Electric power material supplier feature analysis method and system based on business information topic model
Through the analysis method of the electric power material supplier's feature analysis method based on the business information theme model, the knowledge graph is constructed and non-text data is processed, and the problems of information silos and low intelligence in the management of electric power material suppliers are solved, and the comprehensive characteristics of suppliers are realized Analyze and evaluate, improving the application and intelligence of data.
Patent Information
- Application Number
- CN202410725082.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-05
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing power supply supplier management has information islands, lack of scientific and efficient quantitative evaluation methods, and low intelligence, resulting in insufficient informationization and insufficient data application.
Using a method based on the Business Information Theme Model (BTM), the knowledge graphs in the field of power supply and the BERT-BiLSTM-CRF model are used to identify named entities, and the non-text data are processed in combination with Dirichlet distribution to realize supplier feature analysis. and assessment.
A comprehensive feature analysis and evaluation of power material suppliers has been realized, data interoperability and intelligence have been improved, standardized and reliable quantitative evaluation methods have been provided, and dependence on manpower has been reduced.
Smart Images

Figure CN119005312A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer technology, and in particular relates to a method and system for analyzing characteristics of power material suppliers based on a business information topic model. Background Art
[0002] With the advancement of the development of digital society, the digital transformation of power companies has continued to develop, which has put forward new requirements for the supply chain management level of enterprises. As an important link in supply chain management, the management of power material suppliers is closely related to the security and reliability of the power grid. Selecting material suppliers with good comprehensive reputation, qualifications and performance has a significant impact on the long-term stable development of power companies.
[0003] At present, most power companies have gradually integrated digital and intelligent means into the supplier management process, which requires power companies to further strengthen data governance capabilities and improve the information management model of supplier data collection, processing, and application. In recent years, the analysis and evaluation methods of power material suppliers have mainly focused on how to build a multi-criteria decision-making model, mine effective information from expert scores, and output evaluation results, including the analytic hierarchy process (AHP), fuzzy comprehensive evaluation method, data envelopment analysis (DEA), and ideal solution method (TOPSIS). However, the material suppliers of power companies are characterized by a wide variety and large number. The traditional extensive management model can no longer meet the needs of supply chain integration and intelligence. It is necessary to optimize the traditional supply chain management method through new information technology means and realize the modern management method of production and supply integration in the upstream of the supply chain. The existing characteristic analysis and evaluation methods for power material suppliers mainly face the following challenges:
[0004] 1. The integration between different information systems is low and data is difficult to communicate and share, resulting in the emergence of "information islands";
[0005] 2. Business links such as supplier qualification evaluation and performance evaluation rely too much on subjective evaluation information from experts, lacking standardized and reliable evaluation standards and scientific and efficient quantitative evaluation methods;
[0006] 3. Faced with complex business scenarios and a wide variety of suppliers, the intelligence level of the information system is not high, and there are still a large number of business scenarios that rely on manpower.
[0007] In short, the current power companies still have problems in the management of material suppliers, such as insufficient informatization, insufficient data application, and redundant module functions. This requires further strengthening of data governance capabilities and improving the information management model of material supplier data collection, processing, and application. Summary of the invention
[0008] In view of the shortcomings of the existing technology, the present invention proposes a method for analyzing the characteristics of power material suppliers based on a business information topic model (BTM, BusinessTopic Model), which can fully extract various types of data in the existing system to analyze and evaluate the characteristics of power material suppliers.
[0009] To achieve the above objectives, the technical solution of the present invention is to embed the supplier's non-text data into the document topic generation model of Dirichlet distribution, and combine it with the knowledge graph to perform semantic disambiguation for the feature analysis task of power material suppliers.
[0010] A method for analyzing the characteristics of power material suppliers based on a business information topic model, characterized in that it includes designing a knowledge graph construction method related to power material suppliers, based on a knowledge base related to the field of power material supply, identifying named entities in the text of the knowledge base related to the field of power material supply through named entity recognition technology, including supply entities, power enterprise entities, material entities, business information entities, etc., and then linking entities with the same name, then extracting the relationship between entities of different categories, and creating entity nodes and relationships between entities in a graph database to link the entities, and finally completing the construction of a knowledge graph in the field of power material supply;
[0011] The BERT-BiLSTM-CRF model with Len_loss added is used to perform named entity recognition on non-text information in documents and databases. The model is divided into three modules from bottom to top: BERT module, BiLSTM (bidirectional long short-term memory network) module and CRF (conditional random field) module;
[0012] The business information data of power material suppliers is divided into qualification level, operation capability, procurement performance, quality management, credit behavior, etc. from the business logic, and is mainly divided into four categories from the data structure, namely time data, character data, numerical data and text data;
[0013] If the supplier's data set contains time, character or numerical data in addition to document data, and the traditional topic model can only process document data but not non-document data, the present invention proposes a non-text data processing method based on PV (attribute-value) pairs, which converts all the supplier's business data into documents by embedding time, character and numerical data;
[0014] Preferably, in the temporal information embedding process, since the data records of a supplier may be in different time periods or time points, the year information is retained, and months and days are used as parameters for training and prediction, but smaller units of hours, minutes and seconds are deleted. Subsequently, all data records of the supplier are placed on the timeline, and the minimum unit is changed from one day to n days, so that n adjacent days are represented as the same time feature.
[0015] Preferably, in the embedding process of character data, the attributes and values after semantic disambiguation are converted into embeddable texts, and then converted into words in the form of "P+V".
[0016] Preferably, in the embedding process of numerical data, the units of attributes and values after semantic disambiguation are converted into embeddable texts, and then converted into words in the form of "P+V". BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the technical solution of the embodiment of the present application, the following is a brief introduction to the drawings required to be used in the embodiment of the present application:
[0018] Figure 1 BTM topic model structure diagram
[0019] Figure 2 Schematic diagram of embedding time-based data in supplier information DETAILED DESCRIPTION
[0020] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0021] In view of the problems existing in the prior art, the present invention provides a method for analyzing the characteristics of power material suppliers based on a business information topic model. The present invention is described in detail below in conjunction with the accompanying drawings.
[0022] 1. The present invention designs a knowledge graph construction method related to power material suppliers, which is used for semantic disambiguation when cleaning the business information data of suppliers. Based on the knowledge base related to the field of power material supply, the named entities in the text of the knowledge base related to the field of power material supply are identified through named entity recognition technology, including supply entities, power enterprise entities, material entities, business information entities, etc., and then the entities with the same name are linked, and then the relationship between entities of different categories is extracted, and entity nodes and relationships between entities are created in the graph database to link the entities, and finally the knowledge graph in the field of power material supply is constructed, which specifically includes the following steps.
[0023] (1) Entity recognition. The entity recognition module uses the BERT-BiLSTM-CRF model with Len_loss added. It is mainly divided into three modules from bottom to top: BERT module, BiLSTM (bidirectional long short-term memory network) module and CRF (conditional random field) module. The BERT module is responsible for generating the word embedding vector of the input sequence, the BiLSTM module is responsible for generating the label probability distribution of each input word through the forward and backward LSTM, and finally the CRF module ensures the correctness of the output continuous label through the state transition matrix. The calculation steps of the BERT-BiLSTM-CRF model are as follows:
[0024] First, the input sequence is divided into X = (X1, X2, X3, ..., X n ) and then converted into a one-hot vector V = (V1, V2, V3…, V n ), where V i =(0,…,0,1,0,…,0), its V i The length of the vector is the size of the vocabulary, the word sequence number is 1, and the rest are 0, and b is the length of the input sequence.
[0025] The one-hot vector V=(V1,V2,V3…,V n ) Input the BERT model and output the wordembedding feature vector E = (E1, E2, E3, …, E n ), where E i ∈R d ,E∈R n×d , d is the output vector dimension of the BERT model, and n is the input sequence length.
[0026] The word embedding feature vector E = (E1, E2, E3, ..., E n ) Input the BiLSTM model in the order of the input sequence, and the hidden layer feature vector is calculated by the forward LSTM model The hidden layer feature vector is calculated by the backward LSTM model The hidden layer feature vectors calculated by the forward LSTM model and the backward LSTM model are concatenated according to the first dimension to obtain the output feature vector of the BiLSTM model in each step. in H t ∈R 2m , m is the dimension of the hidden layer feature vector of the LSTM model, and the BiLSTM model calculation output of the complete sequence is H = (H1, H2, H3, ..., H n ), where H∈R n×2m , n is the length of the input sequence.
[0027] The BiLSTM model output H = (H1, H2, H3, ..., H n ), input into the fully connected neural network, and H t ∈R 2m Mapped to the labeled probability distribution P of the sequence t ∈R k , where k is the number of labeled categories for named entity recognition, and finally the probability distribution of named entity recognition of the sequence is obtained as P = (P1, P2, P3, ..., P n ), where P∈R n×k .
[0028] The CRF layer stores the label state transition probability matrix A∈R (k+2)×(k+2) , the entire A matrix is used as the model training parameters, where A i,j represents the probability of transferring from the i-th label to the j-th label. For the input sequence X = (X1, X2, X3, ..., X n ), and obtain any label sequence Y=(Y1,Y2,Y3,…,Y n ) is shown in Formula 1.
[0029]
[0030] Formula 1 can be used to calculate that for an input sequence X=(X1,X2,X3,…,X n ) of any label sequence Y=(Y1,Y2,Y3,…,Y n ), so we can finally use a softmax activation function to convert any output label sequence Y = (Y1, Y2, Y3, ..., Y n ) is converted into a probability value, as shown in Formula 2, where Y' is the sequence of all labels.
[0031]
[0032] Therefore, the objective function of model training is to maximize P(Y X |X), where Y X is taken as the input sequence X=(X1,X2,X3,…,X n ) is the ground truth, that is, the correct label sequence. After the log(x) function is calculated, the final objective function of the model training is shown in Formula 3. The goal is to maximize logP(Y X |X).
[0033] logP(Y X |X)=score(X,Y X )-log∑ e(score(X,Y′)) (3)
[0034] Due to the particularity of the power material supply field, the supply entities in the dataset are often twice or even three times longer than other entities. Therefore, Len-loss is added to the objective function of the general named entity recognition model to make the supply entities have more words and longer lengths, as shown in Formula 4, where It represents the probability that the i-th word is recognized as the label I-TAS after the BiLSTM model, n is the input sequence length, and α is a hyperparameter slightly greater than 0 to avoid log(0). The final loss function is shown in Formula 5, where λ is a hyperparameter ranging from 0 to 1, indicating the proportion of Len-loss.
[0035]
[0036] loss=-logP(Y X |X)-λLen_loss (5)
[0037] When the model training completes convergence, a new sequence X=(X1,X2,X3,…,X n ), we can get a label probability distribution matrix corresponding to each word in the sequence, and then we can use the Viterbi algorithm to analyze and get the label sequence Y with the maximum total conditional probability, that is, the most correct one. * =(Y1,Y2,Y3,…,Y n ). Among them, the Viterbi algorithm is mainly a shortest path planning algorithm based on dynamic programming. First, the sequence-state probability matrix W is initialized, where W∈R n×m , n represents the sequence length, m represents the number of states. Then calculate W according to the sequence order i,j , when i=1, Where i represents the i-th moment, x i represents the observation number at the i-th moment in the sequence, and j represents the state number. When i ≥ 1, And record the state value at this moment to reach the maximum value W i,j The state value y of the last moment selected i-1 After calculating W, the final sequence probability is max(W n,j ,j=1,2,...,m), since each state at each moment records its state value y at the previous moment i-1 , so after backtracking, we can get the complete state sequence Y=(y1,y2,…,y n ).
[0038] (2) Affiliation reasoning. For the same paragraph, if the supply entity and the business entity appear at the same time, the supply entity is considered to be the source of the business entity. At the same time, if it is the first time that they appear at the same time, the relationship weight is set to 1, otherwise the relationship weight is increased by 1. The extraction method of other relationship types is the same.
[0039] (3) Synonym relationship reasoning. The business information of suppliers cannot be completely determined in advance, that is, it is impossible to use a synonym table to unify entities. It is necessary to use a simple string similarity calculation method to determine the similarity between two entities. When a new entity is identified, the similarity is calculated with the existing entities of the same type. If there are entities with high similarity, the entity with shorter text length is used as the unified entity. If there are no entities with high similarity, it is used as a new entity.
[0040] (4) Store the discovered entities and relationships in a graph database.
[0041] 2. Extract the internal data of the data center and business systems (including ECP, ERP, PMS, infrastructure management and control systems) and the full amount of external data (including social credit, finance, administrative penalties, etc.) to ensure the real-time, accuracy and uniqueness of the data and build a data system. The business information data of power material suppliers can be divided into qualification level, operational capability, procurement performance, quality management, credit behavior, etc. from the business logic. This information is mainly divided into four categories from the data structure, namely time data, character data, numerical data and text data. If there is only text data in a supplier's data set, the following process is used:
[0042] (1) Select a document m from the dataset. Construct a Dirichlet distribution for the hyperparameters and sample the topic distribution of document m Right now The obtained Dirichlet distribution is a three-layer Bayesian model, which mainly includes the three-layer basic structure of words, topics and documents, as shown in Formula 6. The P(word|document) in the formula is known, and P(word|topic) and P(topic|document) can be obtained through iterative training, that is, the topic probability distribution corresponding to each document in the entire document collection and each word in the document.
[0043]
[0044] The document generation of the entire Dirichlet model is shown in Formula 7, where the calculation of the Dirichlet distribution is shown in Formula 8: According to Formula 9, we get
[0045]
[0046] (2) Distribution by topic The topic z of the nth word in document m is calculated m,n Multinomial distribution, that is get
[0047] (3) Construct a Dirichlet distribution for the hyperparameters and sample the generated topic z m,n Corresponding word distribution Right now According to formula 9, we can get
[0048] (4) From word distribution Sampling generated words w m,n Multinomial distribution, that is get
[0049] (5) Finally, the four probabilities of all documents are multiplied to obtain The whole process is to generate the topic distribution θ from α, and then determine the specific topic z from the topic distribution θ; and generate the word distribution corresponding to the topic z from β Then by word distribution Determine the specific word w in the document, so formula 8 and the above process are equivalent to formula
[0050] The joint probability distribution expressed in formula 9 is:
[0051]
[0052] 3. Processing non-text data. If the supplier's data set contains time, character and numerical data in addition to document data, and the traditional topic model can only process text data but not non-text data, the present invention proposes a non-text data processing method based on PV (attribute-value) pairs, which converts all the supplier's business data into documents by embedding time, character and numerical data. However, the PV pair method first needs to eliminate the ambiguity of the meaning of attributes or characters caused by different data sources.
[0053] (1) Use the knowledge graph of power material suppliers to unify the naming in non-text data to eliminate ambiguity. The first is attribute unification. For example, in the supplier level field, there may be two ways to represent the same value, such as L1 and LEVEL1. Create a mapping table for this field and store it in the data processing class to ensure that they are mapped to the same word in data processing. The second is attribute disambiguation. For example, the field types of "registered capital" and "turnover" are both numeric values, so the same value may appear in different fields, such as "registered capital 50 million" and "turnover 5000 yuan" to represent 5000, and only the prefix of the field name is used to distinguish different "contexts".
[0054] After obtaining the business information entities and material entities with a high degree of relevance to the supply entity "Xianning Transformer Co., Ltd." through knowledge graph reasoning, the combined query conditions of these multiple entity attributes can be generated. Since only the relationship weights between the supply entity and the relevant business entity are considered when obtaining the relevant multiple entities, and the combined relationship between the business entities is not considered, the combined query cannot be performed directly. Instead, it is combined through various entity link relationships of the knowledge graph, supplier entity: "Xianning Transformer Co., Ltd." and material entities "transformer" and "relay", business entities "turnover" and "registered capital". In the power material supply knowledge graph, entities without relationship cannot be combined to generate conditions. Because the correct data cannot be queried, various condition combination forms of query and reasoning can be provided based on the power material supply knowledge graph.
[0055] (2) In the embedding process of character data, the attributes and values after semantic disambiguation are converted into embeddable text and converted into words in the form of "P+V", such as "the enterprise type is a high-tech enterprise".
[0056] (3) In the embedding process of numerical data, the units of attributes and values after semantic disambiguation are converted into embeddable text and converted into words in the form of "P+V", such as "last year's sales volume was 100 million yuan."
[0057] (4) In the embedding process of time-type data, since the data records of suppliers may be in different time periods or time points, the year information is retained, and months and days are used as training and prediction parameters, but the smaller units of hours, minutes and seconds are deleted. Then all the data records of the supplier are placed on the timeline, and the minimum unit is changed from one day to n days, so that the adjacent n days are represented as the same time feature. In this way, the time information of the supplier's business data can be converted into word text through "Date+time number", such as Date1, Date2, etc. Figure 2 shown.
[0058] (5) Use non-text data embedding to convert various business information of suppliers into documents, as shown in formulas 10, 11, 12, 13 and 14. Replace the "word" in formula 6 with character information embedding S, time information embedding T and numerical information embedding C. Then use Dirichlet's training method to obtain the embedding representation of each user, that is, the document-topic probability distribution and word-topic probability distribution. Then use Dirichlet's prediction method to obtain the embedding representation of new data, that is, the topic probability distribution. Finally, obtain the probability of suppliers on topics (features) such as basic capabilities, process control, finished product quality, service support, and credit level. The model framework of BITM is as follows: Figure 1 shown.
[0059]
[0060] W=[S,T,C] (11)
[0061] B=Business_embedding(s,g) (12)
[0062] T=temporal_embedding(t,d) (13)
[0063] C = content_embedding (c) (14)
[0064] in, is the topic distribution of document i, which follows the Dirichlet distribution, is the word distribution of topic k, which follows the Dirichlet distribution, z i,j is the topic corresponding to the jth word in document i, W i,j is the jth word in document i, S i,j is the jth character embedding word in document i, T i,j is the jth time embedding word in document i, C i,j is the jth numeric embedding word in document i, Is used to generate is the hyperparameter of Dirichlet distribution, s is the supplier’s character information, t is the supplier’s time information, c is the supplier’s numerical information, and g is the character segmentation. i,j The hyperparameters, d is the time interval, generate T i,j The hyperparameters are: M is the number of documents, K is the number of topics, and N is the m is the number of words in document m, and V is the vocabulary size.
Claims
1. A method for analyzing the characteristics of power material suppliers based on a business information topic model, characterized in that: include, Based on the knowledge base related to the field of power material supply, the named entities in the text of the knowledge base related to the field of power material supply are identified through named entities, entities with the same name but different names are linked, and the relationship between entities of different categories is extracted. In addition, entity nodes and relationships between entities are created in the graph database to link entities and build a knowledge graph in the field of power material supply. After extracting internal data and external full data from the data center and business system and eliminating ambiguity in the meaning of attributes or characters due to different data sources, all business data of suppliers are converted into documents through time, character and numerical data embedding based on the constructed knowledge graph in the field of power material supply and the non-text data processing method of PV (attribute-value) pairs.
2. The method for analyzing the characteristics of power material suppliers based on the business information topic model according to claim 1 is characterized in that: When building a knowledge graph, we include The entity recognition module uses the BERT-BiLSTM-CRF model with Len_loss added. The BERT module is used to generate the word embedding vector of the input sequence, the BiLSTM module is responsible for generating the label probability distribution of each input word through the forward and backward LSTM, and finally the CRF module ensures the correctness of the output continuous labels through the state transition matrix. For the same paragraph, if the supply entity and the business entity appear at the same time, the supply entity is considered to be the source of the business entity. At the same time, if it is the first time that they appear at the same time, the relationship weight is set to 1, otherwise the relationship weight is increased by 1. The extraction method of other relationship types is the same; The similarity between two entities is determined by using a simple string similarity calculation method. When a new entity is identified, if there is an entity with high similarity, the entity with shorter text length is used as the unified entity. If there is no entity with high similarity, it is used as a new entity. Store the discovered entities and relationships in a graph database.
3. The method for analyzing the characteristics of power material suppliers based on the business information topic model according to claim 1 is characterized in that: Use the constructed knowledge graph of power material suppliers to unify the naming in non-text data to eliminate ambiguity; After obtaining the business information entities and material entities with high relevance to the supply entities through knowledge graph reasoning, the combined query conditions of these multiple entity attributes can be generated, specifically: Embedding processing of character data, converting attributes and values after semantic disambiguation into embeddable text, and converting them into words in the form of "P+V"; Embedding processing of numerical data: converting the units of attributes and values after semantic disambiguation into embeddable text, and converting them into words in the form of "P+V"; The embedding process of time-type data retains the year information, uses months and days as training and prediction parameters, deletes smaller units of hours, minutes and seconds, and then puts all the data records of the supplier on the timeline, and changes the minimum unit from one day to n days, so that n adjacent days can be represented as the same time feature. In this way, the time information of the supplier's business data can be converted into word text through "Date+time number".
4. The method for analyzing the characteristics of power material suppliers based on the business information topic model according to claim 1 is characterized in that: When extracting internal data and external full data from the data center and business system, it specifically includes Select a document m from the dataset, Construct a Dirichlet distribution for the hyperparameters and sample the topic distribution of document m Right now The obtained Dirichlet distribution is a three-level Bayesian model, Distribution by topic The topic z of the nth word in document m is calculated m,n Multinomial distribution, that is get by Construct a Dirichlet distribution for the hyperparameters and sample the generated topic Z m,n Corresponding word distribution Right now According to the joint probability distribution formula, we can get From word distribution Sampling generated words w m,b Multinomial distribution, that is get Multiply the four probabilities of all documents to get The whole process is to generate the topic distribution θ from α, and then determine the specific topic z from the topic distribution θ; and generate the word distribution corresponding to the topic z from β Then by word distribution Determine the specific word w in the document, so Formula 8 and the above process are equivalent to the joint probability distribution expressed by the following formula: In the temporal information embedding process, since the supplier's data records may be in different time periods or time points, the year information is retained, and months and days are used as training and prediction parameters, but the smaller units of hours, minutes and seconds are deleted. Then all the data records of the supplier are placed on the timeline, and the minimum unit is changed from one day to n days, so that the adjacent n days can be represented as the same time feature. In this way, the time information of the supplier's business data can be converted into word text through "Date+time number".
5. The method for analyzing the characteristics of power material suppliers based on the business information topic model according to claim 1 is characterized in that: The three-layer Bayesian model contains three basic structures of words, topics and documents, as shown in the following formula. If P(word|document) in the formula is known, P(word|topic) and P(topic|document) can be obtained through iterative training, that is, the topic probability distribution corresponding to each document in the entire document collection and each word in the document; The entire document generation of the Dirichlet model is shown as follows: The calculation of Dirichlet distribution is as follows: According to the topic probability distribution formula, The non-text data embedding method is used to convert various business information of suppliers into documents, and then the Dirichlet training method is used to obtain the embedding representation of each user, namely the document-topic probability distribution and the word-topic probability distribution. Subsequently, the Dirichlet prediction method is used to obtain the embedding representation of new data, namely the topic probability distribution, and finally the probability of suppliers based on the topic features of basic capabilities, process control, finished product quality, service support, and credit level is obtained.
6. A power material supplier feature analysis system based on business information topic model, characterized in that: include, The first module is configured to identify named entities in the text of the knowledge base related to the field of power material supply based on the knowledge base related to the field of power material supply through named entities, link entities with the same name, extract the relationship between entities of different categories, create entity nodes and relationships between entities in the graph database to link the entities, and build a knowledge graph in the field of power material supply; The second module is configured to extract internal data and external full data from the data center and business system, eliminate the ambiguity of attribute or character meanings caused by different data sources, and then convert all business data of suppliers into documents through time, character and numerical data embedding based on the constructed knowledge graph in the field of power material supply and the non-text data processing method of PV (attribute-value) pairs.
7. The system according to claim 1, characterized in that When building a knowledge graph, we include The entity recognition module uses the BERT-BiLSTM-CRF model with Len_loss added. The BERT module is used to generate the word embedding vector of the input sequence, the BiLSTM module is responsible for generating the label probability distribution of each input word through the forward and backward LSTM, and finally the CRF module ensures the correctness of the output continuous labels through the state transition matrix. For the same paragraph, if the supply entity and the business entity appear at the same time, the supply entity is considered to be the source of the business entity. At the same time, if it is the first time that they appear at the same time, the relationship weight is set to 1, otherwise the relationship weight is increased by 1. The extraction method of other relationship types is the same; The similarity between two entities is determined by using a simple string similarity calculation method. When a new entity is identified, if there is an entity with high similarity, the entity with shorter text length is used as the unified entity. If there is no entity with high similarity, it is used as a new entity. Store the discovered entities and relationships in a graph database.
8. The system according to claim 6, characterized in that Use the constructed knowledge graph of power material suppliers to unify the naming in non-text data to eliminate ambiguity; After obtaining the business information entities and material entities with high relevance to the supply entities through knowledge graph reasoning, the combined query conditions of these multiple entity attributes can be generated, specifically: Embedding processing of character data, converting attributes and values after semantic disambiguation into embeddable text, and converting them into words in the form of "P+V"; Embedding processing of numerical data: converting the units of attributes and values after semantic disambiguation into embeddable text, and converting them into words in the form of "P+V"; The embedding process of time-type data retains the year information, uses months and days as training and prediction parameters, deletes smaller units of hours, minutes and seconds, and then puts all the data records of the supplier on the timeline, and changes the minimum unit from one day to n days, so that n adjacent days can be represented as the same time feature. In this way, the time information of the supplier's business data can be converted into word text through "Date+time number".
9. The system according to claim 6, characterized in that When extracting internal data and external full data from the data center and business system, it specifically includes Select a document m from the dataset, Construct a Dirichlet distribution for the hyperparameters and sample the topic distribution of document m Right now The obtained Dirichlet distribution is a three-level Bayesian model; Distribution by topic The topic z of the nth word in document m is calculated m,n Multinomial distribution, that is get by Construct a Dirichlet distribution for the hyperparameters and sample the generated topic Z m,n Corresponding word distribution Right now According to the joint probability distribution formula, we can get From word distribution Sampling generated words w m,b Multinomial distribution, that is get Multiply the four probabilities of all documents to get The whole process is to generate the topic distribution θ from α, and then determine the specific topic z from the topic distribution θ; and generate the word distribution corresponding to the topic z from β Then by word distribution Determine the specific word w in the document, so Formula 8 and the above process are equivalent to the joint probability distribution expressed by the following formula: In the temporal information embedding process, since the data records of suppliers may be in different time periods or time points, the year information is retained, and months and days are used as training and prediction parameters, but the smaller units of hours, minutes and seconds are deleted. Then all the data records of the supplier are placed on the timeline, and the minimum unit is changed from one day to n days, so that the adjacent n days are represented as the same time feature. In this way, the time information of the supplier's business data can be converted into word text through "Date+time number".
10. The system according to claim 6, characterized in that The three-layer Bayesian model contains three basic structures of words, topics and documents, as shown in the following formula. If P(word|document) in the formula is known, P(word|topic) and P(topic|document) can be obtained through iterative training, that is, the topic probability distribution corresponding to each document in the entire document collection and each word in the document; The entire document generation of the Dirichlet model is shown as follows: The calculation of Dirichlet distribution is as follows: According to the topic probability distribution formula, The non-text data embedding method is used to convert various business information of suppliers into documents, and then the Dirichlet training method is used to obtain the embedding representation of each user, namely the document-topic probability distribution and the word-topic probability distribution. Subsequently, the Dirichlet prediction method is used to obtain the embedding representation of new data, namely the topic probability distribution, and finally the probability of suppliers based on the topic features of basic capabilities, process control, finished product quality, service support, and credit level is obtained.
Citation Information
Patent Citations
LDA-based user consumption forecasting method for electronic commerce
CN106919997A
Article classification method and device, computer equipment and storage medium
CN108399228A
Knowledge graph fusion method based on entity merging
CN115438188A
