Engineering entity alignment method based on generative semantic enhancement and structural feature reasoning

Abstract

The invention relates to an engineering entity alignment method based on generative semantic enhancement and structural feature reasoning, and belongs to the technical field of knowledge graph construction and multi-source data fusion. Comprising the following steps: constructing a semantic enhanced entity text sequence; embedding the semantic enhanced entity text sequence into a prompt template containing an alignment task instruction, and inputting a large language model to obtain an entity semantic vector containing engineering context logic; utilizing a graph neural network to capture structural topological features among entities, fusing the structural topological features with the entity semantic vectors, calculating preliminary similarity among the entities, and screening out a candidate alignment set; the entity pairs in the candidate alignment set and neighborhood difference information of the entity pairs are converted into a natural language reasoning task, a big language model is used for generating a discriminant reason, and the final alignment probability is calculated. According to the method, a semantic gap caused by sparse engineering data can be effectively broken through, and the accuracy and robustness of cross-professional engineering entity alignment are remarkably improved.

Description

Technical Field

[0001] This invention relates to an engineering entity alignment method based on generative semantic enhancement and structural feature reasoning, belonging to the field of knowledge graph construction and multi-source data fusion technology. Background Technology

[0002] With the rapid development of smart engineering and digital twin technologies, massive amounts of multi-source heterogeneous data have accumulated throughout the entire lifecycle of engineering projects. This data records the structural forms, design parameters, and operational status of entities, forming the core foundation for building knowledge graphs in the engineering field. However, engineering data exhibits significant cross-disciplinary and cross-stage characteristics. Different disciplines differ greatly in entity naming and description granularity, and the focus of the same entity differs drastically in the design, construction, and operation and maintenance phases. This results in low overlap of attribute fields in cross-source knowledge bases, exhibiting high data sparsity and heterogeneous representation. This makes it difficult to accurately identify the same engineering entity during cross-source alignment, severely restricting the deep integration of engineering knowledge. Existing entity alignment technologies mostly rely on string similarity or shallow semantic features based on representation learning, which have obvious limitations. Traditional string matching struggles to handle engineering abbreviations and aliases; while existing graph neural networks or pre-trained language models, when faced with long-tail entities with severe attribute deficiencies, lack sufficient semantic context support, making it difficult to generate discriminative feature representations. Furthermore, existing models based on vector space distance lack logical reasoning mechanisms similar to those of human experts, often failing to effectively disambiguate complex entities with similar names but different functions, or those with identical topologies but different names. Therefore, there is an urgent need for an engineering entity alignment method based on generative semantic enhancement and structural feature reasoning. This method should intelligently bridge semantic gaps and introduce deep logical discrimination to significantly improve the accuracy and robustness of cross-disciplinary entity alignment in the engineering field. Summary of the Invention

[0003] To address the problems mentioned in the background section, this invention provides an engineering entity alignment method based on generative semantic enhancement and structural feature reasoning, which significantly improves the accuracy and robustness of cross-disciplinary entity alignment in the engineering field.

[0004] The technical solution of this invention is: an engineering entity alignment method based on generative semantic enhancement and structural feature reasoning, comprising:

[0005] Step 1: For sparse attribute entities in the graph to be aligned, use a large language model to infer and generate auxiliary attribute descriptions based on the known context, and construct a semantically enhanced entity text sequence.

[0006] Step 2: Embed the semantically enhanced entity text sequence into a prompt template containing semantic alignment task instructions, and input it into the large language model to obtain entity semantic vectors containing engineering context logic;

[0007] Step 3: Use graph neural networks to capture the structural topological features between entities and fuse them with the entity semantic vectors to calculate the preliminary similarity between entities and filter out candidate alignment sets.

[0008] Step 4: Transform the entity pairs and their neighborhood differences in the candidate alignment set into a natural language reasoning task, use a large language model to generate discriminative reasons and calculate the final alignment probability.

[0009] Further, Step 1 includes:

[0010] Filter sparse entities in the graph to be aligned whose number of attributes is lower than a preset threshold, extract the name of the sparse entity and its first-order neighborhood structure information, and construct a reasoning context that includes engineering professional categories and local topological features.

[0011] The reasoning context is input into the large language model, and supplementary descriptive text implicit in the sparse entity is generated based on engineering common sense. The supplementary descriptive text is then concatenated with the original entity name to form a semantically enhanced entity text sequence.

[0012] Furthermore, Step 2 includes:

[0013] Construct a prompt template that includes engineering expert role settings and semantically aligned task instructions, fill the prompt template with the enhanced entity text sequence, and generate input instructions that include task context;

[0014] The input instructions are passed into the large language model, and the output features of the last hidden state or instruction fine-tuning layer of the model are extracted as entity semantic vectors containing engineering context logic.

[0015] Furthermore, Step 3 includes:

[0016] A graph attention network is used to model the neighborhood nodes of an engineering entity. The structural attention weights are calculated based on the connection relationships between nodes, and the neighbor features are aggregated in a weighted manner to obtain a structural embedding representation.

[0017] The structure embedding representation is concatenated or weighted and fused with the entity semantic vector obtained in Step 2 to construct a unified entity feature vector. The top several entity pairs with the highest similarity are selected by calculating the vector cosine similarity to form a candidate alignment set.

[0018] Furthermore, Step 4 includes:

[0019] From the candidate alignment set, entity pairs with similarity scores within a preset fuzzy range are selected as difficult examples, and the attribute differences and neighborhood structure differences of the paired entities are transformed into natural language descriptions to construct entity alignment inference instructions.

[0020] The entity alignment inference instruction is input into the large language model, and the logical basis and confidence score for judging whether two entities point to the same object are generated through the thinking chain mechanism. The confidence score is then weighted and fused with the preliminary similarity calculated in Step 3 to determine the final alignment result.

[0021] Furthermore, the specific steps of Step 1 include:

[0022] Step 1.1: Define the source project knowledge graph to be aligned. and target engineering knowledge graph Arbitrary graph Represented as ,in Represents a set of entities. Represents a set of relations. Represents a collection of attribute names. Represents a collection of attribute values;

[0023] For any entity in the graph Its original attribute set is represented as Its first-order neighborhood structure set is represented as ;

[0024] Step 1.2: Calculate the attribute density of each entity in the graph. , This indicates that statistics have been compiled. of Quantity, This is equivalent to taking numerical values ​​and setting a sparsity threshold. ,like Then mark It is a sparse entity;

[0025] For sparse entities Constructing contextual prompts for generative reasoning, i.e., reasoning context. :

[0026] (1)

[0027] in, Indicates the engineering specialty category to which the entity belongs. For entity name, This is a text description of the serialized neighbor node information;

[0028] Step 1.3: Construct attribute generation instructions Connect it with the reasoning context Joint input to a large language model (LLM) to generate virtual attribute text sequences. :

[0029] (2)

[0030] in, This indicates a text concatenation operation. Using a generative model with engineering common sense, the generated It contains the implicit technical parameters of the entity;

[0031] Step 1.4: Merge the original entity information with the generated virtual attribute text sequence to construct the final semantically enhanced entity text sequence. :

[0032] (3)

[0033] in, This indicates a flattening operation, which transforms a structured collection of attributes... It is converted into a single-dimensional continuous text sequence to facilitate subsequent text concatenation and model input.

[0034] Furthermore, the specific steps of Step 2 include:

[0035] Step 2.1: Design a prompt template for alignment tasks. It includes character settings, task descriptions, and input slots; for semantically enhanced entity text sequences Construct input instructions :

[0036] (4)

[0037] Step 2.2, Input the encoder of the pre-trained large language model and extract the vectors corresponding to the [CLS] tag or instruction tag from the last hidden state of the model. These vectors serve as the semantic embedding representation of the entity, i.e., the entity semantic vector. :

[0038] (5)

[0039] in, For semantic vector dimension, This represents the encoder of a pre-trained large language model.

[0040] Furthermore, the specific steps of Step 3 include:

[0041] Step 3.1: Use GAT to analyze entities Aggregate the neighborhood of the entity; first calculate the entity with neighboring entities Attention coefficient between :

[0042] (6)

[0043] (7)

[0044] in, It is a learnable linear transformation matrix. For attention vectors, This represents vector concatenation. The first-order neighborhood structure of an entity is represented by a set of its neighborhood structures. For entities semantic vectors, For entities Neighboring entities Entity semantic vector; It is a nonlinear activation function used to introduce nonlinear features and prevent gradient vanishing; Representing the central entity with neighboring entities The original attention scores between, and This generally refers to the strength of the association between the central entity and any neighbor k;

[0045] Based on attention weights, neighbor features are aggregated to obtain the structural embedding representation. :

[0046] (8)

[0047] in, This represents the sigmoid activation function;

[0048] Step 3.2, transfer the entity semantic vectors With structural embedding representation Adaptive weighted fusion is performed to obtain the final entity feature vector. :

[0049] (9)

[0050] in, Indicates weight, For biased weighting;

[0051] Step 3.3: For entities in the source engineering knowledge graph and entities in the target map Calculate the cosine similarity of vectors :

[0052] (10)

[0053] The top K most similar entity pairs are selected to form a candidate alignment set. , The source engineering knowledge graph to be aligned. For the target engineering knowledge graph, For similarity threshold, For entities in the target map The entity feature vector.

[0054] Furthermore, the specific steps of Step 4 include:

[0055] Step 4.1: Define the fuzzy interval If the entity pair cosine similarity If so, mark it as a difficult example to be verified;

[0056] Step 4.2: Extract entity pairs The differences in characteristics, including attribute difference sets Difference set with neighbors Constructing inference prompts The process of constructing this prompt word can be formally represented as follows:

[0057] (11)

[0058] in, This is a text concatenation operation. This represents the task instruction text, used to explicitly require the model to perform the entity alignment inference task; This represents the thought chain guiding text; by integrating task instructions, feature difference data, and thought chain guidance, it is used to enable the model to output alignment results more accurately.

[0059] Step 4.3, Input a large model, make it output the inference steps and the judgment conclusion, and map the conclusion to the confidence probability. ;

[0060] Calculate the final alignment score :

[0061] (12)

[0062] in, For the balance coefficient, if If the value exceeds the set threshold, then the entity is determined to be an entity. and Alignment.

[0063] The beneficial effects of this invention are:

[0064] 1. This invention solves the problem of entity alignment difficulties caused by attribute sparsity and deep logical discontinuity in cross-disciplinary and cross-stage data in the engineering field; it utilizes the generative completion capability of a large language model to construct semantically enhanced sequences, effectively solving the semantic gap problem caused by attribute sparsity and heterogeneous expression in engineering documents.

[0065] 2. This invention integrates the structural features of graph attention networks into the prompt-driven semantic vector to construct multi-view fusion vectors, enabling the model to simultaneously learn the deep functional logic and spatial topological features of engineering entities.

[0066] 3. The multi-view fusion vector obtained in this invention, combined with the thinking chain reasoning mechanism, can perform logical disambiguation on complex and difficult examples, thereby more accurately and robustly determining the alignment relationship of cross-disciplinary engineering entities; this invention can effectively support the automated construction and fusion of cross-stage knowledge bases in the engineering field. Attached Figure Description

[0067] Figure 1 This is a flowchart from the present invention;

[0068] Figure 2 This is a schematic diagram of the alignment model structure corresponding to the method of the present invention. Detailed Implementation

[0069] Example 1: As Figures 1-2 As shown, an engineering entity alignment method based on generative semantic enhancement and structural feature reasoning, targeting existing multi-source engineering knowledge graphs, includes the following steps:

[0070] Step 1: For sparse attribute entities in the graph to be aligned, use a large language model to infer and generate auxiliary attribute descriptions based on the known context, and construct a semantically enhanced entity text sequence.

[0071] This step aims to address the issue of sparse entity attributes in documents such as engineering operation and maintenance or construction logs. By introducing the generative capabilities of large language models, virtual semantic features are generated for long-tail entities without altering the original graph topology.

[0072] Furthermore, for any engineering entity, its corresponding information includes at least the entity name, the engineering discipline to which the entity belongs, the entity's existing attribute information, and the connection relationship with other entities; the relationship between entities reflects the structural dependency or functional connection relationship between equipment, components, or pipelines in the engineering system;

[0073] In actual engineering operation and maintenance logs, construction records, or design documents, some entities have only a very small amount of attribute information, resulting in incomplete semantic expression. To address this issue, this implementation first counts the number of attributes of the entities and then applies a preset sparsity threshold. (For example (i.e., the number of attributes is less than 3) to determine whether an entity belongs to semantically sparse entities.

[0074] When the number of attributes of an entity is lower than the threshold When the entity is identified as a semantically sparse entity, generative semantic enhancement operations are performed on it.

[0075] Specifically, for semantically sparse entities, we first extract their engineering specialty category, entity name, and directly connected first-order neighbor entities (the maximum number of neighbor samples is set to...). Preferred range The information includes neighbor information and relational information, and the neighbor information is converted into a structural description in text form. This information together constitutes the context input for generative reasoning, and its context length is limited to [missing information]. token (preferred range) (token) to adapt to large language model input windows and ensure stable inference.

[0076] Based on this, the context input is combined with preset attribute generation instructions and then input into a large language model with engineering domain knowledge: generating temperature. (Preferred range) This is used to improve the diversity of generated attributes; maximum generated length. token (preferred range) ) is used to constrain the size of attribute text.

[0077] The model, based on engineering common sense and contextual reasoning, generates a description of the entity's implicit but not explicitly recorded technical attributes. The generated attributes may include engineering-related information such as material type, specifications, pressure rating, and installation method (e.g., "Material: Seamless steel pipe", "Pressure rating: 1.6MPa").

[0078] Subsequently, the original attribute information of the entity is merged with the generated virtual attribute text to form a semantically enhanced entity text description, thereby significantly improving the information completeness of semantically sparse entities without changing the original knowledge graph topology.

[0079] Furthermore, the specific steps of Step 1 include:

[0080] Step 1.1: Define the source project knowledge graph to be aligned. and target engineering knowledge graph Arbitrary graph Represented as ,in Represents a set of entities. Represents a set of relations. Represents a collection of attribute names. Represents a collection of attribute values;

[0081] For any entity in the graph Its original attribute set is represented as Its first-order neighborhood structure set is represented as ;

[0082] Step 1.2: Calculate the attribute density of each entity in the graph. , This indicates that statistics have been compiled. of Quantity, This is equivalent to taking numerical values ​​and setting a sparsity threshold. ,like Then mark It is a sparse entity;

[0083] For sparse entities Constructing contextual prompts for generative reasoning, i.e., reasoning context. :

[0084] (1)

[0085] in, This indicates the engineering specialty category to which the entity belongs (e.g., "HVAC Engineering"). For entity name, To serialize the neighbor node information into a text description (e.g., "connected to <cooling tower return water pipe>");

[0086] Step 1.3: Construct attribute generation instructions Connect it with the reasoning context Joint input to a large language model (LLM) to generate virtual attribute text sequences. :

[0087] (2)

[0088] in, This indicates a text concatenation operation. Using a generative model with engineering common sense, the generated It includes the implicit technical parameters of the physical object (such as "material: seamless steel pipe", "pressure rating: 1.6MPa").

[0089] Step 1.4: Merge the original entity information with the generated virtual attribute text sequence to construct the final semantically enhanced entity text sequence. :

[0090] (3)

[0091] in, This indicates a flattening operation, which transforms a structured collection of attributes... Convert it into a single-dimensional continuous text sequence to facilitate subsequent text concatenation and model input;

[0092] Through this step, sparse entities achieve semantic richness equivalent to that of dense entities.

[0093] Step 2: Embed the semantically enhanced entity text sequence into a prompt template containing semantic alignment task instructions, and input it into the large language model to obtain entity semantic vectors containing engineering context logic;

[0094] This step utilizes cue learning techniques to transform the enhanced text sequence into a dense vector containing engineering logic.

[0095] Furthermore, after completing entity semantic enhancement, this implementation method further adopts a prompt-driven approach to convert the enhanced entity text into an engineering semantic vector representation.

[0096] Specifically, a unified prompt template is pre-designed for entity alignment tasks. This template includes role settings, task description, and an entity information input area (the total template length is truncated to a threshold). token (preferred range) By filling the template with semantically enhanced entity text, input instructions that conform to engineering semantic understanding tasks are constructed.

[0097] The constructed input instructions are fed into the pre-trained large language model encoder, and vectors representing entity semantic features are extracted from the model's output as entity semantic embeddings: semantic vector dimension. Depending on the encoder configuration, the preferred option is... or Semantic vectors not only contain the surface textual information of entities, but also integrate the model's deep understanding of engineering terminology, contextual relationships, and implicit logic;

[0098] Furthermore, the specific steps of Step 2 include:

[0099] Step 2.1: Design a prompt template for alignment tasks. It includes character settings, task descriptions, and input slots; for semantically enhanced entity text sequences Construct input instructions :

[0100] (4)

[0101] Step 2.2, Input the encoder of the pre-trained large language model and extract the vectors corresponding to the [CLS] tag or instruction tag from the last hidden state of the model. These vectors serve as the semantic embedding representation of the entity, i.e., the entity semantic vector. :

[0102] (5)

[0103] in, For semantic vector dimension, This vector represents the encoder of a pre-trained large language model, compared to traditional Word2Vec or BERT. It integrates LLM's deep understanding of engineering terminology with contextual reasoning information.

[0104] Step 3: Use graph neural networks to capture the structural topological features between entities and fuse them with the entity semantic vectors to calculate the preliminary similarity between entities and filter out candidate alignment sets.

[0105] Furthermore, in order to further characterize the structural positional features of entities in the engineering system, this embodiment introduces a graph attention mechanism (GAT) to model the neighborhood structure of entities;

[0106] In the specific implementation process, the number of layers in the GAT network is set to... (Preferred range) The number of heads in a multi-head attention mechanism is... (Preferred range) ); Dropout rate set to (Preferred range) To prevent overfitting.

[0107] Using the semantic vector of an entity as the initial feature of a node, and combining the connection relationships between the entity and its neighboring entities, the attention weights between the entity and each neighboring node are calculated. The neighboring features are then weighted and aggregated to obtain the structural feature representation, i.e., the structural embedding representation. .

[0108] Subsequently, the semantic and structural features of the entity are fused through concatenation and linear transformation to obtain the final entity feature vector: the dimension after fusion is set to... (That is, maintaining consistency with the semantic vector dimension for easier similarity calculation). The final entity feature vector is generated. This fused feature simultaneously contains engineering semantic information and structural location information, providing a more comprehensive feature foundation for subsequent entity alignment.

[0109] Recall of candidate entity alignment pairs:

[0110] In one specific implementation, the cosine similarity between the feature vectors of entities in the source engineering knowledge graph and entities in the target engineering knowledge graph is calculated respectively. The cosine similarity is used to measure the degree of consistency between two entities at the level of engineering semantics and structural features.

[0111] Based on the cosine similarity, samples with a cosine similarity higher than the initial screening threshold are selected. Entity pairs, and select the pair with the highest similarity. indivual( Entity pairs constitute a candidate set. This candidate set is used for subsequent refined discrimination processing.

[0112] Furthermore, the specific steps of Step 3 include:

[0113] Step 3.1: Use GAT to analyze entities Aggregate the neighborhood of the entity; first calculate the entity with neighboring entities Attention coefficient between :

[0114] (6)

[0115] (7)

[0116] in, It is a learnable linear transformation matrix. For attention vectors, This represents vector concatenation. The first-order neighborhood structure of an entity is represented by a set of its neighborhood structures. For entities semantic vectors, For entities Neighboring entities Entity semantic vector; It is a nonlinear activation function used to introduce nonlinear features and prevent gradient vanishing; Representing the central entity with neighboring entities The original attention scores between, and This generally refers to the strength of the association between the central entity and any neighbor k;

[0117] Based on attention weights, neighbor features are aggregated to obtain the structural embedding representation. :

[0118] (8)

[0119] in, This represents the sigmoid activation function;

[0120] Step 3.2, transfer the entity semantic vectors With structural embedding representation Adaptive weighted fusion is performed to obtain the final entity feature vector. :

[0121] (9)

[0122] in, Indicates weight, For biased weighting;

[0123] Step 3.3: For entities in the source engineering knowledge graph and entities in the target map Calculate the cosine similarity of vectors :

[0124] (10)

[0125] The top K most similar entity pairs are selected to form a candidate alignment set. , The source engineering knowledge graph to be aligned. For the target engineering knowledge graph, For similarity threshold, For entities in the target map The entity feature vector.

[0126] Step 4: Transform the entity pairs and their neighborhood differences in the candidate alignment set into a natural language reasoning task, use a large language model to generate discriminative reasons and calculate the final alignment probability.

[0127] Furthermore, the specific steps of Step 4 include:

[0128] Step 4.1: Define the fuzzy interval (e.g., [0.65, 0.85]), if entity pairs cosine similarity If so, mark it as a difficult example to be verified;

[0129] Step 4.2: Extract entity pairs The differences in characteristics, including attribute difference sets Difference set with neighbors Constructing inference prompts The prompt contains the following natural language instruction: "Based on attribute differences" Difference set with neighbors Analyze entities in the engineering field With entity Is it the same object? The formal process of constructing this prompt is as follows:

[0130] (11)

[0131] in, This is a text concatenation operation. This represents the task instruction text, used to explicitly require the model to perform the entity alignment inference task; This represents the thought chain guiding text; by integrating task instructions, feature difference data, and thought chain guidance, it is used to enable the model to output alignment results more accurately.

[0132] Step 4.3, Input a large model, make it output the reasoning steps and the judgment conclusion (Yes / No), and map the conclusion to the confidence probability. ;

[0133] Calculate the final alignment score :

[0134] (12)

[0135] in, For the balance coefficient, if If the value exceeds the set threshold, then the entity is determined to be an entity. and Alignment.

[0136] In actual alignment processes, the vector similarity of some entity pairs falls within the intermediate range, making it difficult to directly determine whether they belong to the same engineering entity based solely on vector similarity. To address this issue, this implementation introduces a thought chain reasoning mechanism based on a large language model for further discrimination.

[0137] Specifically, defining fuzzy intervals When the similarity of an entity pair falls within this range, it is marked as a difficult entity pair to be verified. Subsequently, the inference input is submitted to the large language model. To ensure the determinism and stability of the inference, an inference temperature is set. (Preferred range) In the final decision-making stage, a weighted fusion strategy is used to calculate the final score: ,in (Preferred range) When the final score exceeds the judgment threshold. When the entity pair is determined to be the same engineering entity, entity alignment is completed.

[0138] Through the above implementation methods, this invention effectively solves the entity alignment problem caused by sparse entity attributes, incomplete semantic expression, and complex structural relationships in knowledge graphs in the engineering field without relying on manual rules and large-scale labeled data. It significantly improves the accuracy and robustness of alignment and has good engineering application value.

[0139] To verify the effectiveness of the proposed engineering entity alignment method based on generative semantic enhancement and structural feature reasoning, the entity feature vector obtained by fusing semantic features and structural features was used as the model input. Under the same experimental dataset and parameter settings, comparative experiments were conducted on various typical entity alignment methods.

[0140] In this embodiment, the following entity alignment model is selected as the comparison method:

[0141] 1) Traditional entity alignment methods based on string similarity rely solely on entity names or surface character matching information;

[0142] 2) Entity alignment model based solely on structural information, which uses graph attention networks to model the structure of entity neighborhoods but does not introduce generative attribute semantic enhancement mechanisms;

[0143] 3) Entity alignment model based solely on semantic information: Entity alignment is performed through textual semantic representation, but the structural dependencies between engineering entities are not explicitly modeled.

[0144] 4) The entity alignment model (GenStruct-EA model) corresponding to the method proposed in this invention also introduces attribute text enhancement, engineering semantic cue vector extraction and graph attention structural feature fusion mechanism.

[0145] Furthermore, to further verify the role of generative semantic enhancement mechanisms in entity alignment tasks, while keeping the other network structures and training parameter configurations consistent, an entity alignment model without introducing generative semantic enhancement mechanisms and based solely on structural attention networks for entity modeling and matching was constructed and compared with the GenStruct-EA model proposed in this invention.

[0146] The experiment uses Hit@1, Hit@10, and Mean Reciprocal Rank (MRR) as evaluation metrics to measure the accuracy and stability of the entity alignment model in the process of ranking candidate entities.

[0147] The entity alignment model described in this invention is trained without introducing a generative semantic enhancement module, and the corresponding optimal parameter configuration of the model is obtained. Table 1 shows a comparison of experimental results for different entity alignment methods, and Table 2 shows a comparison of experimental results with and without the introduction of a generative semantic enhancement mechanism.

[0148] In this embodiment, to further verify the role of generative semantic enhancement mechanism and structural feature modeling method in entity alignment tasks in the engineering field, while keeping the other network structures, training data and parameter configurations consistent, an entity alignment model (hereinafter referred to as the structural attention model) without generative semantic enhancement module and based solely on structural attention network for entity representation and matching was constructed, and a comparative experiment was conducted with the GenStruct-EA model proposed in this invention.

[0149] Table 1 shows a comparison of the results of different entity alignment methods.

[0150] method Hit@1 Hit@10 MRR Traditional entity alignment methods 0.3166 0.1482 0.2519 Semantic-based entity alignment model 0.3227 0.1470 0.2650 GenStruct-EA 0.3824 0.2270 0.3069

[0151] As can be seen from the experimental results in Table 1, among different entity alignment methods, the GenStruct-EA model proposed in this invention significantly outperforms the comparative methods in terms of Hit@1, Hit@10, and MRR. Specifically, the GenStruct-EA model achieves experimental results of 0.3824, 0.2270, and 0.3069 for Hit@1, Hit@10, and MRR, respectively.

[0152] The results show that by introducing a generative semantic enhancement mechanism into the entity alignment process and combining it with the joint modeling of engineering semantic cue vectors and structural features, the overall performance of cross-disciplinary entity alignment in the engineering field can be effectively improved. The performance improvement mainly comes from the ability to complete the implicit semantic information of engineering entities and the ability to perform deep semantic modeling.

[0153] Table 2 compares the experimental results with and without the introduction of generative semantic enhancement mechanisms.

[0154] method Hit@1 Hit@10 MRR GenStruct-EA 0.3824 0.2270 0.3069 No generative semantic enhancement introduced 0.3618 0.2159 0.3041

[0155] As can be seen from the experimental results in Table 2, the entity alignment model achieves superior performance across all evaluation metrics after introducing the generative semantic enhancement mechanism. Compared with the model that does not introduce the generative semantic enhancement mechanism and only uses a structural attention network for entity modeling, the GenStruct-EA model achieves stable improvements in Hit@1, Hit@10, and MRR metrics.

[0156] The experimental results further demonstrate that by generatively completing the implicit semantic information of engineering entities and jointly modeling the enhanced semantic representation with the structural features of the entity's neighborhood, the model's ability to discriminate the overall semantic consistency of engineering entities is improved. Especially in cross-disciplinary and cross-stage data scenarios in the engineering field, this method can effectively alleviate the difficulties in entity alignment caused by the sparsity of entity attributes and differences in semantic expression, thereby improving the accuracy and robustness of entity alignment results.

[0157] In summary, comparative experimental results demonstrate that the proposed entity alignment method for engineering domains based on generative semantic enhancement and structural feature reasoning achieves superior performance in entity alignment tasks. By utilizing a large language model to generatively complete the implicit semantic information of engineering entities and combining it with a graph attention network to jointly model entity structural features, the proposed method outperforms the comparative models in both entity alignment accuracy and result stability. This method effectively alleviates the difficulties in entity alignment caused by the sparsity of entity attributes and differences in semantic expression in cross-disciplinary and cross-stage data in the engineering field, providing a technical solution with significant practical value for the construction and fusion of knowledge graphs in the engineering field.

[0158] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An engineering entity alignment method based on generative semantic enhancement and structural feature reasoning, characterized in that, include: Step 1: For sparse attribute entities in the graph to be aligned, use a large language model to infer and generate auxiliary attribute descriptions based on the known context, and construct a semantically enhanced entity text sequence. Step 2: Embed the semantically enhanced entity text sequence into a prompt template containing semantic alignment task instructions, and input it into the large language model to obtain entity semantic vectors containing engineering context logic; Step 3: Use graph neural networks to capture the structural topological features between entities and fuse them with the entity semantic vectors to calculate the preliminary similarity between entities and filter out candidate alignment sets. Step 4: Transform the entity pairs and their neighborhood differences in the candidate alignment set into a natural language reasoning task, use a large language model to generate discriminative reasons and calculate the final alignment probability.

2. The engineering entity alignment method based on generative semantic enhancement and structural feature reasoning according to claim 1, characterized in that: Step 1 includes: Filter sparse entities in the graph to be aligned whose number of attributes is lower than a preset threshold, extract the name of the sparse entity and its first-order neighborhood structure information, and construct a reasoning context that includes engineering professional categories and local topological features. The reasoning context is input into the large language model, and supplementary descriptive text implicit in the sparse entity is generated based on engineering common sense. The supplementary descriptive text is then concatenated with the original entity name to form a semantically enhanced entity text sequence.

3. The engineering entity alignment method based on generative semantic enhancement and structural feature reasoning according to claim 1, characterized in that: Step 2 includes: Construct a prompt template that includes engineering expert role settings and semantically aligned task instructions, fill the prompt template with the enhanced entity text sequence, and generate input instructions that include task context; The input instructions are passed into the large language model, and the output features of the last hidden state or instruction fine-tuning layer of the model are extracted as entity semantic vectors containing engineering context logic.

4. The engineering entity alignment method based on generative semantic enhancement and structural feature reasoning according to claim 1, characterized in that: Step 3 includes: A graph attention network is used to model the neighborhood nodes of an engineering entity. The structural attention weights are calculated based on the connection relationships between nodes, and the neighbor features are aggregated in a weighted manner to obtain a structural embedding representation. The structure embedding representation is concatenated or weighted and fused with the entity semantic vector obtained in Step 2 to construct a unified entity feature vector. The top several entity pairs with the highest similarity are selected by calculating the vector cosine similarity to form a candidate alignment set.

5. The engineering entity alignment method based on generative semantic enhancement and structural feature reasoning according to claim 1, characterized in that: Step 4 includes: From the candidate alignment set, entity pairs with similarity scores within a preset fuzzy range are selected as difficult examples, and the attribute differences and neighborhood structure differences of the paired entities are transformed into natural language descriptions to construct entity alignment inference instructions. The entity alignment inference instruction is input into the large language model, and the logical basis and confidence score for judging whether two entities point to the same object are generated through the thinking chain mechanism. The confidence score is then weighted and fused with the preliminary similarity calculated in Step 3 to determine the final alignment result.

6. The engineering entity alignment method based on generative semantic enhancement and structural feature reasoning according to claim 1, characterized in that: The specific steps of Step 1 include: Step 1.1: Define the source project knowledge graph to be aligned. and target engineering knowledge graph Arbitrary graph Represented as ,in Represents a set of entities. Represents a set of relations. Represents a collection of attribute names. Represents a collection of attribute values; For any entity in the graph Its original attribute set is represented as Its first-order neighborhood structure set is represented as ; Step 1.2: Calculate the attribute density of each entity in the graph. , This indicates that statistics have been compiled. of Quantity, This is equivalent to taking numerical values ​​and setting a sparsity threshold. ,like Then mark It is a sparse entity; For sparse entities Constructing contextual prompts for generative reasoning, i.e., reasoning context. : （1）； in, Indicates the engineering specialty category to which the entity belongs. For entity name, This is a text description of the serialized neighbor node information; Step 1.3: Construct attribute generation instructions Connect it with the reasoning context Joint input to a large language model (LLM) to generate virtual attribute text sequences. : （2）； in, This indicates a text concatenation operation. Using a generative model with engineering common sense, the generated It contains the implicit technical parameters of the entity; Step 1.4: Merge the original entity information with the generated virtual attribute text sequence to construct the final semantically enhanced entity text sequence. : （3）； in, This indicates a flattening operation, which transforms a structured collection of attributes... Convert to a single-dimensional continuous text sequence.

7. The engineering entity alignment method based on generative semantic enhancement and structural feature reasoning according to claim 1, characterized in that: The specific steps of Step 2 include: Step 2.1: Design a prompt template for alignment tasks. It includes character settings, task descriptions, and input slots; for semantically enhanced entity text sequences Construct input instructions : （4）； Step 2.2, Input the encoder of the pre-trained large language model and extract the vectors corresponding to the [CLS] tag or instruction tag from the last hidden state of the model. These vectors serve as the semantic embedding representation of the entity, i.e., the entity semantic vector. : （5）； in, For semantic vector dimension, This represents the encoder of a pre-trained large language model.

8. The engineering entity alignment method based on generative semantic enhancement and structural feature reasoning according to claim 1, characterized in that: The specific steps of Step 3 include: Step 3.1: Use GAT to analyze entities Aggregate the neighborhood of the entity; first calculate the entity with neighboring entities Attention coefficient between : （6）； （7）； in, It is a learnable linear transformation matrix. For attention vectors, This represents vector concatenation. It is a set of first-order neighborhood structures representing entities. For entities semantic vectors, For entities Neighboring entities Entity semantic vector; It is a non-linear activation function; Representing the central entity with neighboring entities The original attention scores between, and This generally refers to the strength of the association between the central entity and any neighbor k; Based on attention weights, neighbor features are aggregated to obtain the structural embedding representation. : （8）； in, This represents the sigmoid activation function; Step 3.2, transfer the entity semantic vectors With structural embedding representation Adaptive weighted fusion is performed to obtain the final entity feature vector. : （9）； in, Indicates weight, For biased weighting; Step 3.3: For entities in the source engineering knowledge graph and entities in the target map Calculate the cosine similarity of vectors : （10）； The top K most similar entity pairs are selected to form a candidate alignment set. , The source engineering knowledge graph to be aligned. For the target engineering knowledge graph, For similarity threshold, For entities in the target map The entity feature vector.

9. The engineering entity alignment method based on generative semantic enhancement and structural feature reasoning according to claim 1, characterized in that: The specific steps of Step 4 include: Step 4.1: Define the fuzzy interval If the entity pair cosine similarity If so, mark it as a difficult example to be verified; Step 4.2: Extract entity pairs The differences in characteristics, including attribute difference sets Difference set with neighbors Constructing inference prompts The process of constructing this prompt word can be formally represented as follows: （11）； in, This is a text concatenation operation. This represents the task instruction text; This indicates a thought process that guides the text; Step 4.3, Input a large model, make it output the inference steps and the judgment conclusion, and map the conclusion to the confidence probability. ; Calculate the final alignment score : （12）； in, For the balance coefficient, if If the value exceeds the set threshold, then the entity is determined to be an entity. and Alignment.

Images