CN116306923A - Evaluation weight calculation method based on knowledge graph - Google Patents

Abstract

The invention discloses a knowledge graph-based evaluation weight calculation method, which comprises the following steps: establishing an evaluation knowledge graph and a keyword vector library according to the evaluation related data; combining the set evaluation target and the calculated adaptation index in the keyword vector similarity retrieval index library to form an evaluation index system; converting the established evaluation index system into a directed graph, and calculating knowledge reasoning probability to replace random walk probability in the conventional PageRank algorithm through keyword vector similarity among nodes of the directed graph to form an improved PageRank algorithm; and calculating the PageRank value of each node according to the improved PageRank algorithm, and taking the PageRank value as the objective weight of each evaluation index. The invention can fully utilize expert knowledge and other data to discover potential relations among evaluation indexes, automatically calculate the weights of all indexes in an evaluation index system, ensure that the determination of the evaluation weights does not depend on subjective judgment of people, and improve the efficiency and objectivity of an evaluation decision process.

Description

An Evaluation Weight Calculation Method Based on Knowledge Graph

technical field

The invention relates to the technical fields of weight calculation and knowledge map, in particular to a method for calculating evaluation weight based on knowledge map.

Background technique

A knowledge graph is a knowledge base established by collecting data information from various sources, and it is a collection of knowledge. The knowledge map was officially proposed by Google in 2012 to enhance the capabilities of search engines and improve users' search experience. The knowledge graph is essentially a semantic network, which represents the entities in the objective world and the relationship between them in the form of a graph. The nodes in the knowledge graph represent entities, and the edges between nodes represent the relationships between entities. Knowledge map technology mainly includes three parts: knowledge representation, knowledge map construction, and knowledge map application. Among them, knowledge map construction is the most critical link in the formation of knowledge map. At present, knowledge graphs have shown great application value in natural language understanding, big data analysis, intelligent question answering, Internet of Things and other fields, and are an important driving force for the development of artificial intelligence in the future.

Evaluation weight calculation is the process of determining the importance of each index in the evaluation index system. Evaluation weight calculation methods are mainly divided into three categories: subjective evaluation weight calculation methods, objective evaluation weight calculation methods, and subjective and objective combined evaluation weight calculation methods. The calculation methods of subjective evaluation weight include analytic hierarchy process, network analysis method, decision-making laboratory analysis method, etc. The calculation methods of objective evaluation weight include entropy weight method, principal component analysis method, etc. combined use to get the comprehensive weight. Evaluation weight calculation is an important link in the evaluation decision-making process, which has a significant impact on the final evaluation results. Reasonable evaluation weights can enhance the scientificity and rationality of the evaluation process.

Most of the current research adopts the combination of subjective and objective evaluation weight calculation methods, mainly based on subjective evaluation weights, supplemented by objective evaluation weights, that is, emphasizing the role of people in the evaluation decision-making process, and making objective corrections to it. However, most of the current objective evaluation weight calculation methods are based on a single indicator data, looking for its distribution and change rules, so as to carry out weighting, and it is difficult to explore deeper connections between indicators.

Contents of the invention

Purpose of the invention: In order to overcome the deficiencies in the prior art, provide a method for calculating evaluation weights based on knowledge graphs, which can make full use of data such as expert knowledge, discover potential connections between evaluation indicators, and automatically calculate each index in the evaluation index system weight, so that the determination of the evaluation weight does not depend on human subjective judgment, improving the efficiency and objectivity of the evaluation decision-making process.

Technical solution: In order to achieve the above purpose, the present invention provides a method for calculating evaluation weights based on knowledge graphs, including the following steps:

S1: Establish the evaluation knowledge graph and keyword vector database based on the evaluation-related data, use the evaluation knowledge graph as the evaluation index database of the evaluation index system, and use the keyword vector database as the basis for evaluation index retrieval;

S2: Combining the set evaluation target and the calculated adaptation index in the keyword vector similarity search index library to form an evaluation index system;

S3: Convert the established evaluation index system into a directed graph, and calculate the knowledge reasoning probability through the keyword vector similarity between the nodes of the directed graph to replace the random walk probability in the existing PageRank algorithm, forming an improved PageRank algorithm ;

S4: Calculate the PageRank value of each node as the objective weight of each evaluation index according to the improved PageRank algorithm.

Further, the establishment method of the evaluation knowledge map in the step S1 is: according to the evaluation task, collect relevant text data, and use the knowledge map construction technology to establish the evaluation knowledge map;

The establishment method of the keyword vector database is as follows: the word embedding algorithm is used to obtain the word vector database for the collected text data, and then the keywords of each node of the knowledge graph are extracted, and the corresponding word vectors are searched to form the keyword vector database.

Further, the formation of the evaluation index system in the step S2 is specifically as follows: according to the set evaluation target, the nodes with specific tags in the knowledge graph are screened out, and the similarity between them and the keyword vector vectors of the evaluation target is calculated for further retrieval, Then, all the evaluation indicators that meet the conditions and the dependencies among them are summarized to form an evaluation index system.

Further, the evaluation target is set by experts according to the evaluation task;

The calculation method of the keyword vector similarity is:

表示评价目标的关键词词向量，/>

表示知识图谱节点的关键词词向量。in,

Indicates the keyword vector of the evaluation target, />

Keyword vectors representing knowledge graph nodes.

Further, the step S3 is specifically:

A1: Construct a directed graph of the evaluation index system

After completing the construction of the evaluation index system, each evaluation index in the evaluation index system is converted into a node in the definition of the directed graph, and the directional dependencies between the indicators are converted into directed edges in the definition of the directed graph to realize The conversion process from the evaluation index system to the directed graph;

A2: Establish an improved random walk model

Define the directed graph of the evaluation index system:

G=(V,E)

Among them, V and E respectively represent the collection of nodes and directed edges, and each directed edge corresponds to a pair of ordered nodes;

The random walk model is essentially a Markov chain defined on a directed graph, and its Markov property is expressed as the state transition probability from the current node state to the next moment node state is only related to the current node state. In the improved random walk model, nodes no longer transfer to the nodes they point to with equal probability, but transfer with knowledge reasoning probability. The directed graph of the evaluation index system is essentially established based on the knowledge graph, so the nodes in the directed graph, that is, the nodes in the knowledge graph corresponding to the evaluation index, can query the keyword vector of the node through the knowledge graph;

Define the knowledge similarity of nodes v _i and v _j in the directed graph G of the evaluation index system as:

为知识图谱中节点v_i、v_j的排序最高的关键词词向量；in,

is the highest-ranked keyword word vector of nodes v _i and v _j in the knowledge graph;

Assuming that node j in the directed graph has k directed edges connecting out, the state transition probability from node j to each node can be expressed by the knowledge similarity as:

Among them, p _i ' _j is the probability of knowledge reasoning, indicating the probability of transferring from evaluation index j to evaluation index i with knowledge as the measure;

Assuming that there are m nodes in the directed graph G, the state transition probability matrix of the improved random walk model is expressed as:

符合状态转移概率矩阵的性质。Obviously, p _i ' _j ≥ 0,

It conforms to the property of the state transition probability matrix.

Further, the calculation method of the PageRank value in the step S4 is:

The directed graph of the evaluation index system established based on the knowledge map may not necessarily meet the conditions of non-periodicity and strong connectivity, so its Markov chain may not have a steady-state distribution.

由P和P'的线性组合构成新的状态转移概率矩阵，在此基础上得到的马尔科夫链可以证明一定具有平稳分布R，R有如下等式：For the directed graph G of the general evaluation index system, assuming that it has m nodes, the state transition probability matrix based on the improved random walk model is P. If it is required to obtain the steady-state distribution of the Markov chain, then in the directed graph In addition to defining its own improved random walk model on G, it is also necessary to define a complete random walk model, and the element values of its state transition probability matrix P' are

A new state transition probability matrix is formed by the linear combination of P and P', and the Markov chain obtained on this basis can prove to have a stationary distribution R, and R has the following equation:

Among them, the linear combination coefficient d (0≤d≤1) is the damping factor, according to empirical values, E is an m×m all-1 matrix; l is an m-dimensional vector with all components being 1, and R indicates that in general The smooth distribution of R, each component of R is the PageRank value of each node, that is, the objective weight corresponding to each evaluation index:

R＝[PR(v ₁ ),PR(v ₂ ),...,PR(v _m )] ^T ，v ₁ ,v ₂ ,...,v _m ∈V

公式第一项表示状态平稳分布时，模型按照状态概率转移矩阵P游走到各节点的概率分布，且这样的概率分布占权重d；第二项其表示状态平稳分布时，模型按照状态概率转移矩阵P'等概率游走到各节点的概率分布，且这样的概率分布占权重1-d。Among them, PR(v _i )>0,

When the first item of the formula indicates the state is in a stable distribution, the model moves to the probability distribution of each node according to the state probability transition matrix P, and such a probability distribution accounts for the weight d; when the second item indicates that the state is in a stable distribution, the model transfers according to the state probability Matrix P' walks to the probability distribution of each node with equal probability, and such probability distribution accounts for the weight 1-d.

转移到任意节点，保证了没有有向边连出的节点也可以实现状态转移，进而确保马尔科夫链存在平稳分布。Intuitively, the second term is introduced as a smoothing term so that nodes have 1-d probability of completely random transition, with equal probability

Transferring to any node ensures that nodes without directed edges can also achieve state transfer, thereby ensuring a stable distribution of the Markov chain.

Further, the calculation process of the PageRank value in the step S4 includes the following steps:

B1: Calculate the state transition probability matrix P of the directed graph G of the evaluation index system based on the knowledge measurement;

B2: let t=0, select the initial state distribution x ₀ ;

B3: Calculate the general state transition probability matrix A of the directed graph

B4: Iterate and normalize the result vector

y _t+1 =Ax _t

B5: When ||x _t+1 -x _t ||≤ε, set R=x _t and stop iteration;

B6: Otherwise, let t=t+1, execute step B4;

B7: Normalize R so that it represents a probability distribution;

The R obtained after completing the above steps is the PageRank value of each node in the directed graph of the evaluation index system, that is, the objective weight of each index.

Beneficial effect: compared with the prior art, the present invention has the following advantages:

1. According to the data related to the evaluation task, the present invention establishes the evaluation knowledge map and the word vector library of the keywords of each node of the knowledge map as the basis for constructing the evaluation index system and weight calculation, which reduces the degree of human participation in the evaluation process and improves the efficiency of evaluation. efficiency.

2. The present invention defines the conversion process from the evaluation index system to the directed graph, so that the algorithm of graph theory can be applied to the evaluation index system, and provides a new idea of weight calculation.

3. The present invention improves the PageRank algorithm, calculates the knowledge reasoning probability by evaluating the similarity of the key word vectors between the nodes in the directed graph of the evaluation index system, and replaces the equal possible transition probability in the PageRank random walk model, so that the calculated The PageRank value of each index can not only reflect the objective weight measured by knowledge, but also avoid the topic deviation problem in the traditional PageRank algorithm.

Description of drawings

Fig. 1 is a schematic diagram of the calculation flow of the evaluation weight based on the knowledge graph in the present invention.

Fig. 2 is a schematic diagram of the knowledge map established by the present invention.

Fig. 3 is a schematic diagram of the evaluation index system established based on the knowledge map in the present invention.

FIG. 4 is a diagram of error changes during the iterative process of the improved PageRank algorithm proposed by the present invention.

Fig. 5 is the bottom evaluation weight obtained based on the improved PageRank algorithm in the present invention.

Detailed ways

Below in conjunction with accompanying drawing and specific embodiment, further illustrate the present invention, should be understood that these embodiments are only for illustrating the present invention and are not intended to limit the scope of the present invention, after having read the present invention, those skilled in the art will understand various aspects of the present invention Modifications in equivalent forms all fall within the scope defined by the appended claims of this application.

The present invention provides a method for calculating evaluation weights based on knowledge graphs, as shown in Figure 1, which includes the following steps:

S1: Establish the evaluation knowledge graph and keyword vector database based on the evaluation-related data, use the evaluation knowledge graph as the evaluation index database of the evaluation index system, and use the keyword vector database as the basis for evaluation index retrieval;

S2: Combining the set evaluation target and the calculated adaptation index in the keyword vector similarity search index library to form an evaluation index system;

S3: Convert the established evaluation index system into a directed graph, and calculate the knowledge reasoning probability through the keyword vector similarity between the nodes of the directed graph to replace the random walk probability in the existing PageRank algorithm, forming an improved PageRank algorithm ;

S4: Calculate the PageRank value of each node as the objective weight of each evaluation index according to the improved PageRank algorithm.

The method for establishing the evaluation knowledge map in step S1 is: according to the evaluation task, collect relevant text data, and use the knowledge map construction technology to establish the evaluation knowledge map;

The establishment method of the keyword vector database is as follows: the word embedding algorithm is used to obtain the word vector database for the collected text data, and then the keywords of each node of the knowledge graph are extracted, and the corresponding word vectors are searched to form the keyword vector database.

The formation of the evaluation index system in step S2 is as follows: according to the set evaluation target, select the nodes with specific labels in the knowledge map, and calculate the similarity between them and the keyword vector vector of the evaluation target for further retrieval, and then, all the nodes that satisfy the The evaluation indicators of the conditions and the dependencies among them are summarized to form an evaluation index system.

Evaluation objectives are set by experts according to the evaluation tasks;

The calculation method of keyword vector similarity is:

表示评价目标的关键词词向量，/>

表示知识图谱节点的关键词词向量。in,

Indicates the keyword vector of the evaluation target, />

Keyword vectors representing knowledge graph nodes.

The specific process of step S3 is:

A1: Construct a directed graph of the evaluation index system

After completing the construction of the evaluation index system, each evaluation index in the evaluation index system is converted into a node in the definition of the directed graph, and the directional dependencies between the indicators are converted into directed edges in the definition of the directed graph to realize The conversion process from the evaluation index system to the directed graph;

A2: Establish an improved random walk model

Define the directed graph of the evaluation index system:

G=(V,E)

Among them, V and E respectively represent the collection of nodes and directed edges, and each directed edge corresponds to a pair of ordered nodes;

The random walk model is essentially a Markov chain defined on a directed graph, and its Markov property is expressed as the state transition probability from the current node state to the next moment node state is only related to the current node state. In the improved random walk model, nodes no longer transfer to the nodes they point to with equal probability, but transfer with knowledge reasoning probability. The directed graph of the evaluation index system is essentially established based on the knowledge graph, so the nodes in the directed graph, that is, the nodes in the knowledge graph corresponding to the evaluation index, can query the keyword vector of the node through the knowledge graph;

Define the knowledge similarity of nodes v _i and v _j in the directed graph G of the evaluation index system as:

为知识图谱中节点v_i、v_j的排序最高的关键词词向量；in,

is the highest-ranked keyword word vector of nodes v _i and v _j in the knowledge graph;

Assuming that node j in the directed graph has k directed edges connecting out, the state transition probability from node j to each node can be expressed by the knowledge similarity as:

Among them, p′ _ij is the probability of knowledge reasoning, which means the probability of transferring from evaluation index j to evaluation index i with knowledge as the measure;

Assuming that there are m nodes in the directed graph G, the state transition probability matrix of the improved random walk model is expressed as:

符合状态转移概率矩阵的性质。Obviously, p′ _ij ≥ 0,

It conforms to the property of the state transition probability matrix.

The calculation method of the PageRank value in step S4 is:

The directed graph of the evaluation index system established based on the knowledge map may not necessarily meet the conditions of non-periodicity and strong connectivity, so its Markov chain may not have a steady-state distribution.

由P和P'的线性组合构成新的状态转移概率矩阵，在此基础上得到的马尔科夫链可以证明一定具有平稳分布R，R有如下等式：For the directed graph G of the general evaluation index system, assuming that it has m nodes, the state transition probability matrix based on the improved random walk model is P. If it is required to obtain the steady-state distribution of the Markov chain, then in the directed graph In addition to defining its own improved random walk model on G, it is also necessary to define a complete random walk model, and the element values of its state transition probability matrix P' are

A new state transition probability matrix is formed by the linear combination of P and P', and the Markov chain obtained on this basis can prove to have a stationary distribution R, and R has the following equation:

Among them, the linear combination coefficient d (0≤d≤1) is the damping factor, according to empirical values, E is an m×m all-1 matrix; l is an m-dimensional vector with all components being 1, and R indicates that in general The smooth distribution of R, each component of R is the PageRank value of each node, that is, the objective weight corresponding to each evaluation index:

R＝[PR(v ₁ ),PR(v ₂ ),...,PR(v _m )] ^T ，v ₁ ,v ₂ ,...,v _m ∈V

公式第一项表示状态平稳分布时，模型按照状态概率转移矩阵P游走到各节点的概率分布，且这样的概率分布占权重d；第二项其表示状态平稳分布时，模型按照状态概率转移矩阵P'等概率游走到各节点的概率分布，且这样的概率分布占权重1-d。Among them, PR(v _i )>0,

When the first item of the formula indicates the state is in a stable distribution, the model moves to the probability distribution of each node according to the state probability transition matrix P, and such a probability distribution accounts for the weight d; when the second item indicates that the state is in a stable distribution, the model transfers according to the state probability Matrix P' walks to the probability distribution of each node with equal probability, and such probability distribution accounts for the weight 1-d.

The calculation process of PageRank value includes the following steps:

B1: Calculate the state transition probability matrix P of the directed graph G of the evaluation index system based on the knowledge measurement;

B2: let t=0, select the initial state distribution x ₀ ;

B3: Calculate the general state transition probability matrix A of the directed graph

B4: Iterate and normalize the result vector

y _t+1 =Ax _t

B5: When ||x _t+1 -x _t ||≤ε, set R=x _t and stop iteration;

B6: Otherwise, let t=t+1, execute step B4;

B7: Normalize R so that it represents a probability distribution;

The R obtained after completing the above steps is the PageRank value of each node in the directed graph of the evaluation index system, that is, the objective weight of each index.

In order to verify the actual effect of the above-mentioned scheme, in this embodiment, the above-mentioned evaluation weight calculation method is applied to an example. Specifically, the evaluation index of the evaluation target of the risk of manipulator failure in the production line of an automobile manufacturing enterprise is calculated. The details are as follows:

The first step: collect data related to the evaluation task, and construct the evaluation knowledge graph and the keyword vector library of each node.

The data in this embodiment comes from an automobile manufacturing company, which produced a large number of equipment problem solving reports during a certain period of production. The cause of the problem and the corresponding maintenance process, etc., contain specific evaluation indicators and rich expert knowledge related to the evaluation objectives.

Part of the data in the company's equipment problem solving report is marked and processed as the training, verification, and test data sets of the knowledge extraction Casrel model, and the other part is input into the trained Casrel model to realize knowledge extraction. In this section, manual labeling is used to ensure the quality of the dataset and improve the training effect of the model.

Analyzing the textual description of production line failures in the equipment problem solving report, the following five triples were identified to describe the entity categories and the relationship between them in the equipment problem solving report, respectively: (specific failure, traceable cause , cause of failure), (cause of failure, cause, cause of failure), (handler, use, maintenance method), (cause of failure, involved, parts), (equipment model, belonging, equipment classification).

The text of the equipment problem solving report is marked according to the above triplet form to obtain the triplet data set, and the marked data set is divided into training set, verification set and test set according to the ratio of 8:1:1. After training, the Casrel model performs knowledge extraction on the remaining data, and stores the extracted triples in the neo4j graph database to form an evaluation knowledge graph, as shown in Figure 2.

For the text data of the equipment problem solving report, the Word2vec word embedding algorithm is used to obtain the word vector library. For the node information in the knowledge graph, refer to the TF-IDF document keyword algorithm to define the keyword algorithm of the knowledge graph node, which is defined as follows:

TF-IDF=TF*IDF

Among them, c represents the number of times a word appears in the node, m represents the total number of words in the node, N represents the total number of nodes in the knowledge graph, and n represents the number of nodes containing the word.

The calculated TF-IDF values are arranged in descending order, and the word with the highest ranking or several higher words are used as keywords, and the keyword is searched for the corresponding word vector in the word vector library to obtain the word vector library.

Step 2: According to the evaluation target and keyword retrieval algorithm, search and evaluate the adaptation index in the knowledge map to construct the evaluation index system.

For the analysis of the evaluation target of the failure risk of the production line manipulator, the specific evaluation indicators subdivided into the next level can correspond to the category of "specific failure" in the knowledge map entity category, so when evaluating knowledge map query and "manipulator" This keyword has a high degree of similarity in word vectors and a node with the label "specific fault". Five nodes with the highest similarity to the evaluation target were found, namely: manipulator clamping fault, manipulator zero return fault, manipulator grasping fault There are five nodes: material failure, manipulator tripping, and manipulator not moving. Take it as a specific fault element set, and then further query the nodes related to these 5 nodes. After analysis, select the node whose label is the cause of the fault as the underlying evaluation index, a total of 30, and use the weight of the underlying index as the evaluation weight that needs to be calculated . The original relationship of the retrieved evaluation indicators in the knowledge graph is retained, and the constructed evaluation index system is shown in Figure 3.

Step 3: Calculate the objective evaluation weight of the underlying evaluation index according to the improved PageRank algorithm.

According to the improved PageRank algorithm, the evaluation index system is transformed into a corresponding directed graph. Firstly, the similarity of the keyword vectors between the indicators with relationship is calculated, and the state transition matrix based on knowledge reasoning probability is obtained.

的向量35维向量，表示初始为等权重状态。在改进PageRank的迭代计算过程中，误差的变化如图所4示，可见误差逐渐减小且趋于平稳，表明最后得的PageRank值符合平稳分布。In the iterative calculation of improved PageRank, the allowable error ε is set to 0.00001, and the damping coefficient d is set to 0.85. At the same time, because the final stable distribution has nothing to do with the initial distribution, in order to be more practical, the initial distribution is set according to the number of evaluation indicators. Components are

The vector of is a 35-dimensional vector, representing the initial state of equal weight. During the iterative calculation process of the improved PageRank, the change of the error is shown in Figure 4. It can be seen that the error gradually decreases and tends to be stable, indicating that the final PageRank value conforms to a stationary distribution.

Finally, a stable distribution is obtained, and the PageRank values of each underlying evaluation index are extracted and normalized to obtain an objective evaluation weight measured by knowledge, as shown in Figure 5.

This embodiment also provides an evaluation weight calculation system based on a knowledge map, the system includes a network interface, a memory, and a processor; wherein, the network interface is used to implement the signal during the process of sending and receiving information with other external network elements receiving and sending; the memory is used to store computer program instructions that can be run on the processor; the processor is used to execute the steps of the above-mentioned consensus method when running the computer program instructions.

This embodiment also provides a computer storage medium, where a computer program is stored in the computer storage medium, and the method described above can be implemented when a processor executes the computer program. The computer readable medium may be considered tangible and non-transitory. Non-limiting examples of non-transitory tangible computer-readable media include non-volatile memory circuits such as flash memory circuits, erasable programmable read-only memory circuits, or masked read-only memory circuits, volatile memory circuits such as static random access memory circuits or dynamic random access memory circuits), magnetic storage media such as analog or digital magnetic tape or hard drives, and optical storage media such as CD, DVD or Blu-ray discs, etc. The computer programs include processor-executable instructions stored on at least one non-transitory tangible computer readable medium. A computer program may also include or rely on stored data. A computer program may include a basic input/output system (BIOS) for interacting with the hardware of a special purpose computer, device drivers for interacting with specific devices of a special purpose computer, one or more operating systems, user application programs, background services, background applications, etc. .

Those skilled in the art should understand that the embodiments of the present application may be provided as methods, systems, or computer program products. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present application is described with reference to flowcharts and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each procedure and/or block in the flowchart and/or block diagram, and a combination of procedures and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a An apparatus for realizing the functions specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions The device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby The instructions provide steps for implementing the functions specified in the flow chart or blocks of the flowchart and/or the block or blocks of the block diagrams.

Claims (7)

1. The evaluation weight calculation method based on the knowledge graph is characterized by comprising the following steps of:

s1: establishing an evaluation knowledge graph and a keyword vector library according to the evaluation related data, taking the evaluation knowledge graph as an evaluation index library of an evaluation index system, and taking the keyword vector library as a basis for evaluation index retrieval;

s2: combining the set evaluation target and the calculated adaptation index in the keyword vector similarity retrieval index library to form an evaluation index system;

s3: converting the established evaluation index system into a directed graph, and calculating knowledge reasoning probability to replace random walk probability in the conventional PageRank algorithm through keyword vector similarity among nodes of the directed graph to form an improved PageRank algorithm;

s4: and calculating the PageRank value of each node according to the improved PageRank algorithm, and taking the PageRank value as the objective weight of each evaluation index.

2. The method for calculating the evaluation weight based on the knowledge graph according to claim 1, wherein the method for establishing the evaluation knowledge graph in step S1 is as follows: according to the evaluation task, collecting related text data, and establishing an evaluation knowledge graph by adopting a knowledge graph construction technology;

the method for establishing the keyword vector library comprises the following steps: and (3) obtaining a word vector library by adopting a word embedding algorithm for the collected text data, extracting keywords of each node of the knowledge graph, and searching for corresponding word vectors to form the keyword vector library.

3. The method for calculating the evaluation weight based on the knowledge graph according to claim 1, wherein the forming of the evaluation index system in the step S2 is specifically: and screening out nodes with specific labels in the knowledge graph according to the set evaluation targets, calculating the keyword vector similarity between the nodes and the evaluation targets for further retrieval, and then summarizing all the evaluation indexes meeting the conditions and the dependency relations between the evaluation indexes to form an evaluation index system.

4. A knowledge-graph-based evaluation weight calculation method according to claim 3, wherein said evaluation target is set by an expert according to an evaluation task;

the keyword vector similarity is calculated in the following way:

wherein,,

keyword vector representing evaluation target, +.>

Keyword word vectors representing knowledge-graph nodes.

5. The method for calculating the evaluation weight based on the knowledge graph according to claim 1, wherein the step S3 is specifically:

a1: constructing a directed graph of an evaluation index system

After the construction of the evaluation index system is completed, each evaluation index in the evaluation index system is converted into a node in the definition of the directed graph, the directional dependency relationship among the indexes is converted into a directed edge in the definition of the directed graph, and the conversion process from the evaluation index system to the directed graph is realized;

a2: establishing an improved random walk model

Defining an evaluation index system directed graph:

G＝(V,E)

wherein V and E represent a set of nodes and directed edges, respectively, each directed edge corresponding to a pair of ordered nodes;

defining node v in evaluation index system directed graph G _i And v _j The knowledge similarity of (2) is:

wherein,,

is node v in the knowledge graph _i 、v _j Keyword word vectors with highest ranks;

assuming that there are k directed edges connecting out to node j in the graph, the probability of state transition from node j to each node can be represented by the knowledge similarity:

wherein p' _ij Is a knowledge reasoning probability representing a slaveThe probability that the evaluation index j is transferred to the evaluation index i is measured by taking knowledge as a measure;

assuming that there are m nodes in the graph G, the state transition probability matrix of the improved random walk model is expressed as:

obviously, p' _ij ≥0，

The properties of the state transition probability matrix are met.

6. The method for calculating the evaluation weight based on the knowledge graph according to claim 5, wherein the method for calculating the PageRank value in step S4 is as follows:

for the evaluation index system directed graph G, assuming that m nodes are provided, obtaining a state transition probability matrix based on the improved random walk model as P, if the steady-state distribution of the Markov chain is required to be obtained, defining a complete random walk model on the directed graph G besides the improved random walk model of the directed graph G, wherein the element values of the state transition probability matrix P' are all

The new state transition probability matrix is formed by the linear combination of P and P', and the Markov chain obtained on the basis can prove that the state transition probability matrix has stable distribution R, wherein the R has the following equation:

wherein, the linear combination coefficient d (d is more than or equal to 0 and less than or equal to 1) is a damping factor, E is an m multiplied by m full 1 matrix according to empirical values; l is an m-dimensional vector with each component being 1, R represents the steady distribution under the general condition, and each component of R is the PageRank value of each node, namely the objective weight corresponding to each evaluation index:

R＝[PR(v ₁ ),PR(v ₂ ),…,PR(v _m )] ^T ，v ₁ ,v ₂ ,...,v _m ∈V

wherein there is PR (v) _i )＞0，

When the first term of the formula represents state stable distribution, the model moves to probability distribution of each node according to a state probability transition matrix P, and the probability distribution occupies a weight d; the second term, which represents the state stationary distribution, models walk to the probability distribution of each node according to the probability such as the state probability transition matrix P', and such probability distribution takes weight 1-d.

7. The method for calculating the evaluation weight based on the knowledge graph according to claim 6, wherein the calculation process of the PageRank value in the step S4 comprises the following steps:

b1: calculating a state transition probability matrix P of the evaluation index system directed graph G based on knowledge measurement;

b2: let t=0, select initial state distribution x ₀ ；

B3: calculating a general state transition probability matrix A of the directed graph

B4: iterating and normalizing result vectors

y _t+1 ＝Ax _t

B5: when x _t+1 -x _t Let r=x when l is less than or equal to ε _t Stopping iteration;

b6: otherwise, let t=t+1, execute step B4;

b7: normalizing R to make it represent probability distribution;

and R obtained after the steps are completed is the PageRank value of each node of the directed graph of the evaluation index system, namely the objective weight of each index.

Images