A cross-language entity alignment method based on meta attribute reduction and attribute embedding

Abstract

The application discloses a cross-language entity alignment method based on meta attribute reduction and attribute embedding and belongs to the technical field of knowledge fusion and artificial intelligence. The method comprises the following steps: firstly, attribute embedding and relation embedding of entities in knowledge graphs of two different natural languages are respectively generated, entity embedding is constructed, and the similarity between the two knowledge graphs is calculated; then, the similarity between the two knowledge graphs is calculated by using a method based on multi-dimensional information measurement; finally, the similarity is fused, and an entity alignment result is output. The cross-language entity alignment method of the application carries out attribute embedding training through construction of attribute pairs with different correlation strengths, captures the semantic relationship between attributes, realizes the fusion of attribute embedding and relation embedding of entities, introduces semantic quantitative information and structural quantitative information of attribute triples and relation triples to model the entities, captures the semantic features and structural features of the entities in the knowledge graphs, and improves the performance of cross-language entity alignment.

Description

Technical Field

[0001] This invention relates to a cross-language entity alignment method based on meta-attribute reduction and attribute embedding, belonging to the fields of knowledge fusion and artificial intelligence technology. Background Technology

[0002] Currently, massive amounts of big data are characterized by multi-source heterogeneity, fragmentation, and dynamism. Knowledge fusion refers to the integration of multi-source heterogeneous data or knowledge to provide high-quality knowledge that meets user needs and is consistently accurate. Entity alignment is an important research topic in knowledge fusion; it is also known as entity disambiguation or instance alignment. Entity alignment refers to identifying, representing, or referring to the same entity in the real world for entities from different data sources, thereby providing users with more accurate and richer high-quality knowledge. Entity alignment technology is widely used in semantic search, question answering systems, information recommendation, and other fields, and has significant application value.

[0003] Existing entity alignment methods mainly include traditional entity alignment methods and representation learning-based entity alignment methods. Traditional entity alignment methods include machine learning-based methods, similarity calculation-based methods, and probability function-based methods. Similarity calculation-based methods include term frequency-inverse document frequency, machine learning, and active learning models.

[0004] Entity alignment methods based on representation learning include those based on translation models, semantic matching models, and deep models. Translation model-based methods include translation embeddings, hyperplane translation, and adaptive metric function models. Semantic matching model-based methods include complex embeddings and holographic embeddings. Deep model-based methods include projection embeddings, convolutional 2D embeddings, and Relational Graph Convolutional Networks (R-GCN). The information utilized in representation learning-based entity alignment methods includes structural information, attribute information, entity name information, and entity description information.

[0005] This invention studies cross-lingual entity alignment methods, specifically entity alignment techniques for knowledge graphs represented in different natural languages. Currently, cross-lingual entity alignment methods suffer from two main problems. First, existing methods rarely capture the common and unique features of attributes in knowledge graph attribute triples, and rarely mine the semantic relationships between entity attributes. Second, cross-lingual entity alignment methods rarely utilize the attribute quantification and structural quantification information of entities. Summary of the Invention

[0006] The purpose of this invention is to propose a cross-language entity alignment method based on meta-attribute reduction and attribute embedding to align entities in knowledge graphs of different languages. Meta-attribute reduction refers to classifying attributes with similar properties into similar attributes to construct associated attribute pairs for attribute embedding training. Furthermore, a method based on multi-dimensional information measurement is designed to calculate the similarity of entities in cross-language knowledge graphs. This method integrates multi-dimensional information from attribute triples, relation triples, and knowledge graph structural information. The characteristics of this entity alignment method are: firstly, it is independent of the representation language of the knowledge graph, mining the meta-properties of entity attributes and capturing the semantic relationships between attributes; secondly, it utilizes the semantic and structural information of attribute triples and relation triples in the knowledge graph to model entities, capturing the semantic and structural quantitative features of entities in the knowledge graph, and integrating entity embedding and relation embedding, thereby improving the performance of cross-language entity alignment.

[0007] A cross-language entity alignment method based on meta-attribute reduction and attribute embedding includes the following steps:

[0008] Step 1: For two knowledge graphs G1 and G2 of different natural languages, generate attribute embeddings of entities in G1 and G2 respectively;

[0009] Knowledge graph G1 includes attribute triples and relation triples, and knowledge graph G2 includes attribute triples and relation triples.

[0010] Step 1.1: For seed entity pairs, generate weakly related attribute pairs, moderately related attribute pairs, and strongly related attribute pairs;

[0011] Scenario 1: All attributes corresponding to the same entity are weakly correlated pairwise;

[0012] Scenario 2: All attributes corresponding to the seed entity pair are weakly correlated pairwise;

[0013] Scenario 3: If the attribute pair is a weakly related attribute pair and the attribute pairs are of the same type, then the attribute pair is a strongly related attribute pair.

[0014] Meta-attribute reduction refers to classifying attributes with similar semantic features into the same category to construct associated attribute pairs for attribute embedding training. Specifically, the attributes in the attribute triple are classified into the following types according to their attribute values: integer, floating-point, time type, numeric string, person name, place name, and other types.

[0015] Case 4: If the attribute pair is a strongly correlated attribute pair, and the attribute value lengths of the attribute pair in the attribute triple are within the same threshold range, then the attribute pair is a strongly correlated attribute pair.

[0016] Step 1.2: For each entity, generate weakly related attribute pairs, moderately related attribute pairs, and strongly related attribute pairs using the method in Step 1.1;

[0017] Step 1.3: Generate attribute embeddings of entities in knowledge graph G1;

[0018] First, randomly initialize each attribute embedding 'a' and add it to the attribute embedding set AE. The value range for each dimension of attribute embedding 'a' is as follows:

[0019] Secondly, the normalized attribute embedding a∈AE is as follows:

[0020]

[0021] Then, the attribute embeddings in AE are updated according to the objective function, which is shown below:

[0022]

[0023] Among them, W a,b This represents the weight corresponding to the associated attribute pair (a, b). Strongly associated attribute pairs have greater weights than moderately associated attribute pairs, and moderately associated attribute pairs have greater weights than weakly associated attribute pairs. SET0 is a set consisting of weakly associated attribute pairs, moderately associated attribute pairs, and strongly associated attribute pairs, while SET0' is the constructed set of corresponding pseudo-associated attribute pairs. The σ(x) function is defined as follows:

[0024]

[0025] Finally, for each entity e 1i According to entity e 1i All attributes are embedded in a1, a2, ..., a k , generate entity e 1i Attribute embedding a The calculation formula is as follows:

[0026]

[0027] Among them, w1, w2, ..., w k They are a1, a2, ..., a k The weighting coefficients. n is w1, w2, ..., w k The sum of.

[0028] Step 1.4: Following the method in Step 1.3, generate the attribute embeddings of entities in knowledge graph G2;

[0029] Step 2: Generate the relation embeddings of entities in G1 and G2 respectively;

[0030] Step 2.1: Generate the relation embeddings of entities in knowledge graph G1;

[0031] First, randomly initialize each relation embedding r in R, with each dimension taking a range of values.

[0032] Secondly, the normalized relation embedding r∈R, where R is the relation set;

[0033]

[0034] Then, the embeddings of entities in the entity set are randomly initialized, with each dimension taking values ​​within a certain range.

[0035] Finally, update the relation embedding r according to the objective function, which is shown below:

[0036]

[0037] Where h and t represent the head entity and tail entity, respectively, and r represents the relationship between the head entity and tail entity. (h,r,t) represents a true relation triple in the knowledge graph, and (h',r,t') represents an incorrect relation triple, which is constructed by randomly replacing the head entity or tail entity with a true triple. T batch Let be the set consisting of true relation triples and false relation triples. d(h+r,t) is the distance function, representing the distance between h+r and t. γ represents the distance between positive and negative samples in the embedding space, and is a pre-set constant. [x] + This means that for x, if x is positive, it is retained; if x is negative, the result is 0.

[0038] Step 2.2: Following the method in Step 2.1, generate the relation embeddings of entities in knowledge graph G2;

[0039] Step 3: Calculate the similarity M1 between knowledge graphs G1 and G2 based on entity embeddings;

[0040] For entity e in knowledge graph G1 1i splicing the entity e 1iEntity embedding and relation embedding are used to construct the entity embedding matrix X1. For entity e in knowledge graph G2... 2j splicing the entity e 2j Calculate the entity embedding and relation embedding, and construct the entity embedding matrix X2. Calculate entity e. 1i and entity e 2j The similarity matrix M2 is calculated using the following formula:

[0041]

[0042] The elements m of the similarity matrix M1 ij Represents entity e in knowledge graph G1 1i And entity e in knowledge graph G2 2j The similarity between them.

[0043] Step 4: Construct additional information matrices A1 and A2 for knowledge graphs G1 and G2 respectively;

[0044] For each entity in the knowledge graph G1, construct an additional information matrix A1. In matrix A1, rows represent entities and columns represent features.

[0045] For entity e 1i The element in the i-th row of the additional information matrix A1 (a i1 ,a i2 ,…,a im ) represents entity e 1i The matrix A1 contains additional information in multiple dimensions. In this invention, each row of matrix A1 contains 6 elements.

[0046] In the i-th row element of matrix A1 (a i1 ,a i2 ,a i3 ,a i4 ,a i5 ,a i6 )middle,

[0047] (a)a i1 Represents entity e 1i The number of neighboring nodes, which is an integer;

[0048] (b)a i2 Represents entity e 1i The number of attributes, with values ​​being integers;

[0049] (c)a i3 Represents entity e 1i The length of the entity name, which takes the value of a floating-point number;

[0050] (d)a i4 Represents entity e 1i The average length of the attribute names, with values ​​being floating-point numbers;

[0051] (e)a i5 Represents entity e 1i The average length of the attribute value, which takes the value of a floating-point number;

[0052] (f)a i6 Represents entity e 1i The most frequently occurring attribute type in the attribute set, with a value range of [missing information].

[0053] {1,2,3,4,5,6,7}; The attribute types in the attribute triplet include integer, floating-point, time type, numeric string, person name, place name and other types;

[0054] Step 5: Calculate the similarity M2 between knowledge graphs G1 and G2 based on the additional information of the entities;

[0055]

[0056] Step 6: Calculate the similarity M3 between knowledge graphs G1 and G2, and output the entity alignment results;

[0057] The similarity matrix M3 between knowledge graphs G1 and G2 is calculated using the following formula:

[0058] M3 = w1M1w2M2,

[0059] Where M1 is the similarity output in step 3, M2 is the similarity output in step 5, and w1 and w2 are weights.

[0060] For entity e in knowledge graph G1 i Based on the elements in the i-th row of the similarity matrix M3, entity e can be obtained. i The similarity ranking is compared with that of entities in the knowledge graph G2. The output then shows the entity pairs with alignment relationships.

[0061] This completes the entire process of this method.

[0062] Beneficial effects

[0063] The method of this invention performs entity alignment between two cross-language knowledge graphs. It employs an entity alignment method based on meta-attribute reduction and attribute embedding, capable of identifying entity pairs referring to the same entity in the real world. Compared with existing technologies, this method has the following advantages:

[0064] (1) The method described above can align entities in these two cross-language knowledge graphs, construct entity alignment knowledge bases in different natural languages, and provide large-scale knowledge fusion services. In addition, the method is independent of the representation language of the knowledge graph.

[0065] (2) To address the problem that existing methods rarely capture the common and unique features of attributes in knowledge graph attribute triples and rarely mine the semantic relationships between entity attributes, the proposed method designs an entity alignment method based on meta-attribute reduction and attribute embedding. First, for seed entity pairs and entities in the knowledge graph, weakly related attribute pairs, moderately related attribute pairs, and strongly related attribute pairs are constructed based on the meta-attribute reduction method. Then, attribute embeddings and relation embeddings of the entities are generated, thereby constructing entity embeddings. Finally, based on the entity embeddings, the similarity of entities in different knowledge graphs is calculated. The entity alignment method of this invention captures the semantic relationships between attributes by constructing attribute pairs with different association strengths for attribute embedding training, realizes the fusion of entity attribute embeddings and relation embeddings, and improves the performance of cross-language entity alignment.

[0066] (3) To address the issue that entity alignment methods often fail to utilize the attribute and structural quantification information of entities, the proposed method designs a multi-dimensional information measurement approach to calculate the similarity of entities in cross-lingual knowledge graphs. This similarity calculation method leverages the multi-dimensional additional information of entities in the knowledge graph, introducing semantic and structural quantification information from attribute triples and relation triples to model entities. It captures the semantic and structural features of entities in the knowledge graph, and integrates knowledge graph similarity calculation methods based on entity embedding and multi-dimensional additional information, thereby improving the performance of cross-lingual entity alignment.

[0067] (4) Experiments were conducted on a public dataset, and the experimental results demonstrate the effectiveness and superiority of the proposed method. The method has broad application prospects in fields such as information recommendation, question answering systems, and information retrieval. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the process for cross-language entity alignment based on meta-attribute reduction and attribute embedding proposed in this invention. Detailed Implementation

[0069] The preferred embodiments of the method of the present invention will be described in detail below with reference to examples.

[0070] Example

[0071] A cross-language entity alignment method based on meta-attribute reduction and attribute embedding, such as Figure 1 As shown, it includes the following steps:

[0072] Step 1: For two knowledge graphs G1 and G2 of different natural languages, generate attribute embeddings of entities in G1 and G2 respectively;

[0073] Knowledge graph G1 includes attribute triples and relation triples, and knowledge graph G2 includes attribute triples and relation triples.

[0074] For example, cross-language knowledge graphs include Chinese and English knowledge graphs. Both Chinese and English knowledge graphs include relation triples and attribute triples. Attribute triples are in the form (entity, attribute name, attribute value). For example, a Chinese attribute triple is represented as (United Arab Emirates, official language, Arabic), and an English attribute triple is in the form (The United Arab Emirates, official language, Arabic). The objective of this invention is to identify entity pairs (United Arab Emirates, The United Arab Emirates) that have an alignment relationship in the Chinese and English knowledge graphs.

[0075] Step 1.1: For seed entity pairs, generate weakly related attribute pairs, moderately related attribute pairs, and strongly related attribute pairs;

[0076] Scenario 1: All attributes corresponding to the same entity are weakly correlated pairwise;

[0077] Scenario 2: All attributes corresponding to the seed entity pair are weakly correlated pairwise;

[0078] Scenario 3: If the attribute pair is a weakly related attribute pair and the attribute pairs are of the same type, then the attribute pair is a strongly related attribute pair.

[0079] Meta-attribute reduction refers to classifying attributes with similar semantic features into the same category to construct associated attribute pairs for attribute embedding training. Specifically, the attributes in the attribute triple are classified into the following types according to their attribute values: integer, floating-point, time type, numeric string, person name, place name, and other types.

[0080] Case 4: If the attribute pair is a strongly correlated attribute pair, and the attribute value lengths of the attribute pair in the attribute triple are within the same threshold range, then the attribute pair is a strongly correlated attribute pair.

[0081] Step 1.2: For each entity, generate weakly related attribute pairs, moderately related attribute pairs, and strongly related attribute pairs using the method in Step 1.1;

[0082] Step 1.3: Generate attribute embeddings of entities in knowledge graph G1;

[0083] First, randomly initialize each attribute embedding 'a' and add it to the attribute embedding set AE. The value range for each dimension of attribute embedding 'a' is as follows:

[0084] Secondly, the normalized attribute embedding a∈AE is as follows:

[0085]

[0086] Then, the attribute embeddings in AE are updated according to the objective function, which is shown below:

[0087]

[0088] Among them, W a,b This represents the weight corresponding to the associated attribute pair (a, b). Strongly associated attribute pairs have greater weights than moderately associated attribute pairs, and moderately associated attribute pairs have greater weights than weakly associated attribute pairs. SET0 is a set consisting of weakly associated attribute pairs, moderately associated attribute pairs, and strongly associated attribute pairs, while SET0' is the constructed set of corresponding pseudo-associated attribute pairs. The σ(x) function is defined as follows:

[0089]

[0090] Finally, for each entity e 1i According to entity e 1i All attributes are embedded in a1, a2, ..., a k , generate entity e 1i Attribute embedding a The calculation formula is as follows:

[0091]

[0092] Among them, w1, w2, ..., w k They are a1, a2, ..., a k The weighting coefficients. n is w1, w2, ..., w k The sum of.

[0093] In this embodiment, w1, w2, ..., w k It can take the value 1, 2 or 3.

[0094] Step 1.4: Following the method in Step 1.3, generate the attribute embeddings of entities in knowledge graph G2;

[0095] Step 2: Generate the relation embeddings of entities in G1 and G2 respectively;

[0096] Step 2.1: Generate the relation embeddings of entities in knowledge graph G1;

[0097] First, randomly initialize each relation embedding r in R, with each dimension taking a range of values.

[0098] Secondly, the normalized relation embedding r∈R, where R is the relation set;

[0099]

[0100] Then, the embeddings of entities in the entity set are randomly initialized, with each dimension taking values ​​within a certain range.

[0101] Finally, update the relation embedding r according to the objective function, which is shown below:

[0102]

[0103] Where h and t represent the head entity and tail entity, respectively, and r represents the relationship between the head entity and tail entity. (h,r,t) represents a true relation triple in the knowledge graph, and (h',r,t') represents an incorrect relation triple, which is constructed by randomly replacing the head entity or tail entity with a true triple. T batch Let be the set consisting of true relation triples and false relation triples. d(h+r,t) is the distance function, representing the distance between h+r and t. γ represents the distance between positive and negative samples in the embedding space, and is a pre-set constant. [x] + This means that for x, if x is positive, it is retained; if x is negative, the result is 0.

[0104] Step 2.2: Following the method in Step 2.1, generate the relation embeddings of entities in knowledge graph G2;

[0105] Step 3: Calculate the similarity M1 between knowledge graphs G1 and G2 based on entity embeddings;

[0106] For entity e in knowledge graph G1 1i Then, concatenate the entity embeddings and relation embeddings to construct the entity embedding matrix X1. For entity e in knowledge graph G2... 2j Concatenate the entity embeddings and relation embeddings to construct the entity embedding matrix X2. Calculate entity e. 1i and entity e 2j The similarity matrix M2 is calculated using the following formula:

[0107]

[0108] The elements m of the similarity matrix M1 ij Represents entity e in knowledge graph G1 1i And entity e in knowledge graph G2 2j The similarity between them.

[0109] Step 4: Construct additional information matrices A1 and A2 for knowledge graphs G1 and G2 respectively;

[0110] For each entity in the knowledge graph G1, construct an additional information matrix A1. In matrix A1, rows represent entities and columns represent features.

[0111] For entity e 1iThe element in the i-th row of the additional information matrix A1 (a i1 ,a i2 ,…,a im ) represents entity e 1i The matrix A1 contains additional information in multiple dimensions. In this invention, each row of matrix A1 contains 6 elements.

[0112] The element in the i-th row of matrix A1 (a i1 ,a i2 ,a i3 ,a i4 ,a i5 ,a i6 )middle,

[0113] (a)a i1 Represents entity e 1i The number of neighboring nodes, which is an integer;

[0114] (b)a i2 Represents entity e 1i The number of attributes, with values ​​being integers;

[0115] (c)a i3 Represents entity e 1i The length of the entity name, which takes the value of a floating-point number;

[0116] (d)a i4 Represents entity e 1i The average length of the attribute names, with values ​​being floating-point numbers;

[0117] (e)a i5 Represents entity e 1i The average length of the attribute value, which takes the value of a floating-point number;

[0118] (f)a i6 Represents entity e 1i The most frequently occurring attribute type in the attribute set, with a value range of [missing information].

[0119] {1,2,3,4,5,6,7}; The attribute types in the attribute triplet include integer, floating-point, time type, numeric string, person name, place name and other types;

[0120] Step 5: Calculate the similarity M2 between knowledge graphs G1 and G2 based on the additional information of the entities;

[0121]

[0122] Step 6: Calculate the similarity M3 between knowledge graphs G1 and G2, and output the entity alignment results;

[0123] The similarity matrix M3 between knowledge graphs G1 and G2 is calculated using the following formula:

[0124] M3 = w1M1w2M2,

[0125] Where M1 is the similarity output in step 3, M2 is the similarity output in step 5, and w1 and w2 are weights.

[0126] For entity e in knowledge graph G1 i Based on the elements in the i-th row of the similarity matrix M3, entity e can be obtained. i The similarity ranking is compared with that of entities in the knowledge graph G2. The output then shows the entity pairs with alignment relationships.

[0127] To illustrate the cross-language entity alignment effect of this invention, this experiment compares the entity alignment effects using three different methods on the same dataset under identical conditions. The first method is a translation-based cross-language knowledge graph embedding entity alignment method, which encodes entities and relations for each language in an independent embedding space. The second method is a cross-language entity alignment method based on an attribute-preserving embedding model, which jointly embeds the structures of two knowledge graphs into a unified vector space. The third method is the cross-language entity alignment method of meta-attribute reduction and attribute embedding of this invention.

[0128] The evaluation metrics used were Hits@1, Hits@10, Hits@50, and Mean Rank. The four evaluation values ​​for the first method were 30.83, 61.41, 79.12, and 154, respectively. For the second method, the four evaluation values ​​were 41.18, 74.46, 88.90, and 64, respectively. The four evaluation values ​​for cross-language entity alignment based on meta-attribute reduction and attribute embedding proposed in this invention were 43.98, 75.60, 89.31, and 50, respectively.

[0129] The above description is merely a preferred embodiment of the present invention, and the present invention should not be limited to the content disclosed in this embodiment and the accompanying drawings. Any equivalent or modified embodiments made without departing from the spirit of the present invention fall within the scope of protection of the present invention.

Claims

1. A cross-language entity alignment method based on meta-attribute reduction and attribute embedding, characterized in that: Step 1: For two knowledge graphs G1 and G2 of different natural languages, generate attribute embeddings of entities in G1 and G2 respectively; Step 1.1: For seed entity pairs, generate weakly related attribute pairs, moderately related attribute pairs, and strongly related attribute pairs; Step 1.2: For each entity, generate weakly related attribute pairs, moderately related attribute pairs, and strongly related attribute pairs using the method in Step 1.1; Step 1.3: Generate attribute embeddings of entities in knowledge graph G1; Step 2: Generate the relation embeddings of entities in G1 and G2 respectively; Step 2.1: Generate the relation embeddings of entities in knowledge graph G1; Step 2.2: Following the method in Step 2.1, generate the relation embeddings of entities in knowledge graph G2; Step 3: Calculate the similarity M1 between knowledge graphs G1 and G2 based on entity embeddings; For entity e in knowledge graph G1 1i splicing the entity e 1i Entity embedding and relation embedding are used to construct the entity embedding matrix X1; for entity e in knowledge graph G2 2j splicing the entity e 2j Entity embedding and relation embedding are used to construct the entity embedding matrix X2; entity e is calculated. 1i and entity e 2j The similarity matrix M2 is calculated using the following formula: The elements m of the similarity matrix M1 ij Represents entity e in knowledge graph G1 1i And entity e in knowledge graph G2 2j The similarity between them; Step 4: Construct additional information matrices A1 and A2 for knowledge graphs G1 and G2 respectively; Step 5: Calculate the similarity M2 between knowledge graphs G1 and G2 based on the additional information of the entities; Step 6: Calculate the similarity M3 between knowledge graphs G1 and G2, and output the entity alignment results; The similarity matrix M3 between knowledge graphs G1 and G2 is calculated using the following formula: Where M1 is the similarity output in step 3, M2 is the similarity output in step 5, and w1 and w2 are weights; For entity e in knowledge graph G1 i Based on the elements in the i-th row of the similarity matrix M3, entity e is obtained. i The similarity ranking is performed with entities in the knowledge graph G2; thus, entity pairs with alignment relationships are output.

2. The cross-language entity alignment method based on meta-attribute reduction and attribute embedding according to claim 1, characterized in that: In step 1, for two knowledge graphs G1 and G2 of different natural languages, attribute embeddings of entities in G1 and G2 are generated respectively; Specifically: Knowledge graph G1 includes attribute triples and relation triples, and knowledge graph G2 includes attribute triples and relation triples; Step 1.1: For seed entity pairs, generate weakly related attribute pairs, moderately related attribute pairs, and strongly related attribute pairs; Scenario 1: All attributes corresponding to the same entity are weakly correlated pairwise; Scenario 2: All attributes corresponding to the seed entity pair are weakly correlated pairwise; Scenario 3: If the attribute pair is a weakly related attribute pair and the attribute pairs are of the same type, then the attribute pair is a strongly related attribute pair. Meta-attribute reduction refers to classifying attributes with similar semantic features into the same category to construct associated attribute pairs for attribute embedding training. Specifically, the attributes in the attribute triple are classified into the following types according to their attribute values: integer, floating-point, time type, numeric string, person name, place name, and other types. Case 4: If the attribute pair is a strongly correlated attribute pair, and the attribute value lengths of the attribute pair in the attribute triple are within the same threshold range, then the attribute pair is a strongly correlated attribute pair. Step 1.2: For each entity, generate weakly related attribute pairs, moderately related attribute pairs, and strongly related attribute pairs using the same method as in Step 1.1; Step 1.3: Generate attribute embeddings of entities in knowledge graph G1; First, randomly initialize each attribute embedding 'a' and add it to the attribute embedding set AE, where the range of values ​​for each dimension of attribute embedding 'a' is: ; Secondly, the normalized attribute embedding a∈AE is as follows: Then, the attribute embeddings in AE are updated according to the objective function, which is shown below: Among them, W a,b This represents the weight corresponding to the associated attribute pair (a, b). The weight of a strongly associated attribute pair is greater than the weight of a moderately associated attribute pair, and the weight of a moderately associated attribute pair is greater than the weight of a weakly associated attribute pair. SET0 is a set consisting of weakly associated attribute pairs, moderately associated attribute pairs, and strongly associated attribute pairs. It is the constructed set of corresponding pseudo-associative attribute pairs. The function is defined as follows: Finally, for each entity e 1i According to entity e 1i All attributes are embedded in a1, a2, ..., a k , generate entity e 1i Attribute embedding a The calculation formula is as follows: Where w1, w2, …, w k They are a1, a2, …, a k The weighting coefficients, n being w1, w2, …, w k The sum of; Step 1.4: Following the method in Step 1.3, generate the attribute embeddings of entities in knowledge graph G2.

3. The cross-language entity alignment method based on meta-attribute reduction and attribute embedding according to claim 1, characterized in that: In step 2, the relation embeddings of entities in G1 and G2 are generated respectively; Specifically: Step 2.1: Generate the relation embeddings of entities in knowledge graph G1; First, randomly initialize each relation embedding r in R, with each dimension taking a range of values. ; Secondly, the normalized relation embedding r∈R, where R is the relation set; Then, the embeddings of entities in the entity set are randomly initialized, with each dimension taking values ​​within a certain range. ; Finally, update the relation embedding r according to the objective function, which is shown below: Where h and t represent the head entity and tail entity respectively, r represents the relationship between the head entity and tail entity, (h,r,t) represents the true relation triple in the knowledge graph, and (h',r,t') represents the erroneous relation triple, which is constructed by randomly replacing the head entity or tail entity with the true triple, T batch Let be a set consisting of true relation triples and false relation triples, and d(h+r, t) be a distance function representing the distance between h+r and t. [x] represents the distance between positive and negative samples in the embedding space, and is a pre-set constant. + This means that for x, if x is positive, it is retained; if x is negative, the result is 0. Step 2.2: Following Step 2.1, generate the relation embeddings of entities in knowledge graph G2.

4. The cross-language entity alignment method based on meta-attribute reduction and attribute embedding according to claim 1, characterized in that: In step 4, additional information matrices A1 and A2 are constructed for knowledge graphs G1 and G2, respectively; Specifically, for entities in knowledge graph G1, construct additional information matrix A1; in matrix A1, rows represent entities and columns represent features. for Entity e 1i The element in the i-th row of the additional information matrix A1 (a i1 , a i2 ,…, a im ) represents entity e 1i The matrix A1 contains 6 elements in each row, providing additional multi-dimensional information. In the i-th row element of matrix A1 (a i1 , a i2 , a i3 , a i4 , a i5 , a i6 )middle, (a)a i1 Represents entity e 1i The number of neighboring nodes, which is an integer; (b)a i2 Represents entity e 1i The number of attributes, with values ​​being integers; (c)a i3 Represents entity e 1i The length of the entity name, which takes the value of a floating-point number; (d)a i4 Represents entity e 1i The average length of the attribute names, with values ​​that are floating-point numbers; (e)a i5 Represents entity e 1i The average length of the attribute value, which takes the value of a floating-point number; (f) a i6 Represents entity e 1i The most frequently occurring attribute type in the attribute set has a value range of {1,2,3,4,5,6,7}; the attribute types in the attribute triple include integer, floating-point, time type, numeric string, person name, place name and other types.

Images