Weakly supervised entity alignment optimization method and system for cross-language knowledge graph

Abstract

The application discloses a weakly supervised entity alignment optimization method and system for a cross-language knowledge graph. The method introduces a lightweight unsupervised module based on neighborhood relations and semantic characters to mine a candidate seed set participating in supervised training under the condition of missing labeled data, comprehensively evaluates the reliability of the candidate seed set by using semantic character level distance measurement, and optimizes the model joint training loss function by taking the reliability as a reference, so that the joint training of the candidate seed set and the labeled data is realized, thereby reducing the dependence of the supervised algorithm on the labeled data. Experiments on DBP15K show that the application can effectively reduce the size of the labeled data, and only 10% of the labeled data can achieve an accuracy similar to that of the Dual-AMN method. Compared with the baseline method, the performance of the application method in the Hit@1 value is at least improved by 3%.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of cross-language knowledge graph entity alignment, and particularly relates to a weakly supervised entity alignment optimization method and system for cross-language knowledge graphs. BACKGROUND

[0002] In recent years, with the research on multilingual knowledge graph embedding, entity potential semantic representation and cross-language knowledge reasoning tasks have achieved certain results, thus promoting many knowledge-driven cross-language work. However, due to the low degree of entity alignment between knowledge graphs of different languages, the accuracy of downstream work such as cross-language reasoning, cross-language knowledge graph completion and cross-language connection prediction is often not satisfactory. In real life, there are inevitably many minority language entities. In the e-commerce, question-answering system and artificial intelligence fields, knowledge graphs of different languages have both intersections and differences. Entity alignment and integration of knowledge graphs of different languages can greatly enrich the professional information of each graph. Therefore, mining more aligned entities between cross-language knowledge graphs is crucial for many downstream tasks and practical applications.

[0003] Cross-language knowledge graphs store real-world knowledge in the form of triples (for example, <entity, relationship, entity>), and entities exist in at least two or more language representations. Among them, the relationship connects two entities. At present, knowledge graphs have been widely used in information retrieval and artificial intelligence fields. Since different knowledge graphs often contain complementary information, these complementary information plays an important role in improving the quality of knowledge graphs. Therefore, many researchers are committed to the alignment of knowledge graphs. Entity alignment, as a key step in knowledge graph alignment, aims to identify equivalent entities in different knowledge graphs, which is of great significance to knowledge fusion and the construction of higher quality knowledge graphs.

[0004] Most of the existing entity alignment methods can be divided into supervised methods, unsupervised methods and semi-supervised methods. The supervised entity alignment task maps multiple knowledge graphs to a unified vector space, and then relies on a large amount of labeled data as training data, and finally calculates the distance between entities to obtain aligned entities. Considering that it is difficult to obtain labeled data, manually collecting enough labeled seeds is very expensive and time-consuming. In order to eliminate the dependence on labeled data, previous studies have researched unsupervised entity alignment methods. The unsupervised entity alignment task generally obtains some approximate candidate seed sets according to the side information of the knowledge graph. However, this method has a noise problem, and how to effectively solve the noise problem to filter out more reliable alignment seeds has become a difficult point that needs to be solved in unsupervised methods. Semi-supervised learning can use labeled data and unlabeled data to complete the entity alignment task, and the existing semi-supervised entity alignment method selects a number of high-confidence alignment entities to add to the training set in each iteration, which can effectively reduce the dependence on labeled data.

[0005] Generally speaking, compared with unsupervised information, labeled data can provide more reliable information, and if a small amount of labeled data is used for training in the entity alignment task, the process often needs the participation of unsupervised data. However, the alignment entity pairs generated by the unsupervised method still have a high probability of error, and it is very challenging to filter out more reliable unsupervised candidate seeds and link them well with labeled data. The existing entity alignment method needs sufficient and reliable labeled data to complete the task, and has strict requirements for the quantity and quality of the labeled data. If unreliable candidate seeds are added to the basic ground truth set, early errors will be reinforced in learning, and eventually these candidate seeds may introduce and amplify noise, reducing the accuracy of entity alignment. It can be seen that labeled data directly participates in the training process and affects the final result, and plays a key role in the field of entity alignment. SUMMARY

[0006] In view of the problem that the supervised method in the entity alignment has high dependence on labeled data, the application provides a weakly supervised entity alignment optimization method and system for cross-language knowledge graphs, which supplements the limited training data in the supervised method. In the whole training process, the weight information generated by the candidate seed optimization selection algorithm determines the importance of each pair of entities, and the loss function based on the weight optimizes the training process. The advantage of this method is that the model can be trained under small labeled data, at this time the seed and labeled data are put into a unified learning framework for iterative training, which is a semi-supervised method. The model can also use the automatically generated candidate seed set for training, which is an unsupervised method.

[0007] In order to achieve the above purpose, the application adopts the following technical scheme:

[0008] The application provides a weakly supervised entity alignment optimization method for a cross-language knowledge graph.

[0009] Step 1: Extract the neighborhood relationship and semantic character information of the entity in the cross-language knowledge graph, and obtain a preliminary candidate seed set by similarity calculation.

[0010] Step 2: Evaluate the reliability of the candidate seed set by comprehensively considering the semantic and character information, and find candidate seeds with high reliability.

[0011] Step 3: Joint training of the candidate seeds selected by optimization and part of the labeled data.

[0012] Step 4: Use the obtained reliability to improve the loss function, and finally optimize the entire training process.

[0013] Further, the step 1 comprises:

[0014] Sparse matrix is used to store the entity and relationship information in the knowledge graph to obtain a neighbor information matrix, wherein the diagonal information in the matrix is the ratio of the number of all triples to the number of entities, representing the weight between entities connected to themselves.

[0015] The weight information between entities is calculated according to the following formula:

[0016]

[0017] Wherein, l t represents the number of triples, represents the relationship r ij between entity e i and entity e j appears in all triples, the higher the weight, the closer the connection between the two entities.

[0018] For the source and target cross-language knowledge graphs, first, use a machine translation system to translate the non-English property names into English; then use Glove word embedding to obtain the semantic-level average word vector of the property name, and use the average word vector to represent the semantic information of the entity name; for each property name, divide it into two characters to obtain the character-level embedding vector; finally, calculate the semantic and character-level similarity of the two knowledge graphs using the cosine similarity function to obtain a preliminary candidate seed set.

[0019] Further, the step 2 comprises:

[0020] Firstly, based on the obtained candidate seed set, entity pairs with each other being the nearest distance are obtained, the entity pairs being calculated by entity neighborhood and semantic information; then semantic level vectors and character level vectors of the entity pairs are extracted respectively, similarity calculation is performed on the two kinds of vectors, and semantic and character level reliabilities between entities are obtained; finally, the two kinds of reliabilities are integrated to evaluate the candidate seeds.

[0021] Further, the step 3 comprises:

[0022] The candidate seed set is stored in a set U, and the labeled data is stored in a set L, w1 proportion of data is selected in the set U, w2 proportion of data is selected in the set L, the selected data satisfies formula (9), and there is no repetition:

[0023] U*w1+L*w2=L*n (9)

[0024] Wherein, each parameter range needs to satisfy the following formula:

[0025] w1∈[0,n],w2∈[0,n],n∈[0,1],w1+w2=n (10)

[0026] Next, different values of w1 and w2 are taken to select different proportions of candidate seeds and labeled data, and finally the two kinds of selected data are jointly trained.

[0027] Further, the step 4 comprises:

[0028] In the training process, each pair of entities (u, v) in the selected candidate seed is assigned weight information D final (u, v), the weight score of the labeled data is set to 1, the sample loss with fixed mean and variance is used to reduce the dependence on sample size and hyperparameters, the improvement of the loss function is completed, and the whole training process is finally optimized.

[0029] Another aspect of the application provides a weakly supervised entity alignment optimization system for cross-language knowledge graph, comprising:

[0030] A preliminary candidate seed set derivation module is used to extract the neighborhood relationship and semantic character information of the entity in the cross-language knowledge graph, and a preliminary candidate seed set is obtained by using similarity calculation;

[0031] A candidate seed optimization selection module is used to evaluate the reliability of the candidate seed set by comprehensively considering semantic and character information, and find candidate seeds with high reliability;

[0032] A joint training module is used to jointly train the candidate seeds selected by optimization and part of the labeled data;

[0033] The loss function improvement module is configured to improve the loss function by using the obtained credibility, and finally optimize the entire training process.

[0034] Further, the preliminary candidate seed set derivation module is specifically configured to:

[0035] Sparse matrix is used to store entity and relationship information in the knowledge graph to obtain a neighbor information matrix, wherein diagonal information in the matrix is a ratio of a total number of triples to a number of entities, representing a weight between entities connected to each other.

[0036] The weight information between entities is calculated according to the following formula:

[0037]

[0038] wherein, l t represents the number of triples, represents a relationship r i between the entity e j and the entity e ij The higher the weight, the closer the connection between the two entities.

[0039] For the source and target cross-language knowledge graphs, first, a machine translation system is used to translate non-English property names into English; then, a semantic-level average word vector of the property name is obtained by using Glove word embedding, and the average word vector is used to represent semantic information of the entity name; for each property name, the property name is divided into character-level embedding vectors in units of two characters; finally, the obtained semantic and character-level vectors are used to calculate the semantic and character-level similarities of the two knowledge graphs by using a cosine similarity function, to obtain a preliminary candidate seed set.

[0040] Further, the candidate seed optimization selection module is specifically configured to:

[0041] First, based on the obtained candidate seed set, entity pairs that are closest to each other are obtained, the entity pairs being calculated based on entity neighborhood and semantic information; then, semantic-level vectors and character-level vectors of the entity pairs are extracted respectively, and similarity calculations are performed on the two types of vectors to obtain semantic and character-level credibility between entities; finally, the two types of credibility are integrated to evaluate the candidate seeds.

[0042] Further, the joint training module is specifically configured to:

[0043] The candidate seed set is stored in a set U, and the labeled data is stored in a set L, w1 proportion of data in the set U is selected, w2 proportion of data in the set L is selected, the selected data satisfies formula (9), and there is no repetition:

[0044] U*w1+L*w2=L*n (9)

[0045] wherein each parameter range needs to satisfy the following formula:

[0046] w1∈[0,n],w2∈[0,n],n∈[0,1],w1+w2=n (10)

[0047] Next, different values are taken for w1 and w2 to select different proportions of candidate seeds and labeled data, and finally the selected two kinds of data are jointly trained.

[0048] Further, the loss function improvement module is specifically used for:

[0049] In the training process, each pair of entities (u, v) in the selected candidate seeds is assigned weight information D final (u, v), the weight score of the labeled data is set to 1, the sample loss with fixed mean and variance is used to reduce the dependence on sample size and hyperparameters, so as to improve the loss function and finally optimize the entire training process.

[0050] Compared with the prior art, the present application has the beneficial effects:

[0051] The present application analyzes the possibility of interaction between labeled data and unlabeled data from the deficiencies of existing entity alignment methods, and proposes a weakly supervised entity alignment optimization method and system for cross-language knowledge graph. The present application uses the candidate seed set generated by the unsupervised method as supplementary data, selects candidate seeds with higher reliability by comprehensively considering semantic and character level distance measurement, and then inputs the candidate seeds and labeled data into the model for joint training, and uses a loss function based on credibility weight to optimize the entire training process. Experimental results show that the method of the present application can achieve experimental results comparable to the Dual-AMN method with only 10% labeled data, and compared with all baseline methods, the performance of the method of the present application on Hit@1 value is at least improved by 3%. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 It is a schematic diagram of the Dual-AMN framework;

[0053] Figure 2 It is an example of labeled data;

[0054] Figure 3 It is a weakly supervised entity alignment optimization method (WSEO) model framework for cross-language knowledge graph according to an embodiment of the present application;

[0055] Figure 4 It is a candidate seed node generation process according to an embodiment of the present application;

[0056] Figure 5 For the overall idea of the distance metric method of the embodiment of the application;

[0057] Figure 6 For the performance comparison of the WSEO and Dual-AMN of the embodiment of the application on three data sets using equivalent labeled data;

[0058] Figure 7 For the embodiment of the application, a weakly supervised entity alignment optimization system architecture diagram for a cross-language knowledge graph. DETAILED DESCRIPTION

[0059] The application will be further explained and described below in conjunction with the accompanying drawings and specific embodiments:

[0060] (a) Problem definition

[0061] To solve the problem of high time complexity in entity alignment, Mao [Mao X, Wang W, Wu Y, et al. Boosting the speed of entity alignment 10x: Dual attention matching network with normalized hard sample mining [C] / / Proceedings of the Web Conference 2021. 2021: 821-832.] et al. proposed a less time-consuming entity alignment method (Dual-AMN), which uses a simplified encoder and a normalized hard sample mining technique to reduce the time complexity of the algorithm, and uses the relationship, neighbor, etc. of the knowledge graph to generate a single entity vector and a cross graph vector, and finally combines the distance metric algorithm to obtain the aligned entity pair.

[0062] The Dual-AMN method mainly includes three steps: single entity vector acquisition, cross graph vector acquisition, and distance metric. Figure 1 The main process of the method is described. The single entity vector acquisition uses a simplified relationship attention layer, which uses a relationship projection operation instead of a standard linear conversion matrix to reduce the additional parameters brought by each entity, and finally splices the embedding of each layer to expand the multi-hop neighbor information, Figure 1 ①The GNN network is used to obtain the entity vector of a single knowledge graph, where e1 and e2 represent the aligned entities; the cross graph vector acquisition uses a limited proxy vector set to represent the alignment relationship information of the cross graph, and these limited proxy vectors are obtained by feeding the entity, relationship and neighbor information into the GNN model, if two entities are equivalent, then their distances to these proxy vectors are also consistent, it can be seen that, Figure 1 ②In Figure 1the agent vector of the cross-graph is further obtained on the basis of the entity vector of the cross-graph; after obtaining the entity vector of the cross-graph, the distance between the entity vector and the agent vector of the cross-graph needs to be calculated, and the entities with consistent distances are considered as aligned entity pairs; in the training process, the L2 distance is used to obtain the similarity between the entity and the agent vector, and in the testing process, the model uses the CSLS distance measurement to reduce the dimension disaster. Figure 1 ③the distance between each entity and the agent vector is compared, and the entities with consistent distances are considered as aligned entity pairs; finally, the Dual-AMN model uses the above three steps to obtain the aligned entity pairs.

[0063] Compared with the previous supervised entity alignment algorithm, the Dual-AMN entity alignment method combines a single entity vector with a cross-graph vector, uses a simplified relationship attention layer and an agent matching attention layer to simplify the encoder, and uses a standard hard sampling technique to select information-rich negative samples. The above steps shorten the running time of the entire algorithm, which not only solves the problem that the entity alignment algorithm often takes a long time, but also ensures that the model has a high accuracy.

[0064] However, this method has two shortcomings:

[0065] 1) The Dual-AMN entity alignment method relies on a large amount of labeled data for training.

[0066] The Dual-AMN entity alignment method relies on 30% of the labeled data for training, and the accuracy of the model significantly decreases when the labeled data is less than 30%. Obviously, the Dual-AMN entity alignment method cannot obtain a high accuracy rate using a small amount of labeled data. However, in real life, it is difficult to obtain sufficient labeled data, and the labeling cost is high, so relying too much on labeled data is unrealistic. Table 1 below shows the experimental results of Dual-AMN under different amounts of labeled data, which clearly shows that the experimental results decrease with the decrease of labeled data, and the experimental results become better and better with the increase of labeled data.

[0067] Table 1 Hit rate comparison of Dual-AMN under different proportions of alignment seeds

[0068]

[0069] 2) The labeled data in the entity alignment data set may also contain error information.

[0070] The existing entity alignment method uses labeled data to connect two knowledge graphs, and if two entities are labeled as aligned entities, they are stored as a single entity in a unified vector space. However, some labeled data may not exist in real life. For example, Figure 2Part of the English knowledge graph and the Spanish knowledge graph are shown, and the entity alignment task aims to identify equivalent entity pairs between the two knowledge graphs. However, Hirokazu Koreeda in English and Hirokazu Koreeda in Spanish cannot be equivalent because they do not exist in real life. When training the model, the iteration process will propagate and amplify these noises, thereby reducing the entity alignment performance.

[0071] The principle and shortcomings of the Dual-AMN entity alignment method are described above, and the main problems involved in the method of the application are defined as follows.

[0072] (b) Entity alignment definition

[0073] Cross-language knowledge graph: A cross-language knowledge graph can be described as a graph G=(E,R,T) containing three sets, The information of the real world is stored in the form of triples <entity, relation, entity>, where triples are used to describe the inherent relationship between entities. Where E represents the entity set, R represents the set of relations, and the entity is a specific thing (such as Peking University, China Beijing, etc.), and the edge represents the connection between things (such as located in), then the triple <Peking University, located in, China Beijing> represents the meaning that Peking University is located in China Beijing.

[0074] Entity alignment: Entity alignment aims to identify equivalent entity pairs in two cross-language knowledge graphs. Given two knowledge graphs G1=(E1,R1,T1) and G2=(E2,R2,T2), where E1, E2 represents the set of entities, R1, R2 represents the relationship between entities, T1, T2 represents the triple set, and the labeled data P={(u,v)|u∈E1,v∈E2,u≡v}, where ≡ represents equivalence, the model ultimately obtains more potential equivalent entity pairs through G1, G2 and P.

[0075] Weakly supervised entity alignment for cross-language knowledge graph: The Dual-AMN method relies too much on labeled data, and the improvement strategy is proposed to address the shortcomings of the Dual-AMN method. This method aims to generate a candidate seed set C using neighborhood relations and semantic character information, and then generate training data from the candidate seed set C and the labeled data P={(u,v)|u∈E1,v∈E2,u≡v}. Finally, the algorithm uses G1=(E1,R1,T1), G2=(E2,R2,T2), C and P to obtain aligned entity pairs.

[0076] 1、The method of the application

[0077] Due to the great similarity of entity semantic names in cross-language knowledge graphs, the similarity is more easily captured by the model after the cross-language entities are translated into the same language by a machine translation system. In order to realize the weakly supervised entity alignment optimization framework and use as little labeled data as possible in the whole model, the present application proposes a weakly supervised entity alignment optimization method (WSEO) for cross-language knowledge graphs. WSEO first generates a candidate seed set using the semantic similarity between entities. Here we do not use a complex model to generate a candidate seed set, but use the entity neighborhood relationship as auxiliary information of the attribute name to obtain a more comprehensive neighborhood-semantic similarity.

[0078] As shown in Figure 3 , our method model includes two parts: an unsupervised candidate seed generation part and a supervised training part. The unsupervised candidate seed generation part mainly includes candidate node generation and optimization selection of candidate seeds, and the supervised training part includes joint training of candidate seeds and labeled data and modification of the loss function.

[0079] Given two cross-language knowledge graphs, first, the neighborhood relationship and semantic character information of the entities in the knowledge graphs are extracted by the unsupervised candidate seed generation part, and the cosine similarity is used to find a preliminary candidate seed set. Then, in the optimization selection algorithm, the semantic and character information are integrated to evaluate the reliability of the candidate seed set, and candidate seeds with higher reliability are found. In the supervised training part, the candidate seeds selected by optimization and part of the labeled data are fed into the entity alignment method model for joint training. The reliability obtained by the optimization selection algorithm is used to improve the loss function, and finally the whole training process is optimized.

[0080] The important role of the above-mentioned candidate seed generation is to generate a preliminary candidate seed set. The main problem of this part is to generate the most similar N nodes of the entity and the corresponding weight information. Since these preliminary candidate seeds contain error information, further optimization selection is performed on these preliminary candidate seeds. The candidate seeds obtained after optimization selection are the training data we want to use. Since the optimization selection weights of each node are different, in order to obtain more scientific training results, we propose a loss function improvement strategy based on reliability weight. In the whole method model, how to select candidate seeds with higher reliability directly affects the experimental performance, therefore, the optimization selection of candidate seeds is an important link to determine the quality of the model.

[0081] 1.1 Candidate seed generation

[0082] This chapter focuses on the generation of candidate seed sets. Candidate seeds are unsupervised data that can supplement labeled data, and are used to construct training data in different proportions during training. We believe that a more optimized candidate seed selection algorithm can achieve higher accuracy, while overly complex models can lead to redundancy. If we can obtain an initial candidate seed set using readily available source data and convenient, efficient methods, and then effectively optimize and select from these candidate sets, achieving accuracy comparable to existing methods, then we consider this work meaningful.

[0083] Knowledge graphs contain a wealth of lateral information, such as entity neighborhoods, attributes, and descriptions. Due to significant differences in the textual representations of cross-lingual knowledge graphs, conventional knowledge graph alignment methods are not applicable to cross-lingual entity alignment. However, cross-lingual entities share similar structural information. Therefore, entity alignment of cross-lingual knowledge graphs requires either determining the similarity between entities based on neighborhood information or converting cross-lingual entities into entities of the same language using specific methods before determining alignment. Neighborhood information plays a crucial role in entity alignment; if two entities are an aligned pair, their corresponding neighborhood information is also similar. Attribute names are widely present in most knowledge graphs; considering data availability, this work uses typical attribute names as source data. Neighborhood information and attribute information are combined to generate an initial candidate seed set.

[0084] 1.1.1 Construction of Neighborhood Information Based on Relationship Evaluation

[0085] To obtain neighbor information for cross-lingual knowledge graphs, we use a sparse matrix to store entity and relation information, resulting in a neighbor information matrix. The diagonal information in this matrix represents the ratio of the number of triples to the number of entities, indicating the weight between entities connected to each other. Since the importance of a relation reflects, to some extent, the importance between entities, the weight information between entities is determined by the relation information. The higher the importance of a pair of entities, the less space there is for entity selection in the knowledge graph. For example, a triple (e... i ,r1,e j ) and (e m ,r2,e n ), of which the entities related to r1 are (e h ,e i ,...,e j There are j entities in total, and the entities related to r2 are (e p ,e q ,...,e n Given n entities, where j < n, then for entity ei In general, the range of tail entities available is greater than that of e. m Small, for entity e i The greater the probability of locating a specific entity, the higher the likelihood of locating entity e. i With entity e j The connection between them compared to e m and e n The relationships between entities are also closer. The weight information between entities is obtained by the following formula:

[0086]

[0087] Among them, l t Indicates the number of triples. Representative entity e i and entity e j The relationship between r ij The higher the frequency of occurrence in all triples, the stronger the connection between the two entities. After obtaining the neighborhood information between entities, we combine the neighbor information matrix and the entity semantic character-level vectors, and calculate the initial candidate seed set based on similarity.

[0088] 1.1.2 Comprehensive Semantic-Character Similarity Calculation

[0089] For the source and target cross-language knowledge graphs, the non-English attribute names are first translated into English using a machine translation system. Then, GloVe word embeddings are used to obtain the semantic-level average word vectors of the attribute names, which represent the semantic information of the entity names. To obtain a more comprehensive candidate seed set, each attribute name is split into two-character units to obtain character-level embedding vectors. Finally, the semantic and character-level similarity of the two knowledge graphs are calculated using the cosine similarity function to obtain a preliminary candidate seed set. When calculating the initial similarity, we obtain the Top N nearest distances and corresponding weights of the entities in the two knowledge graphs. These Top N entities can be used for the optimization selection of candidate seeds, and the weight information is used to adjust the corresponding loss function so that the aligned entities can receive more accurate guidance during the training process. Figure 4 The process of generating the initial candidate seed set is described.

[0090] 1.2 Optimization of Candidate Seed Selection

[0091] The weakly supervised algorithm for cross-lingual entity alignment needs initial candidate seeds, which gradually generate more aligned entities in the iterative process of the model. Therefore, it is crucial to optimize the selection of the initial candidate seeds. Good seeds need to meet two conditions: similar neighbor information and similar semantics. Such an optimized selection algorithm selects a seed set that is more reliable during training. Section 1.1 generates an initial candidate seed set from the neighborhood and semantic character information. Based on this, this section designs a candidate seed optimization selection algorithm from the perspective of semantic-character two-level distance measurement.

[0092] The preliminary candidate seeds generated from entity neighborhood information and attribute names still have many noises. In order to obtain a more high-quality candidate seed set, we designed an optimization selection strategy based on semantic-character distance measurement. Specifically, the algorithm first obtains entity pairs that are the closest distance to each other, which are calculated from entity neighborhood and semantic information; then the algorithm extracts the semantic-level vector and character-level vector of these entity pairs, respectively, and calculates the similarity of the two vectors to obtain the semantic and character-level confidence between entities; finally, we integrate the two confidences to evaluate the candidate seeds; Figure 5 The overall idea of distance measurement is shown.

[0093] 1.2.1 Semantic-Character Level Distance Measurement of Entities

[0094] Figure 5 The semantic-character level distance measurement of entities is described. Wherein, u and v are the closest distance to each other, the semantic similarity is calculated as d1; starting from entity u in KG1, find the second closest entity v' of u in KG2, calculate its semantic similarity as d2; finally, calculate the difference D1 between d1 and d2; then start from entity v in KG2, get the difference D2 of semantic similarity; the average of the two differences is the overall semantic similarity (Ds). The similarity at the character level is similar to the calculation method of the semantic level, and the overall character level similarity (Dc) can be calculated. The combination of semantic level similarity (Ds) and character level similarity (Dc) can obtain the weight of the entity pair. The greater the weight, the greater the possibility that the two entity pairs are aligned entities.

[0095] The algorithm first finds the two closest distances v and v' of entity u in the source knowledge graph in the target knowledge graph, and calculates the distance between entities u and v and v', respectively. Here we use cosine similarity to calculate:

[0096] D1 = |Cos(u, v) - Cos(u, v')| (2)

[0097] Then, find the two closest distances of entity u and u' of the target knowledge graph in the source knowledge graph, and calculate the distance between target entity v and u and u', respectively:

[0098] D2 = |Cos(v, u) - Cos(v, u')| (3)

[0099] Finally, the average of the two distances is calculated, which is the angle of integrating the two knowledge graphs to get the entity semantic level similarity.

[0100] D s = (D1 + D2) / 2 (4)

[0101] Next we get the character level distance measure. Here we use Jaccard similarity to calculate the character level similarity. Similarly, the distance between the source entity u and the target entity v and v' is calculated:

[0102] D3 = |Jaccard(u, v) - Jaccard(u, v')| (5)

[0103] Then, find the two nearest distance entities u and u' of the target knowledge graph in the source knowledge graph, respectively calculate the Jaccard distance of the target entity v and u and u':

[0104] D4 = |Jaccard(v, u) - Jaccard(v, u')| (6)

[0105] Integrate the above node distance measure results, calculate the average of the two distances:

[0106] D c = (D3 + D4) / 2 (7)

[0107] Thus we get the semantic and character level measure results Ds and Dc of the two knowledge graphs respectively.

[0108] 1.2.2 Candidate seed node evaluation

[0109] The value of Ds and Dc takes into account the multi-angle information, and the most used in the whole process is the calculation of similarity, without introducing more hyperparameters. In this way, we use fewer parameters to complete the design of the optimization selection algorithm. Finally, the evaluation value of the candidate seed is obtained by formula (8).

[0110] D final = D s + D c (8)

[0111] The optimization selection algorithm can select better candidate seeds because it matches the candidate seeds that integrate the neighborhood and semantic character multi-angle distance measurement information, and the results obtained are optimal for each other, and the distance between the matched entity pairs needs to be lower than a certain preferred factor. Through similarity calculation, we limit the final result to [0, 1], and the larger the value, the higher the reliability of the aligned entity.

[0112] 1.3 Joint training of candidate seeds and labeled data

[0113] Processing all cross-language entities into entities of the same language can only process a small amount of data, and this method is impractical when facing massive data. Fortunately, aligned entities generally have similar neighborhood information, so we can start from the structure information to mine cross-language entity embedding, and perform alignment work through these vectors processed by the neural network. The Dual-AMN entity alignment method selects part of the labeled data to be input into the model for training, which relies too much on labeled data. The method of the present application increases the candidate seed set on the basis of the labeled data. These candidate seed sets are stored in set U, and the labeled data is stored in set L. We need to select w1 proportion of data in set U and w2 proportion of data in set L. These selected data need to satisfy formula (9) and cannot be repeated. Next, we take different values of w1 and w2 in the two sets to select different proportions of candidate seeds and labeled data, and finally input the selected two kinds of data into the WSEO method for joint training.

[0114] U*w1+L*w2=L*n (9)

[0115] wherein the ranges of the parameters need to satisfy the following formula:

[0116] w1∈[0,n],w2∈[0,n],n∈[0,1],w1+w2=n (10)

[0117] 1.3.1 Relationship-aware attention layer

[0118] The input of the model is two matrices, H e ∈R |E|×d represents the initial entity vector, H r ∈R |R|×d represents the initial relationship feature. The model uses a relationship attention mechanism to aggregate the neighbor information of the entity. The entity e i The output vector of the lth layer can be obtained by the following formula:

[0119]

[0120] In the above formula, tanh is an activation function, and the relationship projection operation generates the relationship vector of each entity, and this operation does not produce additional parameters, where a ijk The calculation is as follows:

[0121]

[0122] where v T is the attention vector, and the softmax operation selects the key path connected with the entity from various types of relationships. In previous studies, GNN can expand multi-hop neighbor information by stacking multiple layers, so the model splices the embedding information of different layers to obtain the multi-hop embedding of entity e i :

[0123]

[0124] In the above formula, || represents the splicing operation. In the test process, the method model uses CSLS [Conneau A, Lample G, Ranzato M A, et al. Word translation without parallel data [J]. arXiv preprint arXiv: 1710.04087, 2017.] to calculate the distance between entity pairs, and the smaller the distance is, the greater the possibility of aligning entities is.

[0125] 1.3.2 Proxy vector attention layer

[0126] The simplified relationship attention layer obtains the embedding vector of a single knowledge graph, and the implementation of entity alignment needs to be completed in a unified vector space, so the model uses the vector matching attention layer. We use limited proxy vectors to represent the alignment relationship information across graphs, and if two entities are equivalent, then their similarity distribution with these proxy vectors should be consistent. In this way, this layer can capture the alignment information across graphs without calculating the interaction between nodes and nodes. The interaction of the proxy vector attention layer is to calculate the similarity between all entities and limited proxy vectors, and this operation is similar to clustering.

[0127] The input of the proxy vector attention layer is two matrices: H e ∈R |E|×ld represents the entity embedding information captured by the simplified relationship attention layer, and Q∈R n×ld represents the proxy vector after random initialization, where n represents the number of proxy vectors. In order to obtain the entity embedding information of the cross graph, the first step is to calculate the similarity between each entity and all proxy vectors:

[0128]

[0129] S pdenotes the set of agent vectors. Here we use cosine similarity to compute the similarity between embedding vectors, entity e i The cross graph vector of entity e

[0130]

[0131] is described as the difference between all agent vectors. Finally, we use gating mechanisms to merge and to control the information flow between individual graphs and cross graphs:

[0132]

[0133]

[0134] M and b denote the weight matrix and gate bias matrix, respectively.

[0135] 1.4 Seed credibility sensitive loss function optimization

[0136] The candidate seeds selected by the optimization selection algorithm have different degrees of credibility, and these credibility degrees have different degrees of influence on the model training as weight information between entity pairs. Therefore, the loss function is modified during the training process of the model, that is, each entity pair (u, v) is assigned weight information D final (u, v) to make the training results of the model more reasonable. In addition to the weight information obtained by the optimization selection algorithm, we have the same reason to believe that the negative samples related to these entities have the same credibility score, and in addition, we set the weight score of the labeled data to 1. We reduce the dependence of the model on sample size and hyperparameters by fixing the sample loss of mean and variance.

[0137]

[0138] where P denotes the positive sample and e' denotes the entity in the negative sample.

[0139] ln(e i ,e j ,e' j ) denotes the normalized loss of the triple (e i ,e j ,e' j ), and τ and λ 2 represent the mean and variance of the normalized loss, respectively. Where ln(e i ,e j ,e' j ) is defined as:

[0140]

[0141] l o (e i ,e j ,e' j )=γ+sim(e i ,e j )-sim(e i ,e' j ) (20)

[0142] In the above formula, l o (e i ,e j ,e' j ) represents the initialization triplet (e i ,e j ,e' j ) loss, and μ and σ 2 represent the mean and variance of the initial loss, which are calculated as follows:

[0143]

[0144]

[0145] The calculation process of ln(e j ,e i ,e i ') is similar to that of ln(e i ,e j ,e' j ).

[0146] In the training process, we select different proportions of candidate seed sets as supplementary data, reserve a small part of labeled data, and use L2 distance to calculate the similarity between entities.

[0147] To verify the effect of the application, the following experiments are carried out

[0148] 2、Experiment

[0149] The application uses the Keras framework to develop our model. The experiment is run on a workstation with an NVIDIA TITAN RTX GPU and 128GB of memory.

[0150] 2.1 Experimental setup

[0151] 2.1.1 Dataset

[0152] In order to verify the effectiveness of the model, we choose three typical cross-language knowledge graphs in the entity alignment field to verify the weakly supervised entity alignment work, and the following three data sets are described in detail.

[0153] DBP15K: contains four specific language knowledge graphs, English (En), Chinese (Zh), French (Fr) and Japanese (Ja), extracts entities from the four knowledge graphs in DBpedia, and each knowledge graph contains about 65-106k entities. The dataset includes three cross-language sets constructed from DBpedia: English-French (DBP EN-FR ), English-Chinese (DBP EN-ZH ), English-Japanese (DBP EN-JA ), each set contains 15,000 pairs of pre-aligned entity pairs for training and testing.

[0154] Table 2 lists the statistical data of the dataset. In this research work, we use 0-30% of the labeled data as training data, and train the candidate seeds in combination with these data, and the data is the test data.

[0155] Table 2 DBP15K dataset

[0156]

[0157] 2.1.2 Benchmark method and evaluation index

[0158] Metrics. In order to make a fair comparison with previous work, we use the same evaluation criteria. We use Hit@k and Mean Inverse Rank (MRR) as evaluation indexes. Hit@k score is determined by the proportion of correctly aligned entity pairs in the alignment result. In particular, Hit@1 represents accuracy. In order to be more convincing, we calculate the average score of the five rounds after the experiment is conducted for five rounds.

[0159] Experimental setup. For all datasets, we use the same setting. Embedding dimension d = 100, GNN layer l = 2; agent vector number n = 64, edge γ = 1; new mean and variance in normalization loss are set to τ = 10, λ = 30; training batch size is 1024, learning rate is set to 0.005; semantic and character balance factor is set to β = 0.5, and the preferred factor is set to δ = 0.3.

[0160] Baseline method. In order to further verify the effectiveness of the weakly supervised algorithm based on the invention, we compare the algorithm with several typical supervised algorithms, which are described as follows:

[0161] BootEA is a typical semi-supervised entity alignment method, which selects entities with higher similarity as aligned entities in each iteration.

[0162] NAEA [Zhu Q, Zhou X, Wu J, et al. Neighborhood-Aware Attentional Representation for Multilingual Knowledge Graphs[C] / / IJCAI.2019:1943-1949.] combines hierarchical information from neighborhood subgraphs and weighted aggregates of neighbor representations to learn entity embeddings.

[0163] TransEdge uses edges as its core to link entities and relationships together as a relationship embedding, used to handle one-to-many and many-to-one relationships.

[0164] MRAEA [Mao X, Wang W, Xu H, et al. MRAEA: an efficient and robust entity alignment approach for cross-lingual knowledge graph [C] / / Proceedings of the 13th International Conference on Web Search and Data Mining. 2020: 420-428.] is a supervised algorithm that considers neighbors and their relationships. It proposes a bidirectional iterative strategy to add new alignment seeds during training.

[0165] GCN-Align combines entity structure and attribute information to learn cross-language embeddings. It uses GCN to embed each language into a unified vector space, which is a typical supervised method using graph convolutional neural networks.

[0166] KECG [Li C, Cao Y, Hou L, et al. Semi-supervised entity alignment via joint knowledge embedding model and cross-graph model[C] / / Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing(EMNLP-IJCNLP).2019:2723-2732.] A semi-supervised entity alignment method using a joint knowledge embedding model and a cross-graph model.

[0167] AliNet introduces distant neighbors to expand the overlapping parts of the neighbor structure and uses an attention mechanism to help distant neighbors reduce noise. It combines information from direct and distant neighbors to complete the entity alignment task.

[0168] 2.2 Experimental Results

[0169] In Table 3, we compare the experimental performance of the proposed method with the Dual-AMN method on three cross-lingual datasets. WSEO uses training data set according to equation (9), where L:U represents the ratio of labeled data to candidate seeds. The bolded portion in the table below indicates the percentage performance improvement of WSEO compared to the Dual-AMN method when the labeled data is only 10%.

[0170] Table 3 shows the experimental performance after adding supplementary data.

[0171]

[0172] As shown in the table above, WSEO's performance on the three metrics improves with the addition of candidate seeds. When the L:U ratio is 1:2, WSEO's performance on Hit%1, Hit@10, and MRR is close to that of Dual-AMN, with the difference between WSEO and Dual-AMN in Hit@1 being approximately 1%. This indicates that the method of this invention can achieve experimental performance comparable to Dual-AMN using only 10% of the labeled data. WSEO can reduce the use of labeled data by 20%, fully realizing the weakly supervised entity alignment task for cross-language knowledge graphs. The ZH-EN dataset in the table shows poor performance in the WSEO model because, under the influence of equal selection factors, there is insufficient data for screening, leading to the use of all data as supplementary data. It is worth noting that when the L:U ratio is 0:3, WSEO uses 30% of the candidate seeds as training data, transforming the method of this invention into an unsupervised entity alignment task. Figure 6 Experimental performance of WSEO and Dual-AMN on the JA-EN, FR-EN, and ZH-EN datasets using identically labeled data is presented.

[0173] Depend on Figure 6 It can be seen that WSEO outperforms the Dual-AMN entity alignment method on the three datasets, which fully demonstrates that the weakly supervised entity alignment algorithm for cross-language knowledge graphs proposed in this invention can effectively reduce the use of labeled data and has stable performance.

[0174] To further verify the superiority of the method in performance, we compare the method with the existing supervised method. Table 4 describes the performance of several typical supervised entity alignment methods in Hit@1, Hit@10 and MRR three indicators. Here, the value of L:U is set to 1:2, and the bold font indicates the best performance data among all methods.

[0175] Table 4 Experimental performance of different entity alignment methods

[0176]

[0177] The experimental results show that our method can achieve superior performance on JA-EN and FR-EN datasets with only 10% labeled data. In terms of Hit@1, Hit@10 and MRR three evaluation indicators, the numerical performance of WSEO is better than the baseline method, in which the performance of WSEO in Hit@1 value is increased by 3%-45.3%, and MRR is increased by 2%-34.4%. Among the above indicators, the Hit@1 value directly reflects the accuracy of entity alignment, so the excellent results of the method in Hit@1 value further prove the effectiveness of the model. The poor performance of WSEO on the ZH-EN dataset is because the ZH-EN dataset does not have enough candidate seeds to supplement the labeled data under the same parameter condition, which leads WSEO to use all candidate seeds as supplementary data. When L:U is 2:1, the ZH-EN dataset has enough candidate seeds for our optimization selection algorithm to evaluate, and at this time the performance of WSEO on the ZH-EN dataset is also better than all baselines.

[0178] The relatively weak performance of GCN-Align in the above table is because it only uses structural information to generate alignment seeds, and TranEdge performs second in all methods because it uses the diversity of relations, which verifies the importance of relationship information in entity alignment. MRAEA performs best among all baseline methods because it uses meta-relation and neighbor information and adds new alignment seeds during training. Compared with the above three methods, the performance of WSEO method in Hit@1 is increased by 3%, 4.6% and 2.2% respectively, which fully shows that the method proposed in the application not only can reduce the use of labeled data, but also is superior in performance.

[0179] 2.3 Ablation experiment

[0180] To verify the effectiveness of the candidate seed optimization selection algorithm, we conduct an ablation experiment on DBP15k dataset. The performance of the candidate seed optimization selection algorithm is compared with the performance of the method without the candidate seed optimization selection algorithm. The results are shown in Table 5. JA-ENAn ablation experiment is performed on the data set. Taking the JA-EN data set as an example, there are 4279 candidate seeds participating in the evaluation of the optimization selection, and 3000 pairs of candidate seeds are selected as supplementary data, 1500 pairs of marked data are selected accordingly, the value of L:U is 1:2, and the two kinds of data are sent into the model for joint training, so that the experimental results of the optimization selection algorithm (NSS) are obtained. Then, the model without the optimization selection algorithm (ASN) is experimented, in order to compare under the same conditions, 4279 initial candidate seeds with high weights in the TopN file are selected as the selection space of the supplementary data, at this time these data do not participate in the evaluation of the optimization selection algorithm, 3000 pairs of entities are randomly selected from the 4279 data as supplementary data, and 1500 pairs of marked data are selected accordingly, and the value of L:U reaches 1:2. Table 5 lists the experimental results of the JA-EN data set before and after the candidate seed optimization selection.

[0181] Table 5 Performance of JA-EN data set on the model after removing the candidate seed optimization selection algorithm

[0182]

[0183] In Table 5, we can see that after removing the optimization selection algorithm of WSEO, the performance of JA-EN data set on four indicators is obviously reduced. When the model of the application does not remove the optimization selection algorithm of the candidate seed, WSEO can reach 78.9% of the accuracy rate relying on 10% of the marked data, and the values of Hit@10 and MRR are 93.8% and 84.4% respectively; and after removing the optimization selection algorithm of the model, WSEO can only reach 67.6% of the accuracy rate on the JA-EN data set, the value of Hit@1 is reduced by about 11%, and the values of Hit@10 and MRR are reduced by 5.6% and 9.5% respectively. The above experimental results show that the optimization selection algorithm can effectively select candidate seeds with higher reliability, and ultimately greatly affect the accuracy of entity alignment.

[0184] On the basis of the above embodiment, as Figure 7 indicated, the application further provides a weakly supervised entity alignment optimization system for a cross-language knowledge graph, comprising:

[0185] A preliminary candidate seed set derivation module is used to extract the neighborhood relationship and semantic character information of the entity in the cross-language knowledge graph, and a similarity calculation is used to obtain a preliminary candidate seed set;

[0186] A candidate seed optimization selection module is used to comprehensively evaluate the reliability of the candidate seed set according to the semantic and character information, and find candidate seeds with higher reliability;

[0187] a joint training module configured to jointly train the selected candidate seeds and the partial labeled data;

[0188] a loss function improvement module configured to improve the loss function by using the obtained confidence, and finally optimize the whole training process.

[0189] Further, the preliminary candidate seed set obtaining module is specifically configured to:

[0190] Sparse matrix is used to store the entity and relationship information in the knowledge graph, and a neighbor information matrix is obtained, wherein diagonal information in the matrix is a ratio of the number of all triples to the number of entities, representing the weight between entities connected to themselves.

[0191] The weight information between entities is calculated according to the following formula:

[0192]

[0193] wherein, l t represents the number of triples, represents the relationship r i between the entity e j and the entity e ij The higher the weight, the closer the connection between the two entities.

[0194] For the source and target cross-language knowledge graphs, first, the attribute names in non-English are translated into English using a machine translation system; then, the semantic-level average word vectors of the attribute names are obtained using Glove word embedding, and the average word vectors are used to represent the semantic information of the entity names; for each attribute name, the attribute name is segmented in units of two characters to obtain character-level embedding vectors; finally, the obtained semantic and character-level vectors are used to calculate the semantic and character-level similarities of the two knowledge graphs respectively using a cosine similarity function, and a preliminary candidate seed set is obtained.

[0195] Further, the candidate seed optimization selection module is specifically configured to:

[0196] First, based on the obtained candidate seed set, entity pairs that are closest to each other are obtained, the entity pairs being calculated based on entity neighborhood and semantic information; then, semantic-level vectors and character-level vectors of the entity pairs are extracted respectively, and similarity calculations are performed on the two types of vectors to obtain semantic and character-level confidence between entities; finally, the two types of confidence are integrated for candidate seed evaluation.

[0197] Further, the joint training module is specifically configured to:

[0198] The candidate seed set is stored in a set U, the labeled data is stored in a set L, w1 proportion of data is selected in the set U, w2 proportion of data is selected in the set L, the selected data satisfies formula (9), and there is no repetition:

[0199] U*w1+L*w2=L*n (9)

[0200] Wherein, each parameter range needs to satisfy the following formula:

[0201] w1 [0,n], w2 [0,n], n [0,1], w1+w2=n (10)

[0202] Next, different values are taken for w1 and w2 to select different proportions of candidate seeds and labeled data, and finally joint training is performed based on the two selected data.

[0203] Further, the loss function improvement module is specifically used for:

[0204] In the training process, each pair of entities (u, v) in the candidate seed selected for optimization is assigned weight information D final (u, v), the weight score of the labeled data is set to 1, the sample loss with fixed mean and variance is used to reduce the dependence on sample size and hyperparameters, so as to improve the loss function and finally optimize the entire training process.

[0205] In summary, the present application analyzes the possibility of interaction between labeled data and non-labeled data from the deficiencies in the existing entity alignment method, and proposes a weakly supervised entity alignment optimization method and system for cross-language knowledge graph. The present application uses the candidate seed set generated by the unsupervised method as supplementary data, selects candidate seeds with higher reliability by combining semantic and character level distance measurement, then inputs the candidate seeds and labeled data into the model for joint training, and uses a loss function based on credibility weight to optimize the entire training process. The experimental results show that WSEO can achieve experimental results comparable to Dual-AMN method with only 10% labeled data, and compared with all baseline methods, the performance of the present application method in Hit@1 value is at least improved by 3%.

[0206] The above only shows the preferred embodiments of the present application, it should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A weakly supervised entity alignment optimization method for cross-language knowledge graphs, characterized in that, The method comprises the following steps: Step 1: extracting neighborhood relationship and semantic character information of entities in cross-language knowledge graph, and obtaining a preliminary candidate seed set by similarity calculation; Step 2: comprehensively evaluating the reliability of the candidate seed set based on semantic and character information, and finding candidate seeds with higher reliability; Step 3: jointly training the candidate seeds selected by optimization and part of the labeled data; Step 4: improving the loss function by using the obtained reliability, and finally optimizing the entire training process; The step 1 comprises: The sparse matrix is used to store the entity and relationship information in the knowledge graph to obtain the neighbor information matrix, wherein the diagonal information in the matrix takes the ratio of the number of all triples to the number of entities, representing the weight between entities connected to themselves; The weight information between entities is calculated according to the following formula: （1） wherein, represents the number of triples, represent entities and the relationship between entities The higher the weight, the more frequent the occurrence in all triples, indicating a closer connection between the two entities. For the source and target cross-language knowledge graphs, first, the attribute names in non-English are translated into English using a machine translation system; then the semantic average word vector of the attribute name is obtained by using the Glove word embedding, and the average word vector is used to represent the semantic information of the entity name; for each attribute name, the attribute name is divided into two characters to obtain the character-level embedding vector; finally, the semantic and character-level vectors obtained are used to calculate the semantic and character-level similarity of the two knowledge graphs respectively by using the cosine similarity function, and a preliminary candidate seed set is obtained; The step 4 comprises: In the training process, each pair of entities in the candidate seeds selected for optimization Assigning weight information The weight score of the labeled data is set to 1, the dependence on sample size and hyperparameters is reduced by fixing the sample loss of mean and variance, the improvement of the loss function is completed, and the whole training process is finally optimized. 2.The cross-lingual knowledge graph oriented weakly supervised entity alignment optimization method according to claim 1, characterized in that, The step 2 comprises: First, the obtained candidate seed set is used to obtain entity pairs that are closest to each other, and the entity pairs are calculated based on the neighborhood and semantic information of the entities; then the semantic and character-level vectors of the entity pairs are extracted, and the similarity of the two vectors is calculated to obtain the semantic and character-level reliability between the entities; finally, the two kinds of reliability are comprehensively evaluated to evaluate the candidate seeds. 3.The cross-lingual knowledge graph oriented weakly supervised entity alignment optimization method according to claim 1, characterized in that, The step 3 comprises: The candidate seed set is stored in set U, and the labeled data is stored in set L. The data in set U is selected in proportion, and the data in set L is selected in proportion. The selected data satisfies formula (9) and cannot be repeated. (9) Wherein, each parameter range needs to satisfy the following formula: (10) Next, we take and different values to select different proportions of candidate seeds and labeled data, and finally perform joint training based on the selected two types of data.

4. A weakly supervised entity alignment optimization system for cross-lingual knowledge graph, characterized in that, The method comprises the following steps: The preliminary candidate seed set obtaining module is configured to extract neighborhood relationship and semantic character information of entities in cross-language knowledge graph, and obtain a preliminary candidate seed set by similarity calculation; The candidate seed optimization selection module is configured to comprehensively evaluate the reliability of the candidate seed set based on semantic and character information, and find candidate seeds with higher reliability; The joint training module is configured to jointly train the candidate seeds selected by optimization and part of the labeled data; The loss function improvement module is configured to improve the loss function by using the obtained reliability, and finally optimize the entire training process; The preliminary candidate seed set obtaining module is specifically configured to: The sparse matrix is used to store the entity and relationship information in the knowledge graph to obtain the neighbor information matrix, wherein the diagonal information in the matrix takes the ratio of the number of all triples to the number of entities, representing the weight between entities connected to themselves; The weight information between entities is calculated according to the following formula: （1） wherein, represents the number of triples, represent entities and entities between which the relationship occurs, the higher the weight, the closer the connection between the two entities. For the source and target cross-language knowledge graphs, firstly, the attribute names in non-English are translated into English using a machine translation system; then the semantic-level average word vectors of the attribute names are obtained using Glove word embedding, and the average word vectors are used to represent the semantic information of the entity names; for each attribute name, the attribute name is segmented in two-character units to obtain character-level embedding vectors; finally, the semantic and character-level vectors obtained are used to calculate the semantic and character-level similarities of the two knowledge graphs respectively using a cosine similarity function, to obtain a preliminary candidate seed set; The loss function improvement module is specifically configured to: In the training process, each pair of entities in the candidate seeds selected for optimization Assigning weight information The weight score of the labeled data is set to 1, the dependence on sample size and hyperparameters is reduced by fixing the sample loss of mean and variance, the improvement of the loss function is completed, and the whole training process is finally optimized. 5.The cross-lingual knowledge graph oriented weakly supervised entity alignment optimization system of claim 4, characterized in that, The candidate seed optimization selection module is specifically configured to: Firstly, entity pairs that are closest to each other are obtained based on the obtained candidate seed set, the entity pairs being calculated based on entity neighborhoods and semantic information; then semantic-level vectors and character-level vectors of the entity pairs are extracted respectively, similarity calculation is performed on the two kinds of vectors, and semantic and character-level reliabilities between entities are obtained; finally, the two kinds of reliabilities are integrated for candidate seed evaluation. 6.The cross-lingual knowledge graph oriented weakly supervised entity alignment optimization system of claim 4, characterized in that, The joint training module is specifically configured to: The candidate seed set is stored in set U, and the labeled data is stored in set L. The data in set U is selected in proportion, and the data in set L is selected in proportion. The selected data satisfies formula (9) and cannot be repeated. (9) Wherein, each parameter range needs to satisfy the following formula: (10) Next, we take and different values to select different proportions of candidate seeds and labeled data, and finally perform joint training based on the selected two types of data.

Images