A knowledge graph construction method and system for scientific and technological achievements management

Abstract

The present invention relates to the field of data processing technology, and in particular to a knowledge graph construction method and system for scientific and technological achievement management, comprising: obtaining an initial knowledge graph for each category of scientific and technological achievements; obtaining a plurality of candidate entities of the target entity in the reference graph based on the similarity between the word vectors and structural feature vectors of the target entity and all entities in the reference graph; screening a plurality of entities to be aligned from all candidate entities based on the difference between the attribute value sequence of the target entity and the predicted attribute value sequence of all candidate entities; performing entity alignment between the target graph and the reference graph based on the repetition of the target entity and the entities to be aligned of each neighboring node in the reference graph, and obtaining a fused knowledge graph. The present invention improves the accuracy of entity alignment between knowledge graphs constructed from different types of scientific and technological achievement data.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for constructing a knowledge graph for scientific and technological achievement management. Background Art

[0002] As scientific research expands and the number of scientific and technological achievements increases, the importance and necessity of scientific and technological achievement management are becoming increasingly prominent. Scientific management strategies must be established to ensure that they align with the content of these achievements and maximize their effectiveness. Due to the diverse nature of scientific and technological achievements, including papers, software, products, and design drawings, these achievements contain large amounts of heterogeneous data, making it difficult for conventional data management systems to effectively manage them in a timely manner. On the other hand, with the interdisciplinary development of science and technology, effective retrieval and correlation analysis of scientific and technological achievements across different fields can promote the application of these achievements and provide a rich visual display of the vast amount of scientific and technological achievement information.

[0003] Existing Problems: Knowledge graphs possess powerful association analysis and retrieval capabilities. Their role in scientific and technological achievement management lies primarily in their ability to automatically extract knowledge from various types of scientific and technological achievements and to explore the inherent connections between them. However, due to the varying evaluation criteria across disciplines and types of scientific and technological achievements, even for related scientific and technological achievements, significant discrepancies can occur between the extracted triplets when constructing knowledge graphs. This results in poor entity alignment when constructing knowledge graphs based on data from different types of scientific and technological achievements, making it difficult to effectively manage scientific and technological achievements using these constructed knowledge graphs. Summary of the Invention

[0004] The present invention provides a knowledge graph construction method and system for scientific and technological achievements management to solve existing problems.

[0005] The present invention provides a knowledge graph construction method and system for scientific and technological achievements management using the following technical solutions:

[0006] An embodiment of the present invention provides a method for constructing a knowledge graph for scientific and technological achievement management, the method comprising the following steps:

[0007] Obtaining an initial knowledge graph of each category of scientific and technological achievements; the initial knowledge graph includes: attribute values ​​of each attribute of each entity;

[0008] Denote any two initial knowledge graphs as the target graph and the reference graph respectively; denote any entity in the target graph as the target entity; obtain the word vector and structural feature vector of each entity, and obtain several candidate entities of the target entity in the reference graph based on the similarity between the word vector and structural feature vector of the target entity and all entities in the reference graph;

[0009] The attribute values ​​of all attributes of the target entity in the target graph are used to form an attribute value sequence; a multi-label classification model of the target graph is constructed, each candidate entity is input into the multi-label classification model of the target graph, and a predicted attribute value sequence is output; based on the difference between the attribute value sequence of the target entity and the predicted attribute value sequences of all candidate entities, a number of entities to be aligned are selected from all candidate entities;

[0010] Based on the repetition of the target entity and the entities to be aligned in each neighbor node in the reference graph, entity alignment is performed between the target graph and the reference graph to obtain a fused knowledge graph; the neighbor nodes are other entities in the target graph that are directly connected to the target entity.

[0011] Furthermore, the step of obtaining several candidate entities of the target entity in the reference graph includes the following specific steps:

[0012] Obtain the similarity between the word vectors of the target entity and each entity in the reference graph, recorded as the first similarity, obtain the similarity between the structural feature vectors of the target entity and each entity in the reference graph, recorded as the second similarity, and determine the matching degree between the target entity and each entity in the reference graph based on the first similarity and the second similarity;

[0013] According to the matching degree between the target entity and each entity in the reference graph, several candidate entities of the target entity in the reference graph are obtained.

[0014] Furthermore, the step of determining the matching degree between the target entity and each entity in the reference graph based on the first similarity and the second similarity includes the following specific steps:

[0015] Obtain the product of the first similarity and the preset first weight, recorded as the first product, obtain the product of the second similarity and the preset second weight, recorded as the second product, and use the sum of the first product and the second product as the matching degree between the target entity and each entity in the reference graph.

[0016] Furthermore, the step of obtaining several candidate entities of the target entity in the reference graph according to the matching degree between the target entity and each entity in the reference graph includes the following specific steps:

[0017] Among the matching degrees between the target entity and all entities in the reference graph, a binary classification method is used to classify all matching degrees and obtain two matching degree categories. The matching degree category with the largest matching degree mean corresponds to the entity in the reference graph and is recorded as a candidate entity.

[0018] Furthermore, the specific steps of selecting a number of entities to be aligned from all candidate entities are as follows:

[0019] In the target graph, obtain the normalized value of the information entropy of each attribute on all entities as the weight of each attribute;

[0020] Determining a difference evaluation parameter between the attribute value distributions of the target entity and each candidate entity based on the difference between the attribute value sequence of the target entity and the predicted attribute value sequence of each candidate entity and in combination with the weight of each attribute;

[0021] According to the size of the difference evaluation parameter between the attribute value distribution of the target entity and each candidate entity, several entities to be aligned are screened out from all candidate entities.

[0022] Furthermore, the step of determining the difference evaluation parameter between the attribute value distributions of the target entity and each candidate entity based on the difference between the attribute value sequence of the target entity and the predicted attribute value sequence of each candidate entity, in combination with the weight of each attribute, includes the following specific steps:

[0023] In the predicted attribute value sequence of each candidate entity and the attribute value sequence of the target entity, obtain the absolute value of the difference between the predicted attribute value and the attribute value corresponding to each attribute, which is recorded as the attribute difference value of each attribute, obtain the product of the attribute difference value of each attribute and the weight of each attribute, which is recorded as the weighted attribute difference value of each attribute, obtain the sum of the weighted attribute difference values ​​of all attributes, which is recorded as the difference evaluation parameter between the attribute value distribution of the target entity and each candidate entity.

[0024] Furthermore, the step of evaluating the size of the parameter based on the difference between the attribute value distribution of the target entity and each candidate entity and selecting a number of entities to be aligned from all candidate entities includes the following specific steps:

[0025] A quantity threshold Q is preset, and among the difference evaluation parameters between the target entity and all candidate entities in the attribute value distribution, the candidate entities corresponding to the smallest first Q difference evaluation parameters are recorded as entities to be aligned.

[0026] Furthermore, the entity alignment between the target graph and the reference graph is performed based on the repetition of the target entity and the entity to be aligned of each neighboring node in the reference graph to obtain a fused knowledge graph, which includes the following specific steps:

[0027] According to the repetition of the target entity and each neighboring node in the reference graph, the attention weight between the target entity and each neighboring node of the target entity in the target graph is obtained when the target graph and the reference graph are aligned;

[0028] Taking the target map and the reference map as input, GNN is used to align the entities between the target map and the reference map. When aligning the target map and the reference map, the attention weight between each entity in the target map and each neighbor node of each entity is used to output the entity alignment result between the target map and the reference map.

[0029] According to the entity alignment results between the target graph and the reference graph, a cross-graph equivalence chain is established between the target graph and the reference graph for the aligned entities to obtain a fused knowledge graph of the target graph and the reference graph.

[0030] Furthermore, the step of obtaining the attention weight between the target entity and each neighboring node of the target entity in the target graph when aligning the target graph with the reference graph based on the repetition of the target entity and each neighboring node in the entity to be aligned in the reference graph includes the following specific steps:

[0031] The set of entities to be aligned is formed by taking all the entities to be aligned in the reference graph for each entity in the target graph;

[0032] The normalized value of the Jaccard similarity coefficient between the target entity and the set of entities to be aligned of each neighboring node of the target entity is obtained as the attention weight between the target entity and each neighboring node of the target entity in the target graph when aligning the target graph with the reference graph.

[0033] The present invention also proposes a knowledge graph construction system for scientific and technological achievement management, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. The processor executes the computer program stored in the memory to implement the steps of the aforementioned knowledge graph construction method for scientific and technological achievement management.

[0034] The beneficial effects of the technical solution of the present invention are:

[0035] In an embodiment of the present invention, an initial knowledge graph of each category of scientific and technological achievements is obtained, and several candidate entities of the target entity in the reference graph are obtained based on the similarity of the word vectors and structural feature vectors of the target entity and all entities in the reference graph. In particular, when obtaining the structural feature vector, a graph convolution model is used to better utilize the data representation of the graph structure type of the knowledge graph itself, which can be integrated with the word vector, attribute features, etc. of the entity to form a multi-dimensional feature table of the entity. When screening the candidate entities, the weighted weights of the term similarity and structural similarity are adaptively adjusted according to the actual situation of the entity, so that the screening of candidate entities can adapt to different types of entities. According to the difference between the attribute value sequence of the target entity and the predicted attribute value sequence of all candidate entities, several entities to be aligned are screened from all candidate entities. Therefore, in view of the fact that the attribute value distribution of the alignable entities is subject to cognitive constraints, the mapping relationship between entities and attribute values ​​is learned from the attribute value sequence of the entity using a multi-label learning method. The obtained multi-label classification model can adapt to the attribute differences between entities, ensuring the prediction performance of the multi-label classification model, thereby ensuring the accuracy of the selection of entities to be aligned. Based on the duplication of the target entity and each neighboring node in the reference graph, entity alignment is performed between the target graph and the reference graph to obtain a fused knowledge graph. Thus, the present invention gradually narrows the scope of entity alignment between the initial knowledge graphs through the entity's term similarity, structural similarity, and attribute value distribution constraint features, and optimizes the entity alignment model, thereby improving the accuracy of entity alignment between knowledge graphs constructed from different types of scientific and technological achievement data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 This is a flowchart of the steps of a method for constructing a knowledge graph for scientific and technological achievements management according to the present invention;

[0038] Figure 2 Schematic diagram of the same attributes of different entities in different fields of basic science. DETAILED DESCRIPTION

[0039] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail a method and system for constructing a knowledge graph for scientific and technological achievement management proposed by the present invention, its specific implementation method, structure, features and effects. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0040] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0041] The following describes in detail a method and system for constructing a knowledge graph for scientific and technological achievements management provided by the present invention in conjunction with the accompanying drawings.

[0042] See also Figure 1 , which shows a flowchart of a method for constructing a knowledge graph for scientific and technological achievements management provided by one embodiment of the present invention, the method comprising the following steps:

[0043] Step S001: Obtain an initial knowledge graph of each category of scientific and technological achievements; the initial knowledge graph includes: attribute values ​​of each attribute of each entity.

[0044] In this example, we first construct an initial knowledge graph for scientific and technological achievements within each category. We then train a multi-label classifier using the attribute distribution of entities within different categories to determine the scope for entity alignment between different knowledge graphs. Finally, we perform entity alignment between the initial knowledge graphs based on the entities to be aligned after secondary screening to construct the final knowledge graph.

[0045] It should be noted that after constructing the initial knowledge graph for each category of scientific and technological achievements, the initial knowledge graphs constructed under different categories are matched by analyzing the attributes and relationships of the entity nodes in the knowledge graph, thereby determining whether there are related objects in the two scientific and technological achievements, and realizing the effective combined management of scientific and technological achievements of different categories. In this embodiment, the categories are divided according to the fields to which the scientific and technological achievements belong, and scientific and technological achievements belonging to the same field are divided into the same category. The fields include but are not limited to energy, aerospace, biology, and astronomy. In other embodiments, they can also be divided into basic theoretical achievements, applied technological achievements, and soft science achievements according to the nature of the research. This is not particularly limited in this embodiment. For each category of scientific and technological achievements, named entity recognition and relationship extraction techniques are used to identify entities from the data of each category of scientific and technological achievements, extract the relationships between entities, identify entity attributes and attribute values, and use the knowledge graph constructed by triples composed of all identified entities as the initial knowledge graph for each category. Among them, the data of the scientific and technological achievements include but are not limited to research reports, academic papers, and technical reports. The construction of the knowledge graph is a well-known technology, and the specific process is not repeated here.

[0046] Thus, an initial knowledge graph of each category of scientific and technological achievements is obtained, wherein the initial knowledge graph includes: attribute values ​​of each attribute of each entity.

[0047] It should be noted that the attribute values ​​of entity attributes in the initial knowledge graph are converted to numerical data through label encoding. Furthermore, the attribute values ​​of the same attribute in the initial knowledge graph are normalized using the max-min normalization method to unify the dimensions. Both label encoding and the max-min normalization method are well-known techniques, and their specific methods will not be described here.

[0048] Step S002: record any two initial knowledge graphs as the target graph and the reference graph respectively; record any entity in the target graph as the target entity; obtain the word vector and structural feature vector of each entity, and obtain several candidate entities of the target entity in the reference graph based on the similarity between the word vector and structural feature vector of the target entity and all entities in the reference graph.

[0049] It should be noted that: since the data of scientific and technological achievements are usually highly professional, the relationships between entity nodes in the initial knowledge graphs of different types will be quite different, and the descriptions of the same entity in data from different fields will also be different. Directly aligning entities between the initial knowledge graphs will cause large errors. In this embodiment, the initial matching between entities is first performed using the term similarity and structural similarity of entities in knowledge graphs of different types, and the matching results of semantic similarity are corrected using multi-label learning. That is, the term similarity and structural similarity between entities are first evaluated using scientific and technological achievement data and the initial knowledge graph.

[0050] Preferably, in one embodiment of the present invention, the method for obtaining candidate entities includes:

[0051] For any entity in the initial knowledge graph of any category of scientific and technological achievements, the text data containing the arbitrary entity is filtered from the data of the arbitrary category of scientific and technological achievements, and then the word vector of the arbitrary entity is output using the text processing model.

[0052] It should be noted that text processing models include but are not limited to Word2Vec, BERT, and the domain pre-training model (Bio-BERT). The acquisition of word vectors is a well-known technology in the field of natural language processing, and the specific process will not be repeated here.

[0053] The initial knowledge graph of any category of scientific and technological achievements is taken as input, and a deep learning model based on graph convolution is used to obtain the structural feature vector of each entity in the initial knowledge graph of any category of scientific and technological achievements.

[0054] It should be noted that graph convolution-based deep learning models include but are not limited to R-GCN (Relational Graph Convolutional Network), GNN (Graph Neural Network), and GAT (Graph Attention Network). This is a well-known technology, and the specific process will not be elaborated on here. Here, graph convolution-based models are used because the essence of extracting structural features of entity nodes from knowledge graphs is to aggregate neighbor information through a message passing mechanism, encoding the topological relationships of the graph structure into low-dimensional vectors. Models using graph convolution can better utilize the graph structure data representation of the knowledge graph itself, and can be integrated with the entity's word vectors, attribute features, etc. to form a multi-dimensional feature representation of the entity. The entity's word vectors and structural feature vectors are then used to align entities between different initial knowledge graphs. Specifically, the degree of matching between entities is measured based on the similarity between the word vectors and structural feature vectors of different entities.

[0055] The initial knowledge graphs of any two categories of scientific and technological achievements are recorded as target graph and reference graph respectively.

[0056] Any entity in the target graph is recorded as the target entity.

[0057] Preferably, a possible implementation method for the matching degree between the target entity and each entity in the reference graph is as follows:

[0058] Obtain the similarity between the word vectors of the target entity and each entity in the reference graph, recorded as the first similarity, obtain the similarity between the structural feature vectors of the target entity and each entity in the reference graph, recorded as the second similarity, obtain the product of the first similarity and the preset first weight, recorded as the first product, obtain the product of the second similarity and the preset second weight, recorded as the second product, and take the sum of the first product and the second product as the matching degree between the target entity and each entity in the reference graph.

[0059] Alternatively, another possible implementation of the matching degree between the target entity and each entity in the reference graph is as follows:

[0060] The word vector of each entity is concatenated with the structural feature vector to obtain a concatenated vector. The similarity between the concatenated vectors of the target entity and each entity in the reference graph is obtained as the matching degree of the target entity and each entity in the reference graph.

[0061] It should be noted that: in this embodiment, the calculation method for obtaining the similarity between vectors includes cosine similarity, Jaccard coefficient, etc. This is a well-known technology and the specific process will not be repeated. The word vector and structural feature vector of each entity are spliced ​​head to tail, and the splicing order of the word vector and the structural feature vector is not restricted. The first weight is preset to 0.5 and the second weight is preset to 0.5. This is described as an example, and the sum of the preset first weight and the preset second weight is 1. The implementer can adjust the weighted weights of the first similarity and the second similarity according to the specific situation of the entity in the scientific and technological achievement data. The adjustment method is: the more professional the entity, the more stable the word vector corresponding to the entity, and the weighted weight of the similarity between the word vectors should be greater than the weighted weight of the similarity between the structural feature vectors. On the contrary, if the professionalism of the entity is weaker, it is easy to form different word vectors in combination with the context, and the word vector has weaker discrimination for the entity. In this case, the weighted weight of the similarity between the word vectors should be less than the weighted weight of the similarity between the structural feature vectors.

[0062] Among the matching degrees between the target entity and all entities in the reference graph, a binary classification method is used to classify all matching degrees and obtain two matching degree categories. The matching degree category with the largest matching degree mean corresponds to the entity in the reference graph and is recorded as a candidate entity.

[0063] It should be noted that binary classification methods include threshold segmentation and data clustering, such as the K-means clustering algorithm, where the K value is set to 2. This is a well-known technique, and the specific process will not be described in detail. The candidate entities selected have a high degree of terminology and structural similarity with the target entity.

[0064] Step S003: Use the attribute values ​​of all attributes of the target entity in the target graph to form an attribute value sequence; build a multi-label classification model of the target graph, input each candidate entity into the multi-label classification model of the target graph, and output a predicted attribute value sequence; based on the difference between the attribute value sequence of the target entity and the predicted attribute value sequences of all candidate entities, select a number of entities to be aligned from all candidate entities.

[0065] What needs to be explained is that in the initial knowledge graph, the attribute value of the entity is usually affected by three directions: one is the shared cognitive knowledge of basic science, the second is the domain characteristics, and the third is the data error of the entity data. If the entity corresponds to a professional term, the professional term shows the characteristics of similar attribute value distribution in the knowledge graph of different fields. This is because the top-level entities share the basic cognitive framework in different fields, and the attribute value distribution of the corresponding entity is constrained due to the semantic inheritance characteristics of the cross-domain of different fields. Schematic diagram of the same attribute representation of different entities in different fields of basic science, such as Figure 2 shown. Figure 2 The same attribute (vibration frequency) of industrial equipment (entity 1) in the engineering field under basic science and medical equipment (entity 2) in the biomedical field has a distribution range of attribute values ​​(from the minimum vibration frequency to the maximum vibration frequency). Vibration frequency, as an attribute of an entity, is constrained by the cognitive knowledge of vibration frequency in basic science, whether it appears in scientific and technological achievement data in the engineering field or in the biomedical field. This means that the attribute value of the vibration frequency attribute of the entity identified in industrial equipment or medical equipment always falls within the specified frequency range. Therefore, based on the above analysis, this embodiment uses the mapping relationship between entities and attribute values ​​to perform a secondary screening of the selected candidate entities.

[0066] Preferably, in one embodiment of the present invention, the method for obtaining the entity to be aligned includes:

[0067] In the target graph, obtain the attribute values ​​of all attributes of the target entity to form an attribute value sequence.

[0068] Here, all attributes are those present in the target graph. Attributes not present in the target entity are set to Null. This is because even for the same entity, the attribute information extracted from different categories of scientific and technological achievement datasets may differ. Setting the number of labels based solely on the number of attributes of a single entity will not accommodate attribute differences between entities, impacting the performance of subsequent multi-label classification models.

[0069] In this embodiment, the multi-label classification task is to learn the mapping relationship between entities and attributes in the initial knowledge graph of each category, and obtain a multi-label classification model corresponding to the initial knowledge graph, that is, to obtain a multi-label classification model of the target graph. Multi-label learning is a well-known technology in the field of machine learning. Commonly used multi-label classification models include Rank-SVM, CML (Conditional Multilabel Classification), convolutional neural network CNN, etc. The specific training process will not be repeated here.

[0070] It should be noted that if an entity in another category can be aligned with an entity node in another initial knowledge graph, then the attribute value sequences of the two entities should have a high degree of similarity. On the other hand, because an entity has multiple attributes and different attributes have different degrees of relevance to the entity, the amount of information carried by the attribute information also varies. Therefore, the similarity is weighted according to the information entropy of the attribute information.

[0071] In the target graph, the normalized value of the information entropy of each attribute on all entities is obtained as the weight of each attribute.

[0072] It should be noted that in this embodiment, the calculation of information entropy is a well-known calculation. In the target graph, the information entropy of the attribute values ​​of each attribute on all entities is obtained as the entropy value of each attribute. The sum of the entropy values ​​of all attributes is obtained, and the ratio of the entropy value of each attribute to the sum of the entropy values ​​is used as the normalized value of the information entropy of the attribute values ​​of each attribute on all entities. Among them, the greater the information entropy, the greater the difference in attribute values ​​of the attribute on different entities, the stronger the correlation between the attributes in the entities of this type of scientific and technological achievements, and the greater the weight should be given.

[0073] Each candidate entity is input into the multi-label classification model of the target graph and a sequence of predicted attribute values ​​is output.

[0074] It should be noted that all attributes in the predicted attribute value sequence are all attributes existing in the target graph, and the attributes corresponding to the same ordinal values ​​in the predicted attribute value sequence and the attribute value sequence of the target entity are the same.

[0075] In the predicted attribute value sequence of each candidate entity and the attribute value sequence of the target entity, obtain the absolute value of the difference between the predicted attribute value and the attribute value corresponding to each attribute, which is recorded as the attribute difference value of each attribute, obtain the product of the attribute difference value of each attribute and the weight of each attribute, which is recorded as the weighted attribute difference value of each attribute, obtain the sum of the weighted attribute difference values ​​of all attributes, which is recorded as the difference evaluation parameter between the attribute value distribution of the target entity and each candidate entity.

[0076] What needs to be explained is that: the smaller the difference evaluation parameter is, the more similar the attribute value distribution of the candidate entity and the target entity is when they are in the initial knowledge graph composed of the same type of scientific and technological achievement data, and the stronger the correlation between the candidate entity and the target entity.

[0077] The preset quantity threshold Q is 20, which is used as an example for description.

[0078] Among the difference evaluation parameters between the target entity and all candidate entities in the attribute value distribution, the candidate entities corresponding to the smallest top Q difference evaluation parameters are recorded as entities to be aligned.

[0079] According to the above method, a number of entities to be aligned in the reference graph can be obtained for each entity in the target graph.

[0080] Step S004: Perform entity alignment between the target graph and the reference graph based on the repetition of the target entity and each neighbor node in the reference graph to obtain a fused knowledge graph; the neighbor nodes are other entities in the target graph that are directly connected to the target entity.

[0081] It should be noted that for the target graph, the entities to be aligned belonging to the target entity and all neighboring nodes of the target entity in the reference graph are counted separately. If the degree of overlap between each neighboring entity node of the target entity and the set of entities to be aligned with the target entity in the reference graph is high, each neighboring entity node should be given more attention when using GNN to extract the features of the target entity, and the node information between each neighboring entity node should be increased when constructing the adjacency matrix.

[0082] Preferably, in one embodiment of the present invention, the method for obtaining the fused knowledge graph includes:

[0083] The set of entities to be aligned is formed by all the entities to be aligned in the reference graph for each entity in the target graph.

[0084] Get several neighbor nodes of the target entity. Neighbor nodes are other entities directly connected to the target entity in the target graph.

[0085] The normalized value of the Jaccard similarity coefficient between the target entity and the set of entities to be aligned of each neighboring node of the target entity is obtained as the attention weight between the target entity and each neighboring node of the target entity in the target graph when aligning the target graph with the reference graph.

[0086] It should be noted that the Jaccard similarity coefficient is a well-known technique that measures the similarity between two sets by calculating the ratio of the intersection size to the union size. In this embodiment, the sum of the Jaccard similarity coefficients between the target entity and the set of entities to be aligned for all of the target entity's neighboring nodes is obtained. The ratio of the Jaccard similarity coefficient between the target entity and the set of entities to be aligned for each of the target entity's neighboring nodes to the sum of the Jaccard similarity coefficients is used as the normalized value of the Jaccard similarity coefficient between the target entity and the set of entities to be aligned for each of the target entity's neighboring nodes.

[0087] Taking the target map and the reference map as input, GNN is used to perform entity alignment between the target map and the reference map. When aligning the target map and the reference map, the attention weight between each entity in the target map and each neighbor node of each entity is used to output the entity alignment result between the target map and the reference map.

[0088] Among them, GNN entity alignment between knowledge graphs is a commonly used technology for knowledge graphs, and the specific process will not be repeated here.

[0089] Based on the entity alignment results between the target graph and the reference graph, a cross-graph equivalence chain is established between the target graph and the reference graph for the aligned entities to obtain a fused knowledge graph of the target graph and the reference graph. This operation is a common step in the field of knowledge graph fusion, and the specific process will not be repeated here.

[0090] What needs to be explained is: when it is necessary to search and analyze scientific and technological achievements in two fields and conduct cross-integration between scientific and technological achievements, the effective management of scientific and technological achievements can be achieved by integrating the initial knowledge graphs in the two fields, screening common entities, and cross-integrating the scientific and technological achievements containing common entities.

[0091] The present invention also provides a knowledge graph construction system for scientific and technological achievement management, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. The processor executes the computer program stored in the memory to implement the steps of the aforementioned knowledge graph construction method for scientific and technological achievement management.

[0092] So far, the present invention is completed.

[0093] In summary, in an embodiment of the present invention, an initial knowledge graph of each category of scientific and technological achievements is obtained, any two initial knowledge graphs are respectively recorded as a target graph and a reference graph, any entity in the target graph is recorded as a target entity, and according to the similarity of the word vector and structural feature vector of the target entity and all entities in the reference graph, several candidate entities of the target entity in the reference graph are obtained, and the attribute values ​​of all attributes of the target entity in the target graph are used to form an attribute value sequence, each candidate entity is input into the multi-label classification model of the target graph, and a predicted attribute value sequence is output. According to the difference between the attribute value sequence of the target entity and the predicted attribute value sequence of all candidate entities, several entities to be aligned are screened from all candidate entities, and according to the repetition of the target entity and the entity to be aligned in the reference graph of each neighbor node, entity alignment is performed between the target graph and the reference graph to obtain a fused knowledge graph. The present invention improves the accuracy of entity alignment between knowledge graphs constructed from different types of scientific and technological achievement data.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A knowledge graph construction method for scientific and technological achievements management, characterized in that: The method comprises the following steps: Obtaining an initial knowledge graph of each category of scientific and technological achievements; the initial knowledge graph includes: attribute values ​​of each attribute of each entity; Denote any two initial knowledge graphs as the target graph and the reference graph respectively; denote any entity in the target graph as the target entity; obtain the word vector and structural feature vector of each entity, and obtain several candidate entities of the target entity in the reference graph based on the similarity between the word vector and structural feature vector of the target entity and all entities in the reference graph; The attribute values ​​of all attributes of the target entity in the target graph are used to form an attribute value sequence; a multi-label classification model of the target graph is constructed, each candidate entity is input into the multi-label classification model of the target graph, and a predicted attribute value sequence is output; based on the difference between the attribute value sequence of the target entity and the predicted attribute value sequences of all candidate entities, a number of entities to be aligned are selected from all candidate entities; Perform entity alignment between the target graph and the reference graph based on the repetition of the target entity and the entities to be aligned in each neighboring node in the reference graph to obtain a fused knowledge graph; the neighboring nodes are other entities in the target graph that are directly connected to the target entity; The entity alignment between the target graph and the reference graph is performed based on the repetition of the target entity and the entity to be aligned of each neighboring node in the reference graph to obtain a fused knowledge graph, which includes the following specific steps: According to the repetition of the target entity and each neighboring node in the reference graph, the attention weight between the target entity and each neighboring node of the target entity in the target graph is obtained when the target graph and the reference graph are aligned; Taking the target map and the reference map as input, GNN is used to align the entities between the target map and the reference map. When aligning the target map and the reference map, the attention weight between each entity in the target map and each neighbor node of each entity is used to output the entity alignment result between the target map and the reference map. According to the entity alignment results between the target graph and the reference graph, a cross-graph equivalence chain is established between the target graph and the reference graph for the aligned entities to obtain a fused knowledge graph of the target graph and the reference graph.

2. A knowledge graph construction method for scientific and technological achievements management according to claim 1, characterized in that: The specific steps of obtaining several candidate entities of the target entity in the reference graph are as follows: Obtain the similarity between the word vectors of the target entity and each entity in the reference graph, recorded as the first similarity, obtain the similarity between the structural feature vectors of the target entity and each entity in the reference graph, recorded as the second similarity, and determine the matching degree between the target entity and each entity in the reference graph based on the first similarity and the second similarity; According to the matching degree between the target entity and each entity in the reference graph, several candidate entities of the target entity in the reference graph are obtained.

3. The method for constructing a knowledge graph for scientific and technological achievements management according to claim 2, characterized in that: The step of determining the matching degree between the target entity and each entity in the reference graph based on the first similarity and the second similarity includes the following specific steps: Obtain the product of the first similarity and the preset first weight, recorded as the first product, obtain the product of the second similarity and the preset second weight, recorded as the second product, and use the sum of the first product and the second product as the matching degree between the target entity and each entity in the reference graph.

4. The method for constructing a knowledge graph for scientific and technological achievements management according to claim 2, characterized in that: The specific steps of obtaining a plurality of candidate entities of the target entity in the reference graph according to the matching degree between the target entity and each entity in the reference graph are as follows: Among the matching degrees between the target entity and all entities in the reference graph, a binary classification method is used to classify all matching degrees and obtain two matching degree categories. The matching degree category with the largest matching degree mean corresponds to the entity in the reference graph and is recorded as a candidate entity.

5. The method for constructing a knowledge graph for scientific and technological achievements management according to claim 1, characterized in that: The specific steps of selecting a number of entities to be aligned from all candidate entities are as follows: In the target graph, obtain the normalized value of the information entropy of each attribute on all entities as the weight of each attribute; Determining a difference evaluation parameter between the attribute value distributions of the target entity and each candidate entity based on the difference between the attribute value sequence of the target entity and the predicted attribute value sequence of each candidate entity and in combination with the weight of each attribute; According to the size of the difference evaluation parameter between the attribute value distribution of the target entity and each candidate entity, several entities to be aligned are screened out from all candidate entities.

6. A knowledge graph construction method for scientific and technological achievements management according to claim 5, characterized in that: The step of determining the difference evaluation parameter between the attribute value distribution of the target entity and each candidate entity based on the difference between the attribute value sequence of the target entity and the predicted attribute value sequence of each candidate entity and combining the weight of each attribute includes the following specific steps: In the predicted attribute value sequence of each candidate entity and the attribute value sequence of the target entity, obtain the absolute value of the difference between the predicted attribute value and the attribute value corresponding to each attribute, which is recorded as the attribute difference value of each attribute, obtain the product of the attribute difference value of each attribute and the weight of each attribute, which is recorded as the weighted attribute difference value of each attribute, obtain the sum of the weighted attribute difference values ​​of all attributes, which is recorded as the difference evaluation parameter between the attribute value distribution of the target entity and each candidate entity.

7. The method for constructing a knowledge graph for scientific and technological achievements management according to claim 5, characterized in that: The specific steps of evaluating the size of the parameter based on the difference between the attribute value distribution of the target entity and each candidate entity and selecting a number of entities to be aligned from all candidate entities are as follows: A quantity threshold Q is preset, and among the difference evaluation parameters between the target entity and all candidate entities in the attribute value distribution, the candidate entities corresponding to the smallest first Q difference evaluation parameters are recorded as entities to be aligned.

8. The method for constructing a knowledge graph for scientific and technological achievements management according to claim 1, characterized in that: The method of obtaining the attention weight between the target entity and each neighboring node of the target entity in the target graph when aligning the target graph with the reference graph based on the repetition of the target entity and each neighboring node in the reference graph includes the following specific steps: The set of entities to be aligned is formed by taking all the entities to be aligned in the reference graph for each entity in the target graph; The normalized value of the Jaccard similarity coefficient between the target entity and the set of entities to be aligned of each neighboring node of the target entity is obtained as the attention weight between the target entity and each neighboring node of the target entity in the target graph when aligning the target graph with the reference graph.

9. A knowledge graph construction system for scientific and technological achievements management, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is executed by a processor, the steps of a method for constructing a knowledge graph for scientific and technological achievement management as described in any one of claims 1 to 8 are implemented.

Images