via wired and / or wireless networks. In addition, it may also include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, and a BUS bus. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0038] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10. As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0039] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the entity alignment method in this embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby implementing the aforementioned entity alignment method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0040] The transmission module 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission module 106 includes a network interface controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission module 106 may be a radio frequency (RF) module, used for wireless communication with the Internet.

[0041] The display can be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10.

[0042] It should be noted here that, in some optional embodiments, the above... Figure 1 The computer terminal shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 1 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computer terminal.

[0043] In the above operating environment, this application provides an embodiment of a method for entity alignment. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0044] Figure 2 This is a flowchart of an entity alignment method according to an embodiment of this application, such as... Figure 2 As shown, the method includes the following steps:

[0045] Step S202: Obtain knowledge graph data corresponding to multiple heterogeneous communication networks, wherein the entities in the knowledge graph data are used to represent network devices in the communication network.

[0046] In step S202 above, heterogeneous communication networks refer to communication networks of different operators or regions, or network systems with different technical architectures (such as 4G and 5G networks). Differences between heterogeneous networks include, but are not limited to, differences in device naming, service configuration, and system architecture. Knowledge graph data is a structured collection of data, which may contain entities, entity attributes, and relationships between entities. It organizes information in the form of a graph to depict the relationships between devices, services, and products in a communication network.

[0047] In some embodiments of this application, knowledge graph data from different communication networks can be collected. This data typically includes the names, attributes, and relationships of devices in their respective network systems. Subsequently, data cleaning and standardization processes can be performed to convert data with different formats, encodings, or naming conventions into a unified standard format, so as to facilitate subsequent entity alignment algorithm recognition and processing.

[0048] In some embodiments of this application, the knowledge graph data includes the raw data of the Product Service Resource (PSR) graph, wherein resource objects in the PSR graph represent network devices of a communication network, network service objects in the PSR graph represent network services established based on multiple network devices, and business service objects in the PSR graph represent business services established based on at least one network service.

[0049] It should be noted that the PSR graph is a graph model specifically designed to describe the relationships between products, services, and resources in a communication network. It unifies network devices, service logic, and product design within a multi-layered architecture, facilitating management and automated operation. The PSR graph includes:

[0050] (1) Resource Object (RES): RES is the most basic building block in the PSR graph, representing network devices in the communication network, including but not limited to routers, switches, servers, etc. RES can be used as the basic element for matching in entity alignment.

[0051] (2) Network Service Object (RFS): RFS is a logical layer composed of RES, which describes how to implement a specific network service by combining multiple RES, such as IP telephony, broadband access service, etc. In entity alignment, RFS can be used to verify whether candidate entity pairs can form a consistent service logic at the network service level.

[0052] (3) Business Service Object (CFS): CFS is located above RFS and is composed of one or more RFS. It is directly facing customers and provides a combination of services to meet specific business needs, such as enterprise cloud solutions and home entertainment packages. In entity alignment, CFS can be used to finally confirm whether entity pairs can be seamlessly connected at the business level, and ensure that the alignment results meet business objectives and user experience.

[0053] Specifically, RES data (i.e., raw data) can be obtained from multiple heterogeneous communication networks and standardized to eliminate issues such as naming differences and missing attributes. For example, a unified RES attribute dictionary can be established, mapping all RES attributes to this dictionary to ensure attribute consistency and completeness. Subsequently, the information from RFS and CFS can be represented in a structured format, such as converting it to JSON, enabling the large language model to better understand the logical relationships between entities. This approach involves deep parsing and reorganization of the raw data, ensuring that the model can capture complex relationships at the service and business levels. Through structured representation, RFS and CFS data can be processed more effectively by the large language model, which is crucial for understanding deep relationships between entities and achieving accurate service alignment.

[0054] In the PSR map, entity alignment can include at least one of the following: resource object alignment, network service object alignment, business service object alignment, and product alignment, wherein:

[0055] (1) Resource object alignment: This is the most basic type of entity alignment, involving physical and virtual resources in the communication network, such as network devices (switches, routers), servers, storage devices, etc. The purpose of RES alignment is to identify and match devices with the same or similar functions in cross-domain communication networks, laying the foundation for subsequent service and product level alignment.

[0056] (2) Network Service Object Alignment: An RFS entity is composed of one or more RES entities, representing specific network functions and services, such as broadband access, virtual private network (VPN) services, etc. RFS alignment is an advanced matching based on RES alignment, which aims to identify homogeneous or similar network services in different networks. This is crucial for cross-carrier service integration and network optimization.

[0057] (3) Business service object alignment: CFS entities are built on one or more RFS entities, directly facing customers and providing a combination of services that meet specific business needs. CFS alignment focuses on the consistency and interoperability of business services, ensuring that business services in different communication networks or product designs can be accurately identified and integrated to support cross-domain business expansion and customer experience improvement.

[0058] (4) Product alignment: Product-level alignment involves identifying and matching the same or similar products in different PSR maps. Products are usually composed of one or more CFS entities, representing the complete communication solutions or services that are eventually launched to the market. Product alignment helps to achieve business integration and resource optimization in cross-domain environments, ensuring that products from different regions or operators can be seamlessly connected to meet national or global business needs.

[0059] In some embodiments of this application, alignment can begin with the most basic RES alignment and gradually move upwards to RFS, CFS, and finally product alignment. Each level of alignment relies on accurate matching at the next level, forming a bottom-up entity matching chain to ensure the integration and optimization of the entire PSR map. For example, in a provincial-level communication network resource integration project, it is first necessary to identify network devices of the same type through RES alignment, then ensure the consistency of service logic based on these devices at the RFS level, and finally confirm the matching of business services and final products at the CFS and product levels to achieve comprehensive resource and service integration and avoid redundant construction and operational conflicts.

[0060] Step S204: Convert the knowledge graph data into target data in a preset format, wherein the preset format includes a code format that the generative model can understand.

[0061] In step S204 above, the target data refers to converting the original knowledge graph data into a specific format, such as JSON, that can be understood and processed by specific types of models (e.g., large language models), in order to better activate the reasoning capabilities of large language models (LLMs). Generative models, such as large pre-trained language models, are capable of understanding and generating human language. By converting knowledge graph data into a code format (e.g., JSON), they can ensure that the model can accurately parse the attributes and relationships of entities, rather than simply treating it as ordinary text data.

[0062] In some embodiments of this application, entity, attribute, and relation information in the PSR graph can be converted into JSON-formatted key-value pairs and arrays to form a data structure that is easy for large language models to understand and process. In the JSON representation of an entity, in addition to basic information, detailed attribute and relation descriptions can also be included. These descriptions can be presented in a coded manner to facilitate logical reasoning and attribute comparison by the model.

[0063] Raw PSR graph data may exist in the form of triples or unstructured text, which is not conducive to large language models directly understanding and inferring relationships between entities. By converting it to JSON format, not only is the detailed information of the entities preserved, but a clear representation of the relationships between entities is also provided, which helps the model to perform multi-step reasoning. Entity attributes and relationship information are key to multi-dimensional entity alignment. Through coded description, the model can access this information in a structured way, perform accurate attribute comparisons and relationship reasoning, especially in cross-carrier equipment management scenarios, where entities in different networks may have similar but differently expressed attributes (such as "bandwidth" and "transmission bandwidth"). Through coded description, the model can identify the equivalence relationships between them, improving the accuracy and efficiency of entity alignment.

[0064] To enable generative models to accurately understand the logical structure and hierarchical relationships between entities, knowledge graph data can be converted into target data in a preset format as follows: network topology information is extracted from the knowledge graph data, including entity information, relationship information describing the connections between network devices, and attribute information describing the characteristics of network devices; the network topology information is then reorganized according to a preset format to obtain the target data, wherein the preset format encapsulates the network topology information according to the hierarchical relationship between entity information, relationship information, and attribute information.

[0065] It should be noted that network topology information includes entity information, connection relationship information, and attribute information. Entity information refers to data describing the network devices themselves, such as name, type, and ID. Relationship information describes the connection between network devices, such as "connected devices" and "used protocols." Attribute information covers detailed parameters describing the characteristics and status of network devices, such as bandwidth, protocol version, and management interface.

[0066] Specifically, entity information, including entity ID, name, and type, can be extracted from the original knowledge graph (taking the PSR graph as an example). The connections between entities and the attribute information describing entity characteristics are encoded into key-value pairs or arrays in a preset format to form a structured data representation. In the target data in JSON or other preset formats, data encapsulation follows the hierarchical relationship of the PSR model. That is, RES, RFS, CFS, and PRODUCT are successively contained in higher-level entities, forming a hierarchical data structure. For example, each RFS entity contains information about its related RES entities, while the CFS entity contains RFS and higher-level attributes and relationships.

[0067] The original data may not explicitly express hierarchical relationships, making it difficult for the model to understand the logical structure and belonging relationships between entities. The embodiments of this application encapsulate hierarchical relationships, which can reflect the contextual connections between entities in the data, making it easier for the model to perform logical reasoning and association judgment during the entity alignment process. Especially in the scenario of heterogeneous knowledge graph integration, hierarchical relationship encapsulation helps the model identify the entity alignment rules across graphs, improving alignment efficiency and accuracy.

[0068] To facilitate understanding of the above data conversion process, the following explanation will be provided with reference to some specific examples.

[0069] Traditional knowledge graph data (such as RDF, triples) is not intuitive for LLMs, while code formats (such as structured dictionaries, natural language descriptions) can help LLMs extract and reason about knowledge more efficiently. Converting the structured data (triples, entities, relationships) of the knowledge graph into a JSON format suitable for large language models to understand and process can enable LLMs to better utilize their reasoning ability for entity alignment. For example:

[0070] Original KG data (triple format):

[0071] <A8-C device RES, connected device, OLT device RES>

[0072] <A8-C device RES, uses protocol, IPsec>

[0073] <A8-C device RES, function, enterprise dedicated line access>

[0074] <A8-C device RES, management protocol, SNMPv3>

[0075] The above format is not easy for LLMs to understand and may be misparsed as ordinary text.

[0076] Converted format (JSON example):

[0077] {

[0078] "entities": {

[0079] "A8-C device RES": {

[0080] "type": "network device",

[0081] "attributes": {

[0082] "connected device": ["OLT device RES"],

[0083] "protocol": ["IPsec", "SNMPv3"],

[0084] "function": "enterprise dedicated line access"

[0085] }

[0086] }

[0087] }

[0088] }

[0089] Step S206: Determine the similarity between entities in multiple knowledge graph data, and identify candidate entity pairs with a similarity greater than a preset similarity from the entities, wherein each candidate entity pair consists of entities from different knowledge graph data.

[0090] In step S206 above, candidate entity pairs refer to entity pairs whose similarity is greater than a preset threshold selected from all entity pairs after calculating the similarity between entities. These pairs are then used as candidate objects for further verification and matching. It should be noted that the main focus of entity alignment is cross-knowledge graph entity matching. Specifically, when two entities are considered similar or potentially matching, this set of entity pairs (i.e., candidate entity pairs) consists of entities belonging to two or more different knowledge graphs. For example, when integrating resources in a communication network, there may be PSR graphs from different provinces or different operators. Each graph describes its own resources, services, and products. The purpose of entity alignment is to find functionally similar entities that can be matched with each other among these multiple knowledge graphs, even if they may use different naming or description methods in different graphs. By identifying these candidate entity pairs composed of entities from different knowledge graph data, consistency and interoperability of cross-domain resources, services, and products can be achieved.

[0091] In some embodiments of this application, a candidate entity pair may consist of two entities or multiple entities, and this is not limited thereto.

[0092] When calculating similarity, it's not enough to consider only the similarity of entity names. A comprehensive evaluation can also be achieved by combining entity attribute and relationship information. For example, by assigning different weights to entity attributes and relationships, a comprehensive similarity evaluation model can be constructed, thus calculating the similarity between entities more comprehensively. Simply calculating name similarity may overlook deep connections between entities, leading to inaccurate entity alignment results. Comprehensive similarity evaluation can take into account the full picture of entities; even if entity names differ, potential entity alignment relationships can be discovered through attribute and relationship matching.

[0093] To quickly filter out obviously dissimilar entity pairs and reduce the workload and computational resource consumption in subsequent fine screening stages, the similarity between entities in multiple knowledge graph datasets can be determined as follows: A knowledge representation learning framework is used to map multiple knowledge graph datasets, resulting in multiple low-dimensional dense vectors. This framework transforms entities and topological relationships in the knowledge graph datasets into mathematical representations, with each entity corresponding to a low-dimensional dense vector. The cosine similarity between the first and second low-dimensional dense vectors is then determined. The first entity corresponding to the first low-dimensional dense vector and the second entity corresponding to the second low-dimensional dense vector are entities from different knowledge graph datasets. This cosine similarity is then used to determine the similarity between the first and second entities.

[0094] It should be noted that a low-dimensional dense vector refers to a vector that compresses the information of an entity into a low-dimensional space through a knowledge representation learning framework. Each dimension represents a certain feature or attribute of the entity. Dense representation means that most dimensions in the vector have non-zero values.

[0095] Specifically, entity embedding can be achieved using the TransE model (a knowledge representation learning framework). TransE assumes that an entity relation can be represented as an addition and subtraction operation between entity vectors; that is, entity A can be transformed into entity B through relation r in vector form: A + r ≈ B. By constructing a loss function, low-dimensional dense vector representations of each entity and relation are obtained through training. In cross-domain PSR graph alignment, entities may have multiple representations in different graphs. The TransE model can capture the mathematical representation of entity relations, and even if the entity names are different, it can calculate similarity based on entity attributes and relations, thereby identifying potential alignment entities.

[0096] Step S208: Using the target data and candidate entity pairs as input, a generative model is used to infer aligned target entity pairs, wherein the target entity pairs include entity pairs among the candidate entity pairs whose confidence is greater than a preset confidence.

[0097] In step S208 above, after screening candidate entity pairs through similarity calculation, generative model reasoning is further used to determine those entity pairs that pass the reasoning verification and have a confidence level greater than a preset threshold as target entity pairs. These entity pairs are considered to be entity combinations that are accurately aligned between different knowledge graphs.

[0098] In some embodiments of this application, candidate entity pairs and their related contexts (such as attributes, relationships, etc.) can be transformed into structured instructions and input into a generative model. The model performs logical reasoning based on information such as entity name similarity, topological association, and attribute consistency through multi-step dialogic interaction, and outputs alignment results, such as providing confidence scores and reasoning basis. Subsequently, based on the confidence scores provided by the generative model, a pre-set confidence threshold is set, and candidate entity pairs with confidence scores greater than this threshold are identified as target entity pairs.

[0099] Traditional entity alignment methods may rely solely on surface similarity, lacking in-depth understanding and reasoning. By introducing conversational reasoning through generative models, even with significant surface differences, the model can identify and verify the validity of entity alignment if the underlying logic is consistent. In cross-carrier equipment management scenarios, different naming rules or description methods may lead to inaccurate similarity calculations, but generative model reasoning can compensate for this deficiency and accurately identify alignment relationships between entities.

[0100] To enable generative models to understand more complex alignment rules and correctly identify matching relationships between entities, target entity pairs can be determined as follows: A query instruction is determined based on the target data and candidate entity pairs. This query instruction guides the generative model to logically verify the candidate entity pairs based on at least one of the following: entity name similarity, topological association, and attribute consistency. For each candidate entity pair, the generative model parses the query instruction corresponding to the candidate entity pair to obtain a verification result. This verification result is used to determine the degree of matching between the candidate entity pairs. The target entity pair is then determined based on the verification result.

[0101] The steps for determining the target entity pair based on the verification results are as follows: for each candidate entity pair, determine the confidence level corresponding to the verification result; if the confidence level is greater than the preset confidence level, determine the candidate entity pair as the target entity pair.

[0102] In addition, the following steps can be performed: determine the alignment criteria for the target entity pair based on the verification results, wherein the alignment criteria are used to describe the matching criteria and reasoning process of the target entities in the target entity pair.

[0103] It should be noted that during the entity alignment process, the query instructions are carefully designed natural language sentences or structured data queries, which are intended to guide the generative model to perform in-depth logical verification of candidate entity pairs. The query instructions can explicitly instruct the model to focus on key alignment factors such as entity name similarity, topological structure association, and attribute consistency.

[0104] Specifically, a series of query commands presented in natural language can be designed, such as "verify whether entity A8-C device RES and government / enterprise gateway match." The entity name and expected matching conditions are integrated into the command, guiding the generative model to consider not only name similarity but also topological relationships and attribute consistency during entity alignment. The generative model does not provide the final alignment result all at once, but rather deepens its understanding of candidate entity pairs through multiple iterations. After each iteration, the focus and strategy of subsequent inference are adjusted based on the verification results output by the model until a satisfactory confidence level is reached. For each candidate entity pair, the generative model provides a confidence score after inference verification, reflecting the degree of matching between the entity pairs. By setting a reasonable pre-confidence threshold, the model can filter out entity pairs with high confidence and identify them as the target entity pairs for alignment.

[0105] Entity alignment tasks may involve complex logical verification. A single inference step may not be able to fully assess the relationship between entity pairs. Multi-round iterative inference can give the model ample opportunity to explore and verify the matching basis between entities. In particular, when entity information is incomplete or uncertain, the model can obtain a more accurate alignment result through multiple attempts and corrections.

[0106] To facilitate understanding of the process of determining the target entity pairs described above, the following explanation is provided in conjunction with some specific embodiments.

[0107] (1) Two-stage entity alignment-embedding similarity coarse screening:

[0108] Using knowledge representation learning (KRL) frameworks (such as TransE and RotatE), joint embedding representation learning is performed on the names and topological relationships of PSR objects to generate low-dimensional dense vectors. By measuring the cosine similarity of cross-graph entity embeddings (such as a threshold > 0.8), Top-K high-confidence candidate matching pairs are screened to construct an initial alignment set for subsequent fine screening.

[0109] (2) Two-stage entity alignment - LLMs multi-step inference verification:

[0110] Using candidate entity pairs obtained from embedding similarity coarse screening and JSON-formatted graph context as input, the LLM's conversational reasoning is triggered by constructing structured query commands (such as 'verify whether the A8-C device RES and the government and enterprise gateway match'). The LLM's multi-dimensional verification is triggered by structured natural language commands, such as multi-round logical verification based on entity name similarity, topological structure association, and attribute consistency. The output is an alignment result (i.e., verification result) with confidence score (such as 'matching probability 92%') and interpretable evidence, which solves the problem of uninterpretable decision-making in traditional black box models.

[0111] To flexibly adjust the search strategy according to different scenarios and entity characteristics, the following steps can also be performed: when the confidence level is less than or equal to the preset confidence level, determine the update query instruction based on the verification results, wherein the update query instruction is used to guide the generative model to expand the search scope of the target data; the generative model is used to parse the update query instruction to obtain the update verification results.

[0112] Specifically, when the model has low confidence in the validation results of candidate entity pairs under the initial query command, the specific feedback information in the validation results can be analyzed (such as which entity attributes or relationship information the model believes do not match), and the query command can be updated based on this feedback to enrich the description of entity pairs by introducing more contextual information, entity attributes or related entities, thereby guiding the model to perform more comprehensive validation.

[0113] For candidate entity pairs with low confidence, updating the query instruction can guide the model to not only focus on directly related entity attributes and relationships, but also consider broader contextual information and indirectly related entities, such as the entity's neighboring nodes and the higher-level PSR type to which the entity belongs, in order to expand the search scope and capture more comprehensive matching criteria.

[0114] During entity alignment, some entity pairs may not be considered a match due to superficial differences (such as names or direct attributes), but in reality, they may be similar in deeper network structures or functions. Dynamically adjusting the search scope can ensure that the model does not miss potential matching entity pairs due to overly restrictive search strategies. For example, in the integration of provincial communication network resources, the naming rules of equipment from different provincial companies may lead to lower direct attribute matching scores. However, by expanding the search scope and considering the position of entities in the topology and their indirect relationships, deeper matching relationships can be discovered, improving the efficiency and accuracy of entity alignment.

[0115] Based on feedback from updated query commands, the model can progressively learn and optimize its search strategy. For example, it can learn how to automatically adjust the depth and breadth of queries based on factors such as entity type, attributes, and topology to find the most relevant and matching entity pairs. Entity alignment is a dynamic process that requires flexible adjustments to the search strategy according to different scenarios and entity characteristics. The iterative optimization of the search strategy described above enables the model to continuously improve itself in practical applications, enhancing alignment efficiency and accuracy.

[0116] In some embodiments of this application, the search range can also be automatically adjusted based on a confidence threshold (i.e., a preset confidence level). Specifically, in the initial stage of entity alignment, the system may adopt a relatively broad search range to ensure that as many potential matching entities as possible are captured. However, as the alignment process progresses, to avoid ineffective calculations on a large number of irrelevant entities, the system can automatically adjust the depth and breadth of the search based on the confidence scores of the entity pairs. The confidence threshold is a decisive factor, which can be dynamically adjusted based on the current alignment results and resource consumption. For example, if the confidence of the current entity pairs is generally high, the threshold can be appropriately increased to reduce the search range; conversely, the threshold can be decreased to expand the search range in order to find more potential matching entity pairs. This approach allows the system to find a balance between resource utilization and alignment accuracy, avoiding resource waste while ensuring that all possible matching entities are found.

[0117] Furthermore, iterative optimization can be performed on conflict cases. Specifically, during entity alignment, some conflict cases may be encountered, where two or more entities match another entity simultaneously, but there may be competition between them. For such cases, the system can perform iterative optimization, meaning it will continuously try different strategies and methods to resolve these conflicts until it finds the most reasonable entity alignment result. For example, this may involve a deeper analysis of entity attributes, topological relationships, and functional details, as well as updating and refining query commands to guide the generative model to more accurately analyze entity pairs, eliminate conflicts, and find the best match. In addition, iterative optimization strategies may also include adjusting model parameters to improve its reasoning ability and accuracy on conflict cases, which is not limited here.

[0118] Through steps S202 to S208, knowledge graph data from multiple heterogeneous communication networks are acquired and integrated, and converted into a preset code format that can be efficiently parsed by a generative model. Similarity analysis technology is used to initially screen out candidate entity pairs with high similarity. Then, through deep conversational reasoning, the generative model is used to screen out target entity pairs with high confidence for the second time, thereby enhancing the accuracy of entity alignment. This achieves the technical effect of avoiding matching errors caused by differences in naming rules, and solves the technical problem that entity alignment for cross-domain knowledge graphs relies on entity name similarity or static rule matching, resulting in low entity alignment accuracy.

[0119] Figure 3 This is a schematic diagram of the overall process of an entity alignment method according to an embodiment of this application, as shown below. Figure 3 As shown, in some embodiments of this application, entity alignment includes the following steps:

[0120] (1) PSR knowledge graph JOSN structured representation.

[0121] Specifically, the entities and relationships in the PSR knowledge graph are converted into JSON format in a structured way to meet the processing requirements of Large Language Models (LLMs). Converting the PSR knowledge graph into JSON format clearly presents information such as entities, attributes, and topological relationships in the form of key-value pairs, making it easier for LLMs to understand and process.

[0122] (2) Use knowledge representation learning to vectorize PSR objects and topological relationships.

[0123] Specifically, through KRL, PSR objects and their topological relationships can be represented as multi-dimensional vectors, i.e., embeddings. This process involves jointly training entities and relationships in the graph to learn vector representations that reflect the topological relationships and attribute features between entities. For example, the TransE model represents the relationships between entities as the vector difference from one entity vector to another, i.e.:

[0124] Entity vector e1 + relation vector r ≈ entity vector e2.

[0125] Here, e1, r, and e2 represent the vector representation of an entity, the vector representing the relationship between them, and the vector representation of another entity, respectively. By minimizing the distance function (such as L1 or L2 distance), the KRL model can learn low-dimensional dense vectors representing PSR objects and topological relationships.

[0126] (3) Use cosine similarity to perform embedding similarity coarse screening.

[0127] Specifically, after obtaining the vector representation of the PSR object, cosine similarity can be used to measure the similarity between entity vectors in different graphs. Its value ranges from -1 to 1, where 1 indicates complete similarity (completely in the same direction), -1 indicates complete dissimilarity (completely opposite direction), and 0 indicates orthogonality (no correlation). By setting an appropriate confidence threshold (e.g., 0.8), entity pairs exhibiting high similarity in the vector space can be selected, constructing a preliminary alignment candidate set and providing a foundation for subsequent fine-tuning and inference verification.

[0128] (4) Using the candidate entity pairs obtained from the embedding similarity coarse screening and the JOSNized graph context as input, perform multi-step reasoning verification of the large model.

[0129] Based on the candidate entity pairs obtained in the coarse screening stage and the JSON-based graph context, large models (such as the GPT series models) perform fine screening and verification of candidate entity pairs through conversational multi-step reasoning. Specifically, structured query commands can be constructed to trigger logical verification of entity alignment in LLMs. Leveraging their rich external knowledge base and powerful reasoning capabilities, comprehensive verification is performed from multiple dimensions such as entity name similarity, topological structure association, and attribute consistency.

[0130] For example, for a pair of candidate entities A and B, the following query instruction can be constructed: "Verify whether entities A and B are matched through a common connection device Z and the same management protocol." Based on this instruction, LLMs will perform verification through multi-turn dialogue (i.e., multi-step reasoning), utilizing the entity's attributes, relationships, and graph context information, and output an alignment result (i.e., verification result) with confidence scores and interpretable evidence. This verification process not only reduces the over-reliance on embedded similarity but also provides transparency and interpretability of decision-making through multi-step reasoning, significantly improving the accuracy and efficiency of entity alignment.

[0131] Figure 4 This is a structural diagram of an entity alignment device according to an embodiment of this application, such as... Figure 4 As shown, the device includes:

[0132] The acquisition module 402 is used to acquire knowledge graph data corresponding to multiple heterogeneous communication networks, wherein the entities in the knowledge graph data are used to represent network devices in the communication network;

[0133] The conversion module 404 is used to convert knowledge graph data into target data in a preset format, wherein the preset format includes a code format that can be understood by the generative model.

[0134] The determination module 406 is used to determine the similarity between entities in multiple knowledge graph data, and to determine candidate entity pairs with a similarity greater than a preset similarity from the entities, wherein each candidate entity pair consists of entities from different knowledge graph data;

[0135] The inference module 408 is used to take the target data and candidate entity pairs as input and use a generative model to infer the aligned target entity pairs, wherein the target entity pairs include entity pairs among the candidate entity pairs with a confidence level greater than a preset confidence level.

[0136] It should be noted that, Figure 4 The solid alignment device shown is used to perform Figure 2 The method of entity alignment shown, therefore Figure 2 The explanations regarding entity alignment methods in [the document] also apply to [other methods]. Figure 4 The apparatus for aligning entities shown will not be described in detail here.

[0137] This application also provides an electronic device, which includes a memory and a processor, wherein the memory is used to store program instructions; the processor is connected to the memory and is used to execute steps of implementing the entity alignment method in various embodiments of this application.

[0138] This application also provides a non-volatile storage medium including a stored computer program, wherein the device containing the non-volatile storage medium executes the steps of the entity alignment method in various embodiments of this application by running the computer program.

[0139] This application also provides a computer program product, including computer instructions that, when executed by a processor, implement the steps of the entity alignment method in various embodiments of this application.

[0140] This application also provides a computer program that, when executed by a processor, implements the steps of the entity alignment method in various embodiments of this application.

[0141] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0142] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0143] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0144] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0145] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0146] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0147] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method of entity alignment, the method comprising: include: Acquire knowledge graph data corresponding to multiple heterogeneous communication networks, wherein entities in the knowledge graph data are used to represent network devices in the communication networks; The knowledge graph data is converted into target data in a preset format, wherein the preset format includes a code format that can be understood by the generative model; Determine the similarity between entities in multiple knowledge graph data, and determine candidate entity pairs from the entities whose similarity is greater than a preset similarity, wherein each candidate entity pair consists of entities from different knowledge graph data; Using the target data and the candidate entity pairs as input, the generative model is used to infer aligned target entity pairs, wherein the target entity pairs include entity pairs among the candidate entity pairs with a confidence level greater than a preset confidence level.

2. The method of claim 1, wherein, Converting the knowledge graph data into target data in a preset format includes: Network topology information is extracted from the knowledge graph data, wherein the network topology information includes entity information, relationship information describing the connections between network devices, and attribute information describing the characteristics of network devices; The network topology information is reorganized according to the preset format to obtain the target data, wherein the preset format encapsulates the network topology information according to the hierarchical relationship between the entity information, the relationship information and the attribute information.

3. The method of claim 1, wherein, Determining the similarity between entities in multiple knowledge graph datasets includes: A knowledge representation learning framework is used to map multiple knowledge graph data to obtain multiple low-dimensional dense vectors. The knowledge representation learning framework is used to transform entities and topological relationships in the knowledge graph data into mathematical representations, with each entity corresponding to a low-dimensional dense vector. Determine the cosine similarity between the first low-dimensional dense vector and the second low-dimensional dense vector among the plurality of low-dimensional dense vectors, wherein the first entity corresponding to the first low-dimensional dense vector and the second entity corresponding to the second low-dimensional dense vector are entities in different knowledge graph data; The cosine similarity is defined as the similarity between the first entity and the second entity.

4. The method of claim 1, wherein, Using the target data and the candidate entity pairs as input, the generative model is used to infer aligned target entity pairs, including: A query instruction is determined based on the target data and the candidate entity pairs, wherein the query instruction is used to guide the generative model to perform logical verification on the candidate entity pairs based on at least one of the following: entity name similarity, topological structure association, and attribute consistency; For each candidate entity pair, the generative model is used to parse the query command corresponding to the candidate entity pair to obtain a verification result, wherein the verification result is used to determine the matching degree of the candidate entity pair; The target entity pair is determined based on the verification results.

5. The method of claim 4, wherein, Determining the target entity pair based on the verification results includes: For each candidate entity pair, determine the confidence level corresponding to the verification result; If the confidence level is greater than the preset confidence level, the candidate entity pair is determined as the target entity pair.

6. The method of claim 5, wherein, The method further includes: determining the alignment criteria of the target entity pair based on the verification result, wherein the alignment criteria are used to describe the matching criteria and reasoning process of the target entities in the target entity pair.

7. The method according to claim 5, characterized in that, The method further includes: If the confidence level is less than or equal to the preset confidence level, an update query instruction is determined based on the verification result, wherein the update query instruction is used to guide the generative model to expand the search range of the target data; The generative model is used to parse the update query command to obtain the update verification result.

8. The method according to claim 1, characterized in that, The knowledge graph data includes the raw data of the Product Service Resource (PSR) graph, wherein the resource objects in the PSR graph represent network devices of the communication network, the network service objects in the PSR graph represent network services established based on multiple network devices, and the business service objects in the PSR graph represent business services established based on at least one network service.

9. A device for aligning entities, characterized in that, include: An acquisition module is used to acquire knowledge graph data corresponding to multiple heterogeneous communication networks, wherein the entities in the knowledge graph data are used to represent network devices in the communication network; A conversion module is used to convert the knowledge graph data into target data in a preset format, wherein the preset format includes a code format that can be understood by a generative model; The determination module is used to determine the similarity between entities in multiple knowledge graph data, and to determine candidate entity pairs from the entities whose similarity is greater than a preset similarity, wherein each candidate entity pair is composed of entities from different knowledge graph data; The inference module is used to take the target data and the candidate entity pairs as input and use the generative model to infer aligned target entity pairs, wherein the target entity pairs include entity pairs among the candidate entity pairs with a confidence level greater than a preset confidence level.

10. An electronic device, characterized in that, include: A memory and a processor, the memory being used to store program instructions; the processor being connected to the memory and used to execute the method for implementing entity alignment as described in any one of claims 1 to 8.

11. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored computer program, wherein the device containing the non-volatile storage medium executes the entity alignment method according to any one of claims 1 to 8 by running the computer program.

12. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the entity alignment method according to any one of claims 1 to 8.

Images