Knowledge graph information domain quality evaluation method and system

Abstract

The invention discloses a knowledge graph information domain quality evaluation method and system, and the method comprises the steps: collecting knowledge in a knowledge graph, and obtaining a data set; the data sets are classified according to 119 relations, and positive and negative example triples containing the same relation are taken out; a Li-Z-C model is established; inputting the positive and negative example triple containing the same relationship into the Li-Z-C model to obtain an embedded vector of each entity and relationship, and performing link prediction of the knowledge graph by using the embedded vector of each entity and relationship; and classifying a link prediction result of the knowledge graph by using a scoring function according to a relationship threshold to obtain a triple accuracy rate. The invention aims to quantify the quality of knowledge in the knowledge graph. And according to a quantitative evaluation result, the knowledge with relatively high confidence is reserved, and the knowledge with relatively low confidence is abandoned, so that the reliability of the knowledge in the knowledge graph is effectively ensured.

Description

Technical Field

[0001] The present invention belongs to the field of knowledge graph quality assessment, and specifically relates to a knowledge graph information domain quality assessment method and system. Background Art

[0002] With the rapid development of information technology, a large number of reconnaissance, detection and sensing equipment have been widely used, providing humans with massive, multi-source, complex and heterogeneous situational data. However, the timeliness and accuracy of the data are still key issues that need to be urgently solved in the field of situational intelligent cognition.

[0003] Knowledge graph technology can achieve relatively intelligent situation recognition to help people make decisions efficiently and scientifically. Knowledge graph technology can discover patterns and trends in data to better understand the evolution of the situation. At the same time, data visualization and interactive technology are used to present complex data to help people understand and analyze data more intuitively.

[0004] By combining human experience and intuition with the computing power of machines, the efficiency of situation analysis and decision-making can be improved. Natural language processing can be used to achieve human-machine collaboration to jointly analyze situations and make decisions. In order to improve the utility of knowledge graphs, researchers have proposed various refinement methods to try to infer and add missing knowledge to the graph, or identify incorrect information fragments.

[0005] After constructing and completing the knowledge graph, evaluating the quality of the generated knowledge graph becomes a key step. The quality of the knowledge graph triples is the quality of the information domain information. At the application level of the knowledge graph, the quality of the knowledge graph largely determines its utility in specific application scenarios.

[0006] At present, there are several possible problems with knowledge graph triples: 1. Data accuracy. Due to the complexity of knowledge, data may be inaccurately classified, affecting the accuracy and reliability of the knowledge graph. 2. Timeliness of knowledge graph updates: With the continuous updating and development of knowledge, knowledge graphs need to be updated in a timely manner to maintain accuracy and reliability in order to provide real-time situation intelligence. However, in the current environment, there are challenges in the timeliness of data. Summary of the invention

[0007] In view of the shortcomings of the prior art, the present invention proposes a knowledge graph information domain quality assessment method and system, which aims to quantify the quality of knowledge in the knowledge graph. According to the results of the quantitative assessment, the knowledge with higher confidence is retained and the knowledge with lower confidence is discarded, thereby effectively ensuring the reliability of the knowledge in the knowledge graph.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A knowledge graph information domain quality assessment method comprises the following steps:

[0010] Collect knowledge from the knowledge graph and obtain a data set;

[0011] The dataset is classified into 119 types of relations, and the positive and negative example triplets containing the same type of relations are extracted;

[0012] Establish Li-ZC model;

[0013] Input the positive and negative example triplets containing the same kind of relationship into the Li-ZC model to obtain the embedding vector of each entity and relationship. Use the embedding vector of each entity and relationship to perform link prediction on the knowledge graph.

[0014] The link prediction results of the knowledge graph are classified according to the relationship threshold using a scoring function to obtain the triple accuracy.

[0015] Preferably, the method of classifying the data set according to 119 kinds of relations and extracting positive and negative example triplets containing the same kind of relations includes:

[0016] Randomly extract batch size triplets from the dataset as a set for loop iteration;

[0017] Replace the head entity or tail entity of each triple in the set to generate a negative sample; when constructing a negative sample, the average number of head nodes and the average number of tail nodes are calculated. When the average number of head nodes is greater than the average number of tail nodes, the head entity tends to be replaced, otherwise the tail entity is replaced;

[0018] The extracted positive example triplets and the generated negative example triplets are combined to form a single training set.

[0019] Preferably, the training set is calculated using a loss function, and the specific formula is:

[0020]

[0021] Preferably, the positive and negative example triples containing the same kind of relationship are input into the Li-ZC model to obtain the embedding vector of each entity and relationship, and the method for link prediction of the knowledge graph using the embedding vector of each entity and relationship includes:

[0022] Select L1 norm or L2 norm to calculate the triple distance; if L2 norm is selected, the derivation method is: first derive the L2 norm, judge the positive and negative of each element, assign 1 if it is positive, and assign -1 if it is negative; the specific formula is:

[0023]

[0024] Preferably, the method of using the scoring function to classify according to the relationship threshold includes: by translating the sigmoid function, when x=0, the value of y is the threshold of the relationship.

[0025] The present invention also provides a knowledge graph information domain quality assessment system, including: an acquisition module, an extraction module, a construction module, a prediction module and a classification module;

[0026] The acquisition module is used to acquire knowledge in the knowledge graph and obtain a data set;

[0027] The extraction module is used to classify the data set according to 119 types of relations and extract positive and negative example triplets containing the same type of relations;

[0028] The building blocks are used to establish the Li-ZC model;

[0029] The prediction module is used to input positive and negative example triplets containing the same kind of relationship into the Li-ZC model to obtain the embedding vector of each entity and relationship, and use the embedding vector of each entity and relationship to perform link prediction on the knowledge graph;

[0030] The classification module is used to classify the link prediction results of the knowledge graph according to the relationship threshold using a scoring function to obtain the triple accuracy.

[0031] Preferably, the extraction module comprises: a random extraction unit, a replacement unit and a combination unit;

[0032] The random extraction unit is used to randomly extract batch size triplets from the data set as a set for loop iteration;

[0033] The replacement unit is used to replace the head entity or the tail entity of each triple in the set to generate a negative sample; wherein, when constructing the negative sample, the average number of head nodes and the average number of tail nodes are calculated, and when the average number of head nodes is greater than the average number of tail nodes, the head entity tends to be replaced, otherwise the tail entity is replaced;

[0034] The combination unit is used to combine the extracted positive example triples and the generated negative example triples into a single training set.

[0035] Preferably, the training set is calculated using a loss function, and the specific formula is:

[0036]

[0037] Preferably, the positive and negative example triplets containing the same kind of relationship are input into the Li-ZC model to obtain the embedding vector of each entity and relationship. The process of link prediction of the knowledge graph using the embedding vector of each entity and relationship includes:

[0038] Select L1 norm or L2 norm to calculate the triple distance; if L2 norm is selected, the derivation method is: first derive the L2 norm, judge the positive and negative of each element, assign 1 if it is positive, and assign -1 if it is negative; the specific formula is:

[0039]

[0040] Preferably, the process of using the scoring function to classify according to the relationship threshold includes: translating the sigmoid function so that when x=0, the value of y is the threshold of the relationship.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1) For the first time, a flexible relationship threshold model is proposed to judge the accuracy. The data is processed in an efficient way, and the huge data set is labeled quickly and accurately. By adjusting the threshold, the classification behavior of the model can be quickly controlled, so as to customize the classification requirements according to the actual needs based on the different requirements of different relationship accuracy in the naval battlefield.

[0043] 2) We improved the TransE model, which is good at handling one-to-one relationships and does not need to consider semantic information. We optimized the positive and negative example judgment thresholds during model verification, and classified relationships as objects to achieve better classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0045] Figure 1 A schematic diagram of a knowledge graph information domain quality assessment method according to an embodiment of the present invention;

[0046] Figure 2 Schematic diagram of the loss function of the Li-ZC classification model based on relationship threshold according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Embodiment 1

[0050] In the field of natural language processing (NLP) and information extraction, entities and relations are two core concepts.

[0051] Entity: An entity refers to a specific "thing" in a text, which can usually be a person's name, place, organization, date, time, event, etc. For example, in the sentence "Zhang San works in Beijing", "Zhang San" and "Beijing" are both entities. Generally speaking, an entity can be a noun or noun phrase with independent meaning, which refers to a specific object or concept in a sentence or paragraph.

[0052] Relationship: Relationship is information that describes the association between entities. It indicates a specific connection or relationship type between entities. For example, in the sentence "Zhang San works in Beijing", the relationship between "Zhang San" and "Beijing" can be described as "place of work". The goal of relationship extraction is to identify and classify the relationships between two or more entities. Relationships usually come with specific labels, such as "place of residence", "organization", "relative relationship", etc.

[0053] A triple is a structured form of information representation. A triple is usually represented as (entity 1, relationship, entity 2) and is used to represent a relationship between two entities.

[0054] The Li-ZC model uses entity relationship encoding technology to vectorize entities and relationships so that it satisfies the head entity vector + relationship vector = tail entity vector. Its main implementation process is as follows:

[0055] The first step is that the distance between vectors is an indicator used to measure the difference or similarity between two vectors. Different types of distances are suitable for different tasks and application scenarios. We use Euclidean distance to define the distance d to represent the distance between vectors. This is the most commonly used distance metric, which represents the straight-line distance between two points in two-dimensional or multi-dimensional space. Generally, the L1 or L2 norm is used. It is expected that the distance of the correct triplet is as small as possible, while the distance of the wrong triplet is as large as possible.

[0056] The second step is to call the deepcopy function to deep copy the entity and relationship vectors, and extract the vectors corresponding to the entity and relationship IDs respectively;

[0057] The third step is to calculate the score function based on the L1 norm or L2 norm.

[0058] Choose L1 norm or L2 norm to calculate the triple distance. If L2 norm is selected, the derivation method is as follows: first derive the L2 norm, determine the positive and negative of each element, assign 1 if it is positive, and assign -1 if it is negative. The formula is as follows:

[0059]

[0060] The fourth step is to calculate the loss value according to the following formula:

[0061]

[0062] γ is a residual hyperparameter, which acts as a gap correction between a positive triple and a negative triple. The larger the value of γ, the larger the gap between the two triples will be, and the stricter the correction of the word vector will be.

[0063] like Figure 1 As shown, the present invention discloses a knowledge graph information domain quality assessment method, comprising the following steps:

[0064] Collect knowledge from the knowledge graph and obtain a data set;

[0065] The dataset is classified into 119 types of relations, and the positive and negative example triplets containing the same type of relations are extracted;

[0066] Establish Li-ZC model;

[0067] Input the positive and negative example triplets containing the same kind of relationship into the Li-ZC model to obtain the embedding vector of each entity and relationship. Use the embedding vector of each entity and relationship to perform link prediction on the knowledge graph.

[0068] The link prediction results of the knowledge graph are classified according to the relationship threshold using a scoring function to obtain the triple accuracy.

[0069] In this embodiment, the method of classifying the data set according to 119 types of relations and extracting positive and negative example triples containing the same type of relations includes:

[0070] Randomly extract batch size triplets from the dataset as a set for loop iteration;

[0071] Replace the head entity or tail entity of each triple in the set to generate a negative sample; when constructing a negative sample, the average number of head nodes and the average number of tail nodes are calculated. When the average number of head nodes is greater than the average number of tail nodes, the head entity tends to be replaced, otherwise the tail entity is replaced;

[0072] The extracted positive example triplets and the generated negative example triplets are combined to form a single training set.

[0073] In this embodiment, the training set is calculated using the loss function, and the specific formula is:

[0074]

[0075] In this embodiment, positive and negative example triplets containing the same kind of relationship are input into the Li-ZC model to obtain the embedding vector of each entity and relationship. The method for link prediction of the knowledge graph using the embedding vector of each entity and relationship includes:

[0076] Select L1 norm or L2 norm to calculate the distance of triples, sort them by value, and judge those with values ​​greater than the threshold as negative examples, otherwise they are judged as positive examples;

[0077] If the L2 norm is selected, the derivation method is: first derive the L2 norm, determine the positive and negative of each element, assign a value of 1 if it is positive, and assign a value of -1 if it is negative; the specific formula is:

[0078]

[0079] In this embodiment, the method for determining the threshold value adopts the sigmoid function which is widely used in neural networks. This function has the following characteristics: when x approaches negative infinity, y approaches 0; when x approaches positive infinity, y approaches 1; when x=0, y=1 / 2. We translate this function so that when x=0, the value of y is the threshold value of this relationship.

[0080] Compared with the traditional relationship threshold model, which combines all positive and negative triples and takes a total threshold, this new model that determines the threshold by relationship has a stronger semantic background and can better fit the sigmoid function of each relationship. After completing the training of the relationship threshold, the entire data set (including the correct triples and the predicted triples) is calculated through the Li-ZC model, and the correctness is judged according to the threshold to obtain the accuracy.

[0081] The reason why the Li-ZC model proposed in this project is to represent knowledge vectorization is that the model is easy to train, can be extended to large data sets, and meets the requirements for information cognition.

[0082] In this embodiment, the process of model quality assessment is mainly as follows:

[0083] In the first step, the embedding vectors of entity e and relation l are randomly initialized, and the embedding vector of relation l is regularized.

[0084] In the second step, the model enters the loop iteration stage, and the embedding vector of entity e is re-regularized (batch norm) in each loop iteration.

[0085] The third step is to randomly select batch size triplets from the dataset as a set for loop iteration.

[0086] The fourth step is to replace the head entity or the tail entity of each triple in the set to generate a negative sample (but not both at the same time). When constructing negative samples, the average number of head nodes and the average number of tail nodes are calculated. When the average number of head nodes is greater than the average number of tail nodes, the head entity tends to be replaced, otherwise the tail entity is replaced. The extracted positive triples and the generated negative triples together constitute a single training set (Tbatch).

[0087] In the fifth step, the training set is calculated using the loss function. The randomly initialized entity relationship vectors are optimized by stochastic gradient descent (in the minimum batch mode), aiming to minimize the loss function through continuous cycles and obtain a high-dimensional vector representation of entities and relationships.

[0088] The sixth step is to complete knowledge representation. Knowledge representation is to structure the information and knowledge of the real world so that computers can understand and process it. It involves converting concepts, entities, events and their relationships into formal models so that machines can reason, query and learn further. Knowledge representation mainly represents information through structured data and symbol systems. Common methods include logic, rules, semantic networks, knowledge graphs, etc. Verify the correctness of the triples. This project proposes the Li-ZC model for verification. After modeling with the Li-ZC model, the embedding vectors of each entity and relationship are obtained. The embedding vectors are used to predict the links of the knowledge graph, and the triples are recorded as (h, r, t). The Li-ZC model uses the second norm calculation. The smaller the value, the closer the triple is to the correct standard.

[0089] (1) After multiple training cycles, the gradient of the loss function converges to a local stable point. This indicates that after thousands of iterations, the Li-ZC model converges and the generated embedding vector is relatively accurate.

[0090] (2) The accuracy of the vector representation was tested when the dimension embedding_dim was 32, 64, 75, 100, and 150. The accuracy was the highest when the dimension was 100, reaching 82%. The comparison results are shown in Tables 1 and 18. The accuracy of the vector representation was tested when the Margin was 0.5, 1, 2, 5, and 10, respectively. The accuracy was the highest when the Margin was 1, reaching 82%. Figure 2 shown.

[0091] Table 1 Dimension accuracy

[0092]

[0093] Table 2 Margin accuracy

[0094]

[0095] (3) For specific data, this patent constructs a specific data set, the entity elements of which include 8 major categories and 148 minor categories of equipment, involving 88 countries and 5,800 equipment entities.

[0096] (4) The accuracy of the model is shown in Table 3. On the WN11 dataset, the Li-ZC model reaches 75.8%, on the FB13 dataset, the Li-ZC model reaches 70.9%, and on the specific dataset, the Li-ZC model reaches 82.2%, reaching the international leading level in this field.

[0097] Table 3 Model accuracy

[0098] Dataset WN11 FB13 Specific datasets ZC 75.8 70.9 82.2

[0099] Embodiment 2

[0100] The following is an extension of the possibilities and implementation details of the Li-ZC classification model based on relationship thresholds in applications, providing specific application methods in actual scenarios:

[0101] 1. Intelligent search and information retrieval

[0102] The Li-ZC model improves the accuracy of the knowledge graph through efficient relationship threshold judgment and can provide more relevant results for intelligent search engines. Through relationship threshold filtering, the Li-ZC model can quickly identify high-quality knowledge triples and retain these triples in the graph, which has significant application value in the following scenarios:

[0103] Personalized search: The Li-ZC model helps search engines understand the semantic information behind user queries and, combined with the triple structure of the knowledge graph, provides personalized search results that are more in line with user intent.

[0104] Information retrieval in professional fields: In professional fields such as medicine and law, the Li-ZC model can ensure that the information in the knowledge graph is accurate and reliable, thereby providing professional users with efficient query and analysis.

[0105] 2. Recommendation System

[0106] In the recommendation system, the Li-ZC model converts user behavior or preferences into high-dimensional vectors by embedding entities and relationships, and determines the relevance between users and recommended content through relationship thresholds. This method is applicable to the following fields:

[0107] Content recommendation: such as news recommendation, video recommendation, book recommendation, etc. The Li-ZC model can help the system automatically filter out content that is irrelevant to user preferences, recommend content that users may be interested in, and improve user experience.

[0108] Social network recommendation: In social media, the Li-ZC model can be used to identify similarity relationships between users, such as recommending possible friends, groups, etc.

[0109] 3. Real-time update and maintenance of knowledge graph

[0110] In a dynamic data environment, data is constantly updated and knowledge graphs need to be maintained in a timely manner. The Li-ZC model can determine the credibility of new triples through relationship thresholds, quickly update knowledge graphs, and ensure the timeliness of information. This is particularly important in the following areas:

[0111] News and public opinion analysis: The Li-ZC model is used to classify and judge new data in real time, filter out accurate triples, and help media and public affairs management departments quickly grasp the latest developments.

[0112] Finance and stock markets: The Li-ZC model can be used to update knowledge graphs related to stocks, companies, and market dynamics in real time, helping financial practitioners quickly understand changes in market trends and support decision-making.

[0113] 4. Real-time monitoring and risk management

[0114] In real-time monitoring and risk management, the knowledge graph built with the help of the Li-ZC model can be used to analyze complex situation data and identify potential risks or abnormal situations. Specific applications include:

[0115] Network security: In the risk detection of network attacks or data leaks, the Li-ZC model can help build and evaluate security knowledge graphs and mark abnormal behavior patterns.

[0116] Supply Chain Management: Through the relationship threshold model, enterprises can discover potential risk factors in the knowledge graph and quickly identify weaknesses or delays in the supply chain.

[0117] 5. Relationship Extraction and Understanding in Natural Language Processing (NLP)

[0118] In NLP tasks, the Li-ZC model can enhance the system's ability to identify and understand relations in text by vectorizing relations and entities. This is crucial for the following applications:

[0119] Automatic text summarization: The Li-ZC model is used to identify key entities and relations in the text, thereby generating more accurate text summaries.

[0120] Question-answering system: The Li-ZC model helps the question-answering system to quickly retrieve and filter irrelevant information in the knowledge graph based on the relationship threshold to provide more accurate answers.

[0121] 6. Further application expansion

[0122] With the development of big data technology, the Li-ZC model is expected to be applied in more fields, such as:

[0123] Smart home system: By building a relationship map between family members and devices, the Li-ZC model can help the smart home system understand users’ daily habits more accurately and provide personalized services.

[0124] Smart city management: By analyzing various relational data in the city (such as transportation, energy, environment, etc.), the knowledge graph constructed using the Li-ZC model helps optimize urban resource management and improve service efficiency.

[0125] Application Summary

[0126] The Li-ZC classification model based on relation threshold provides an efficient solution for the quality assessment of knowledge graphs due to its flexible relation threshold mechanism, fast data processing capability and wide application scenarios. Its applications in intelligent search, recommendation systems, real-time monitoring and other aspects demonstrate the powerful potential of the model.

[0127] Embodiment 3

[0128] The present invention also provides a knowledge graph information domain quality assessment system, including: an acquisition module, an extraction module, a construction module, a prediction module and a classification module;

[0129] The acquisition module is used to collect knowledge in the knowledge graph and obtain a data set;

[0130] The extraction module is used to classify the data set according to 119 types of relations and extract the positive and negative example triplets containing the same type of relations;

[0131] The building blocks are used to build the Li-ZC model;

[0132] The prediction module is used to input positive and negative example triplets containing the same kind of relationship into the Li-ZC model to obtain the embedding vector of each entity and relationship, and use the embedding vector of each entity and relationship to perform link prediction on the knowledge graph;

[0133] The classification module is used to classify the link prediction results of the knowledge graph according to the relationship threshold using a scoring function to obtain the triple accuracy.

[0134] In this embodiment, the extraction module includes: a random extraction unit, a replacement unit and a combination unit;

[0135] The random extraction unit is used to randomly extract batch size triplets from the data set as a set for loop iteration;

[0136] The replacement unit is used to replace the head entity or tail entity of each triple in the set to generate a negative sample; wherein, when constructing the negative sample, the average number of head nodes and the average number of tail nodes are calculated, and when the average number of head nodes is greater than the average number of tail nodes, the head entity tends to be replaced, otherwise the tail entity is replaced;

[0137] The combination unit is used to combine the extracted positive example triplets and the generated negative example triplets into a single training set.

[0138] In this embodiment, the training set is calculated using the loss function, and the specific formula is:

[0139]

[0140] In this embodiment, the positive and negative example triples containing the same kind of relationship are input into the Li-ZC model to obtain the embedding vector of each entity and relationship. The process of link prediction of the knowledge graph using the embedding vector of each entity and relationship includes:

[0141] Select L1 norm or L2 norm to calculate the triple distance; if L2 norm is selected, the derivation method is: first derive the L2 norm, judge the positive and negative of each element, assign 1 if it is positive, and assign -1 if it is negative; the specific formula is:

[0142]

[0143] In this embodiment, the process of using the scoring function to classify according to the relationship threshold includes: by translating the sigmoid function, when x=0, the value of y is the threshold of the relationship.

[0144] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A knowledge graph information domain quality assessment method, characterized in that: The following steps are involved: Collect knowledge from the knowledge graph and obtain a data set; The dataset is classified into 119 types of relations, and the positive and negative example triplets containing the same type of relations are extracted; Establish Li-ZC model; Input the positive and negative example triplets containing the same kind of relationship into the Li-ZC model to obtain the embedding vector of each entity and relationship. Use the embedding vector of each entity and relationship to perform link prediction on the knowledge graph. The link prediction results of the knowledge graph are classified according to the relationship threshold using a scoring function to obtain the triple accuracy.

2. The knowledge graph information domain quality assessment method according to claim 1 is characterized in that: The method of classifying the data set according to 119 relations and extracting the positive and negative example triplets containing the same relation includes: Randomly extract batch size triplets from the dataset as a set for loop iteration; Replace the head entity or tail entity of each triple in the set to generate a negative sample; when constructing a negative sample, the average number of head nodes and the average number of tail nodes are calculated. When the average number of head nodes is greater than the average number of tail nodes, the head entity tends to be replaced, otherwise the tail entity is replaced; The extracted positive example triplets and the generated negative example triplets are combined to form a single training set.

3. The knowledge graph information domain quality assessment method according to claim 2 is characterized in that: The training set is calculated using the loss function, and the specific formula is:

4. The knowledge graph information domain quality assessment method according to claim 1 is characterized in that: Input the positive and negative example triplets containing the same kind of relationship into the Li-ZC model to obtain the embedding vector of each entity and relationship. The method of link prediction of the knowledge graph using the embedding vector of each entity and relationship includes: Select L1 norm or L2 norm to calculate the triple distance; if L2 norm is selected, the derivation method is: first derive the L2 norm, judge the positive and negative of each element, assign 1 if it is positive, and assign -1 if it is negative; the specific formula is:

5. The knowledge graph information domain quality assessment method according to claim 1 is characterized in that: The method of using the scoring function to classify according to the relationship threshold includes: by translating the sigmoid function, when x=0, the value of y is the threshold of the relationship.

6. A knowledge graph information domain quality assessment system, characterized in that: include: Acquisition module, extraction module, construction module, prediction module and classification module; The acquisition module is used to acquire knowledge in the knowledge graph and obtain a data set; The extraction module is used to classify the data set according to 119 types of relations and extract positive and negative example triplets containing the same type of relations; The building blocks are used to establish a Li-ZC model; The prediction module is used to input positive and negative example triplets containing the same kind of relationship into the Li-ZC model to obtain the embedding vector of each entity and relationship, and use the embedding vector of each entity and relationship to perform link prediction on the knowledge graph; The classification module is used to classify the link prediction results of the knowledge graph according to the relationship threshold using a scoring function to obtain the triple accuracy.

7. The knowledge graph information domain quality assessment system according to claim 6, characterized in that: The extraction module includes: a random extraction unit, a replacement unit and a combination unit; The random extraction unit is used to randomly extract batch size triplets from the data set as a set for loop iteration; The replacement unit is used to replace the head entity or the tail entity of each triple in the set to generate a negative sample; wherein, when constructing the negative sample, the average number of head nodes and the average number of tail nodes are calculated, and when the average number of head nodes is greater than the average number of tail nodes, the head entity tends to be replaced, otherwise the tail entity is replaced; The combination unit is used to combine the extracted positive example triples and the generated negative example triples into a single training set.

8. The knowledge graph information domain quality assessment system according to claim 7, characterized in that: The training set is calculated using the loss function, and the specific formula is:

9. The knowledge graph information domain quality assessment system according to claim 6, characterized in that: Input the positive and negative example triplets containing the same kind of relationship into the Li-ZC model to obtain the embedding vector of each entity and relationship. The process of link prediction of the knowledge graph using the embedding vector of each entity and relationship includes: Select L1 norm or L2 norm to calculate the triple distance; if L2 norm is selected, the derivation method is: first derive the L2 norm, judge the positive and negative of each element, assign 1 if it is positive, and assign -1 if it is negative; the specific formula is:

10. The knowledge graph information domain quality assessment system according to claim 6, characterized in that: The process of using the scoring function to classify according to the relationship threshold includes: by translating the sigmoid function, when x=0, the value of y is the threshold of the relationship.

Images