Knowledge representation learning method fusing market subject dishonesty information knowledge graph auxiliary information

Abstract

The invention discloses a knowledge representation learning method fusing market subject dishonesty information knowledge graph auxiliary information, which comprises the following steps: acquiring knowledge triple related auxiliary information of market subject dishonesty information according to a market subject dishonesty information knowledge graph; according to the knowledge triad, embedding entities and relationships of the triad into a representation learning model; according to entity description information in the related auxiliary information, performing vectorization representation on all semantic information of entity description; according to multiple types of entities in the related auxiliary information and hierarchy of the types, performing vectorization representation on hierarchical type information; splicing and training the representation learning model by combining translation model embedding according to the knowledge triad and related auxiliary information; according to the method, the vectorization representation performance of entities and relationships can be effectively improved, and the accuracy during semantic association is improved.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph technology, and in particular to a knowledge representation learning method that integrates auxiliary information of a knowledge graph of dishonest information of market entities. Background Art

[0002] With the continuous improvement of the business environment, the number of market players has grown rapidly, and the market formats have become more complex and diversified, resulting in an increasing number of market dishonest cases and a heavier regulatory task. Therefore, it is urgent to carry out research on the construction of knowledge graphs of dishonest information of market players and intelligent fusion technology. This study aims to develop a knowledge expression model of dishonest information that integrates extraction, mining, analysis, fusion, and reasoning, targeting the characteristics of large data volume, diverse data types, wide functional areas, numerous technologies, and high-dimensional heterogeneous redundancy of knowledge. This will provide strong support for market regulatory authorities and enhance their credit-based market supervision efficiency.

[0003] Knowledge Graph (KG) is a technical method that uses a graph model to describe knowledge and model the relationship between all things in the world. Its main goal is to describe various entities and concepts that exist in the real world and the relationship between them. It is a tool for representing knowledge in a graphical way. Knowledge Representation Learning is a widely used knowledge representation method. It is an abstract expression of the real world. By embedding entities and relationships in the knowledge graph into a continuous vector space, it is convenient for calculation while retaining the structural information in the knowledge graph. Knowledge representation learning can explicitly capture the similarity between entities and relationships by measuring the similarity of low-dimensional embeddings of entities and relationships. This technology can be applied to many downstream tasks, such as knowledge question answering, recommendation systems, etc.

[0004] Since Google used knowledge graphs to improve the search experience and search quality in 2012, knowledge graphs have attracted much attention from all walks of life, and many large-scale knowledge graphs have emerged, such as Freebase, WordNet, DBpedia, and YAGO. In recent years, learning vector representations of entities and relations in knowledge graphs has become extremely popular. Models such as TransE regard relations as translation vectors between entities, that is, vectors from the head entity to the tail entity. Due to its simplicity and effectiveness, TransE has become a typical model for knowledge representation learning. In addition, based on TransE, other extended models such as TransH, TransR, RESCAL, and ANALOGY have been improved on the basis of TransE to better handle different types of relations or improve the performance of the model.

[0005] However, existing knowledge representation methods usually only focus on the triples themselves, ignoring the external auxiliary information around the triples during modeling, and failing to fully utilize this external information to enrich the semantic content of knowledge representation. Entities and relationships in knowledge graphs usually have rich additional auxiliary information, such as type, description, name, logical rules, and visual information. This information characterizes the semantics of entities and relationships from multiple perspectives and is very valuable prior knowledge. By modeling this auxiliary information and embedding it into the representation learning model, the vectorization representation performance of entities and relationships can be improved, and the accuracy of semantic association can be improved. Summary of the invention

[0006] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a knowledge representation learning method that integrates the auxiliary information of the knowledge graph of the market entity's default information. The present invention can effectively improve the vectorization representation performance of entities and relationships and improve the accuracy of semantic association.

[0007] To achieve the above purpose, the technical solution adopted by the present invention is: a knowledge representation learning method integrating auxiliary information of knowledge graph of dishonest information of market entities, comprising the following steps:

[0008] Step 1: According to the knowledge graph of the market subject's dishonest information, obtain the knowledge triplet related auxiliary information of the market subject's dishonest information;

[0009] Step 2: Based on the knowledge triples, embed the entities and relations of the triples themselves into a representation learning model;

[0010] Step 3: vectorize all semantic information of the entity description according to the entity description information in the relevant auxiliary information;

[0011] Step 4: vectorize the hierarchical type information according to the multiple types of entities in the relevant auxiliary information and the hierarchy of the types;

[0012] Step 5: Based on the knowledge triples and related auxiliary information, the translation model is embedded in combination to perform concatenation and train the representation learning model.

[0013] As a further improvement of the present invention, the step 1 is specifically as follows:

[0014] Extract the structured information of the knowledge graph from the knowledge graph of market subject dishonesty information, and obtain the knowledge triple S = (h s ,r,t s ); extract entity description information from the market subject dishonesty information corpus to obtain the Chinese entity description information set D; extract textual relationship information of entity pairs from the market subject dishonesty information corpus to obtain hierarchical information c = {c(1) ,c (2) ,…,c (m)}.

[0015] As a further improvement of the present invention, the step 2 specifically comprises the following steps:

[0016] Step 2.1: Connect the entity of the triple to the relation S = (h s ,r,t s ) is modified through the attention mechanism;

[0017] For each specific triple (h i ,r k ,t j )∈S, the linear transformation representation of the triple is learned through the linear transformation matrix, and the calculation method is as follows:

[0018] c ijk =W[h i ||r k ||t j ]

[0019] Where W represents the linear transformation matrix;

[0020] Step 2.2: Input the linear transformation representation of the triplet and the linear transformation matrix into the nonlinear activation function to process and obtain the absolute attention coefficient b of the triplet. ijk , the calculation method is as follows:

[0021] b ijk =LeakyReLU(Vc ijk )

[0022] Wherein, LeakyReLU() represents a nonlinear activation function, and V represents the linear transformation matrix;

[0023] The absolute attention coefficient b ijk Input the softmax function to get the attention coefficient a of the triple ijk , the calculation method is as follows:

[0024]

[0025] Among them, N i is the neighboring node of entity i, r p represents the relationship between the link entity i and its adjacent node n, b inp represents the absolute attention coefficient between node i and its neighboring node n;

[0026] Step 2.3, sum the representations of each triple weighted by the attention coefficient, introduce a multi-head attention mechanism, multiply each adjacent node with its corresponding modified attention weight, and finally calculate the average value to obtain the embedding vector of the entity. The calculation method is as follows:

[0027]

[0028] Among them, M represents M independent attention calculation mechanisms;

[0029] Step 2.4: The entities and relations S of the triples modified by the attention mechanism α =(h α ,r,t α ) Embedded translation model processing;

[0030] For each triple, a specific transformation matrix M is introduced r , the head entity vector h α and tail entity vector t α Mapped to the corresponding relational space, we get the entity vector representation in the relational space, that is: h r =h α , t r =t α ;

[0031] Step 2.5: Replace the head entity or tail entity in the positive triple with another randomly selected entity to generate a negative triple, that is:

[0032] S'={(h',r,t)|h'∈E}∪{(h,r',t)|r'∈R}∪{(h,r,t')|t'∈E}

[0033] Among them, h' is the negative sample corresponding to the head node, r' is the negative sample corresponding to the relationship, and t' is the negative sample corresponding to the tail node; positive triplets and negative triplets are combined together to construct a data set for model training;

[0034] Step 2.6: Optimize the objective function L(h α ,r,t α ), obtain the embedding E of the triple entity based on the TransR model S ; The calculation method is as follows:

[0035]

[0036] Among them, γ represents the margin boundary parameter.

[0037] As a further improvement of the present invention, the step 3 specifically comprises the following steps:

[0038] Step 3.1: preprocess the preset entity description text information set D, use jieba to segment the text, remove stop words and irrelevant information, and obtain a corpus d = {w1, w2, ..., w m}, where w i represents the words in the corpus, and m is the total length of the text;

[0039] Step 3.2: Use the entity linking tool AIDA to annotate the entities in the corpus d and obtain a new corpus d' = {x1, x2, ..., x m}; where x is an entity in the knowledge graph or a word in the original corpus d;

[0040] Step 3.3: Construct a co-occurrence network G = (X, Y), where x i ∈X, represents a word or entity; y ij ∈Y, represents x i and the context word y i The co-occurrence frequency of

[0041] Step 3.4: Use the global vector word embedding method on the entity-annotated text corpus d' to learn contextual representations of entities and relations from text; the word x is obtained by minimizing the dot product between two word vectors and the squared difference of their co-occurrence times in the word co-occurrence network. i The word vector of ; the loss function is:

[0042]

[0043] where χ and ψ are the i and the context word y i The word vector of b i and b j is the bias term of the word vector, f(y ij ) is a weight function, which is used to adjust the weight of the co-occurrence frequency to prevent high-frequency words from excessively affecting the loss function; the calculation method is as follows:

[0044]

[0045] The stochastic gradient descent optimization algorithm is used to adjust the parameters of the word vector to minimize the loss function. After the training is completed, the representation χ∈R of each node x in the co-occurrence network G is obtained. k ;

[0046] Step 3.5: For each entity x i , the context of its text n(x i ) is defined as its neighborhood in the co-occurrence network, representing the words that frequently co-occur with the entity in the text corpus; the calculation method is as follows:

[0047] n(x i )={x j |y ij >θ}

[0048] Among them, θ is the threshold, and the adjacent nodes with co-occurrence frequency lower than θ are filtered;

[0049] x i The text context embedding is defined as n i The weighted average of the midpoint vectors; the weighted average is calculated as follows:

[0050]

[0051] If n(x i ) is empty, and η(x i ) is assigned a value of 0;

[0052] Step 3.6: Based on the co-occurrence network and text representation, build an entity representation model:

[0053] h d =η(h)M+h'

[0054] t d =η(t)M+t'

[0055] where h d and t d is the linear transformation of the text context embedding of h and t, M is a k×k matrix representing the weight of the text context; h' and t' represent deviations;

[0056] According to the entity representation embedding and TransR model, we obtain the embedding E based on the entity description information. d ; The calculation method is as follows:

[0057]

[0058] As a further improvement of the present invention, the step 4 specifically comprises the following steps:

[0059] Step 4.1: For a specific triple (h, r, t), the head entity mapping matrix is ​​calculated as:

[0060]

[0061] Among them, M c is the projection matrix of type c, C rh represents the set of relation types of the head entity under the relation r given by the relation-specific type information;

[0062] The calculation formula of the tail entity mapping matrix is:

[0063]

[0064] Among them, M c is the projection matrix of type c, C rt Represents the set of relation types of the tail entity under relation r given by relation-specific type information;

[0065] Step 4.2: Add the projection matrices with different weights to get the projection matrix M c , the calculation method is as follows:

[0066]

[0067] Where m is the number of levels of the hierarchical type, M c (i) Is type c (i) The projection matrix, β i It is c (i) The corresponding weights; where c (i) With c (i+1) There is a weight calculation method with decreasing proportion between them, and the calculation method is as follows:

[0068]

[0069] Among them, subtype c (i) The more accurate, the weight β i The larger the c (i) To M c The greater the impact;

[0070] Step 4.3: M rh 、M rt Multiply it with the triple entity embedding obtained by TransR to get the embedding E of the entity type l ; The calculation method is as follows:

[0071] h l =h s M rh ,t l =t s M rt

[0072]

[0073] As a further improvement of the present invention, the step 5 is specifically as follows:

[0074] Through the embedding E of the triple s , embedding of entity description information d , Description of entity type information E l , we get the embedding model integrating the auxiliary information of the knowledge graph of the market subject’s dishonesty information. The calculation method is as follows:

[0075] E=E s ⊕E d ⊕E l

[0076] Among them, ⊕ represents the concatenation operation;

[0077] The stochastic gradient descent method is used to optimize the training model, and the loss function is:

[0078]

[0079] Among them, h, r, t are the head entity, relation and tail entity respectively, γ represents the margin boundary parameter, E(h, r, t) is the overall energy function of the model, S represents the set of positive triples, S' represents the set of negative triples, which is obtained by randomly replacing the head entity or tail entity of the positive triple, that is:

[0080] S'={(h',r,t)|h'∈E}∪{(h,r',t)|r'∈R}∪{(h,r,t')|t'∈E}.

[0081] The beneficial effects of the present invention are:

[0082] In view of the challenges that data sparsity brings to the knowledge graph embedding performance of training and learning, especially the problem of being unable to accurately encode sparse entities, the present invention proposes a solution of introducing external auxiliary information. The present invention not only elaborates on how to model this auxiliary information, but also discusses how to effectively integrate it with the existing representation learning model. Through this improvement, the vectorized representation of entities and relationships can be optimized, thereby improving the accuracy in semantic association analysis. This strategy of combining external information aims to use richer data sources to make up for the shortcomings of original data, thereby improving the overall performance of knowledge graph embedding. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 is a flow chart of an embodiment of the present invention;

[0084] Figure 2 It is an overall framework diagram of an embodiment of the present invention;

[0085] Figure 3 This is a schematic diagram of the TransR model in an embodiment of the present invention;

[0086] Figure 4 is a schematic diagram of an attention correction method in an embodiment of the present invention;

[0087] Figure 5 It is a schematic diagram of the entity description representation method in an embodiment of the present invention;

[0088] Figure 6It is a schematic diagram of a hierarchical type representation method in an embodiment of the present invention. DETAILED DESCRIPTION

[0089] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0090] Example

[0091] like Figure 1 As shown, the embodiment of the present invention provides a knowledge representation learning method integrating auxiliary information of knowledge graph of dishonest information of market entities, including the steps of:

[0092] Step 1: Obtain the market entity’s dishonest information knowledge triples and related auxiliary information based on the market entity’s dishonest information knowledge graph.

[0093] Step 2: Based on the market entity’s default information knowledge triplet, the entities and relationships of the triplet itself are embedded into the representation learning model.

[0094] Step 3: vectorize all semantic information of the entity description according to the entity description information in the auxiliary information related to the market entity's default information.

[0095] Step 4: Vectorize the hierarchical type information according to the multiple types of entities existing in the auxiliary information related to the market entity's default information and the hierarchy of the types.

[0096] Step 5: Based on the triple information and related auxiliary information, the translation model is embedded and concatenated to train the representation learning model.

[0097] The overall framework of the knowledge representation learning method that integrates the auxiliary information of the knowledge graph of market subject dishonesty information is as follows: Figure 2 As shown in the figure, it specifically includes: for structured data, semantic encoding and structural encoding are performed through a representation learning model modified by an attention mechanism; for text description information, a corpus is obtained through preprocessing operations, and after entity annotation and text context embedding, the text context embedding is finally combined with the entity and relationship representation of the knowledge graph; for hierarchical type information, the hierarchical type is regarded as the projection matrix of the entity, and the hierarchical structure is modeled through the type encoder, so that the entity representation is projected and transformed through its hierarchical type. Finally, information fusion calculation is performed to obtain the final embedding vector.

[0098] The step 1 specifically includes:

[0099] Extract the structured information of the knowledge graph from the knowledge graph of market subject dishonesty information, and obtain the triple of market subject dishonesty information S = (h s ,r,t s); extract entity description information from the market subject dishonesty information corpus to obtain the Chinese entity description information set D; extract textual relationship information of entity pairs from the market subject dishonesty information corpus to obtain hierarchical information c = {c (1) ,c (2) ,…,c (m)}.

[0100] Step 2 is as follows Figure 3 , Figure 4 As shown, specifically including:

[0101] Step 2.1: Associate the entity of the triple with the relation S = (h s ,r,t s ) is modified through the attention mechanism.

[0102] For each specific triple (h i ,r k ,t j )∈S, the linear transformation representation of the triple is learned through the linear transformation matrix, and the calculation method is as follows:

[0103] c ijk =W[h i ||r k ||t j ]

[0104] Where W represents the linear transformation matrix.

[0105] Step 2.2: Input the linear transformation representation of the triplet and the linear transformation matrix into the nonlinear activation function to process and obtain the absolute attention coefficient b of the triplet ijk , the calculation method is as follows:

[0106] b ijk =LeakyReLU(Vc ijk )

[0107] Wherein, LeakyReLU() represents a nonlinear activation function, and V represents the linear transformation matrix.

[0108] The absolute attention coefficient b ijk Input the softmax function to get the attention coefficient a of the triple ijk , the calculation method is as follows:

[0109]

[0110] Among them, N i is the neighboring node of entity i, r p represents the relationship between the link entity i and its adjacent node n, b inprepresents the absolute attention coefficient between node i and its neighboring node n.

[0111] Step 2.3: Sum the representations of each triplet weighted by the attention coefficient, introduce a multi-head attention mechanism, multiply each adjacent node with its corresponding modified attention weight, and finally calculate the average value to obtain the embedding vector of the entity. The calculation method is as follows:

[0112]

[0113] Here, M represents M independent attention calculation mechanisms.

[0114] Step 2.4: The entities and relations S of the triples modified by the attention mechanism α =(h α ,r,t α ) Embedded translation model processing.

[0115] For each triple, a specific transformation matrix M is introduced r , the head entity vector h α and tail entity vector t α Mapped to the corresponding relational space, we get the entity vector representation in the relational space, that is: h r =h α , t r =t α .

[0116] Step 2.5: Replace the head entity or tail entity in the positive triple with other randomly selected entities to generate a negative triple, that is:

[0117] S'={(h',r,t)|h'∈E}∪{(h,r',t)|r'∈R}∪{(h,r,t')|t'∈E}

[0118] Among them, h' is the negative sample corresponding to the head node, r' is the negative sample corresponding to the relationship, and t' is the negative sample corresponding to the tail node.

[0119] Positive and negative triplets are combined together to construct a dataset for model training.

[0120] Step 2.6: Optimize the objective function L(h α ,r,t α ), obtain the embedding E of the triple entity based on the TransR model S The calculation method is as follows:

[0121]

[0122] Among them, γ represents the margin boundary parameter.

[0123] Step 3 is as follows Figure 5 As shown, specifically including:

[0124] Step 3.1: Preprocess the preset entity description text information set D, use Jieba to segment the text, remove stop words and irrelevant information. Get the corpus d = {w1, w2, ..., w m}, where w i Represents the words in the corpus, and m is the total length of the text.

[0125] Step 3.2: Use the entity linking tool AIDA to annotate the entities in corpus d and obtain a new corpus d' = {x1, x2, ..., x m}. Where x is an entity in the knowledge graph or a word in the original corpus d.

[0126] Step 3.3: Construct a co-occurrence network G = (X, Y), where x i ∈X, represents a word or entity; y ij ∈Y, represents x i and the context word y i The co-occurrence window is set to 64.

[0127] Step 3.4: Use the global vector word embedding method on the entity-annotated text corpus d' to learn contextual representations of entities and relations from text. The word x is obtained by minimizing the dot product between two word vectors and the squared difference of their co-occurrence counts in the word co-occurrence network. i The word vector of . The loss function is:

[0128]

[0129] where χ and ψ are the i and the context word y i The word vector of b i and b j is the bias term of the word vector, f(y ij ) is a weight function, which is used to adjust the weight of co-occurrence frequency to prevent high-frequency words from excessively affecting the loss function. The calculation method is as follows:

[0130]

[0131] Among them, take x max =100.

[0132] The stochastic gradient descent optimization algorithm is used to adjust the parameters of the word vector to minimize the loss function. After the training is completed, the representation χ∈R of each node x in the co-occurrence network G is obtained. k .

[0133] Step 3.5: For each entity x i , the context of its text n(x i ) is defined as its neighborhood in the co-occurrence network, representing the words that frequently co-occur with the entity in the text corpus. The calculation method is as follows:

[0134] n(x i )={x j |y ij >θ}

[0135] Among them, θ is the threshold, and the adjacent nodes with co-occurrence frequency lower than θ are filtered.

[0136] Based on the above expression, x i The text context embedding is defined as n i The weighted average of the node vectors in the . The weighted average is calculated as follows:

[0137]

[0138] If n(x i ) is empty, and η(x i ) is assigned a value of 0.

[0139] Step 3.6: Based on the co-occurrence network and text representation, build an entity representation model.

[0140] h d =η(h)M+h'

[0141] t d =η(t)M+t'

[0142] where h d and t d is the linear transformation of the text context embedding of h and t, M is a k×k matrix, which can be regarded as the weight of the text context; h' and t' represent deviations.

[0143] According to the entity representation embedding and TransR model, we obtain the embedding E based on the entity description information. d The calculation method is as follows:

[0144]

[0145] Step 4 is as follows Figure 6 As shown, specifically including:

[0146] Step 4.1: For a specific triple (h, r, t), the head entity mapping matrix is ​​calculated as:

[0147]

[0148] Among them, M c is the projection matrix of type c, C rh Represents the set of relation types of the head entity under relation r given by relation-specific type information.

[0149] The calculation formula of the tail entity mapping matrix is:

[0150]

[0151] Among them, M c is the projection matrix of type c, C rt Represents the set of relation types of the tail entity under relation r given by relation-specific type information.

[0152] Step 4.2: Add the projection matrices with different weights to get the projection matrix M c , the calculation method is as follows:

[0153]

[0154] Where m is the number of levels of the hierarchical type, M c (i) Is type c (i) The projection matrix, β i It is c (i) The corresponding weight. (i) With c (i+1) There is a weight calculation method with decreasing proportion between them, and the calculation method is as follows:

[0155]

[0156] The subtype c (i) The more accurate, the weight β i The larger the c (i) To M c The greater the impact.

[0157] Step 4.3: M rh 、M rt Multiply it with the triple entity embedding obtained by TransR to get the embedding E of the entity type l The calculation method is as follows:

[0158] h l =h s M rh ,t l =t s M rt

[0159]

[0160] The step 5 specifically includes:

[0161] The embedding E of the triple is obtained through step 2 s , through step 3, we get the embedding E of entity description information d , get the description of entity type information E through step 4 l , we get the embedding model integrating the auxiliary information of the knowledge graph of the market subject’s dishonesty information. The calculation method is as follows:

[0162] E=E s ⊕E d ⊕E l

[0163] Among them, ⊕ represents the concatenation operation.

[0164] The stochastic gradient descent method is used to optimize the training model, and the loss function is:

[0165]

[0166] Among them, h, r, t are the head entity, relation and tail entity respectively, γ represents the margin boundary parameter, E(h, r, t) is the overall energy function of the model, S represents the set of positive triples, S' represents the set of negative triples, which is obtained by randomly replacing the head entity or tail entity of the positive triple, that is:

[0167] S'={(h',r,t)|h'∈E}∪{(h,r',t)|r'∈R}∪{(h,r,t')|t'∈E}.

[0168] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A knowledge representation learning method integrating market subject dishonesty information knowledge graph auxiliary information, characterized in that: The following steps are involved: Step 1: According to the knowledge graph of the market subject's dishonest information, obtain the knowledge triplet related auxiliary information of the market subject's dishonest information; Step 2: Based on the knowledge triples, embed the entities and relations of the triples themselves into a representation learning model; Step 3: vectorize all semantic information of the entity description according to the entity description information in the relevant auxiliary information; Step 4: vectorize the hierarchical type information according to the multiple types of entities in the relevant auxiliary information and the hierarchy of the types; Step 5: Based on the knowledge triples and related auxiliary information, the translation model is embedded in combination to perform concatenation and train the representation learning model.

2. The knowledge representation learning method integrating market subject dishonesty information knowledge graph auxiliary information according to claim 1 is characterized in that: The step 1 is specifically as follows: Extract the structured information of the knowledge graph from the knowledge graph of market subject dishonesty information, and obtain the knowledge triple S = (h s ,r,t s ); extract entity description information from the market entity dishonesty information corpus to obtain the Chinese entity description information set D; Extract the textual relationship information of entity pairs from the market subject dishonesty information corpus, and obtain hierarchical information c = {c (1) ,c (2) ,…,c (m) }.

3. The knowledge representation learning method integrating market subject dishonesty information knowledge graph auxiliary information according to claim 2 is characterized in that: The step 2 specifically includes the following steps: Step 2.1: Connect the entity of the triple to the relation S = (h s ,r,t s ) is modified through the attention mechanism; For each specific triple (h i ,r k ,t j )∈S, the linear transformation representation of the triple is learned through the linear transformation matrix, and the calculation method is as follows: c ijk =W[h i ||r k ||t j ] Where W represents the linear transformation matrix; Step 2.2: Input the linear transformation representation of the triplet and the linear transformation matrix into the nonlinear activation function to process and obtain the absolute attention coefficient b of the triplet. ijk , the calculation method is as follows: b ijk =LeakyReLU(Vc ijk ) Wherein, LeakyReLU() represents a nonlinear activation function, and V represents the linear transformation matrix; The absolute attention coefficient b ijk Input the softmax function to get the attention coefficient a of the triple ijk , the calculation method is as follows: Among them, N i is the neighboring node of entity i, r p represents the relationship between the link entity i and its adjacent node n, b inp represents the absolute attention coefficient between node i and its neighboring node n; Step 2.3, sum the representations of each triple weighted by the attention coefficient, introduce a multi-head attention mechanism, multiply each adjacent node with its corresponding modified attention weight, and finally calculate the average value to obtain the embedding vector of the entity. The calculation method is as follows: Among them, M represents M independent attention calculation mechanisms; Step 2.4: The entity and relationship S of the triple modified by the attention mechanism α =(h α ,r,t α ) Embedded translation model processing; For each triple, a specific transformation matrix M is introduced r , the head entity vector h α and tail entity vector t α Mapped to the corresponding relational space, we get the entity vector representation in the relational space, that is: h r =h α , t r =t α ; Step 2.5: Replace the head entity or tail entity in the positive triple with another randomly selected entity to generate a negative triple, that is: S'={(h',r,t)|h'∈E}∪{(h,r',t)|r'∈R}∪{(h,r,t')|t'∈E} Among them, h' is the negative sample corresponding to the head node, r' is the negative sample corresponding to the relationship, and t' is the negative sample corresponding to the tail node; positive triplets and negative triplets are combined together to construct a data set for model training; Step 2.6: Optimize the objective function L(h α ,r,t α ), obtain the embedding E of the triple entity based on the TransR model S ; The calculation method is as follows: Among them, γ represents the margin boundary parameter.

4. The knowledge representation learning method integrating market subject dishonesty information knowledge graph auxiliary information according to claim 3 is characterized in that: The step 3 specifically comprises the following steps: Step 3.1: pre-process the preset entity description text information set D, use jieba to segment the text, remove stop words and irrelevant information, and obtain the corpus d = {w1, w2, ..., w m }, where w i represents the words in the corpus, and m is the total length of the text; Step 3.2: Use the entity linking tool AIDA to annotate the entities in the corpus d and obtain a new corpus d' = {x1, x2, ..., x m }; where x is an entity in the knowledge graph or a word in the original corpus d; Step 3.3: Construct a co-occurrence network G = (X, Y), where x i ∈X, represents a word or entity; y ij ∈Y, represents x i and the context word y i The co-occurrence frequency of Step 3.4: Use the global vector word embedding method on the entity-annotated text corpus d' to learn contextual representations of entities and relations from text; the word x is obtained by minimizing the dot product between two word vectors and the squared difference of their co-occurrence times in the word co-occurrence network. i The word vector of ; the loss function is: where χ and ψ are the i and the context word y i The word vector of b i and b j is the bias term of the word vector, f(y ij ) is a weight function, which is used to adjust the weight of the co-occurrence frequency to prevent high-frequency words from excessively affecting the loss function; the calculation method is as follows: The stochastic gradient descent optimization algorithm is used to adjust the parameters of the word vector to minimize the loss function. After the training is completed, the representation χ∈R of each node x in the co-occurrence network G is obtained. k ; Step 3.5: For each entity x i , the context of its text n(x i ) is defined as its neighborhood in the co-occurrence network, representing the words that frequently co-occur with the entity in the text corpus; the calculation method is as follows: n(x i )={x j |y ij >θ} Among them, θ is the threshold, and the adjacent nodes with co-occurrence frequency lower than θ are filtered; x i The text context embedding is defined as n i The weighted average of the midpoint vectors; the weighted average is calculated as follows: If n(x i ) is empty, and η(x i ) is assigned a value of 0; Step 3.6: Based on the co-occurrence network and text representation, build an entity representation model: h d =η(h)M+h' t d =η(t)M+t' where h d and t d is the linear transformation of the text context embedding of h and t, M is a k×k matrix representing the weight of the text context; h' and t' represent deviations; According to the entity representation embedding and TransR model, we obtain the embedding E based on the entity description information. d ; The calculation method is as follows:

5. The knowledge representation learning method of integrating market subject dishonesty information knowledge graph auxiliary information according to claim 4 is characterized in that: The step 4 specifically comprises the following steps: Step 4.1: For a specific triple (h, r, t), the head entity mapping matrix is ​​calculated as: Among them, M c is the projection matrix of type c, C rh represents the set of relation types of the head entity under the relation r given by the relation-specific type information; The calculation formula of the tail entity mapping matrix is: Among them, M c is the projection matrix of type c, C rt Represents the set of relation types of the tail entity under relation r given by relation-specific type information; Step 4.2: Add the projection matrices with different weights to get the projection matrix M c , the calculation method is as follows: Where m is the number of levels of the hierarchical type, M c (i) Is type c (i) The projection matrix, β i It is c (i) The corresponding weights; where c (i) With c (i+1) There is a weight calculation method with decreasing proportion between them, and the calculation method is as follows: Among them, subtype c (i) The more accurate, the weight β i The larger the c (i) To M c The greater the impact; Step 4.3: M rh 、M rt Multiply it with the triple entity embedding obtained by TransR to get the embedding E of the entity type l ; The calculation method is as follows: h l =h s M rh ,t l =t s M rt 6. The knowledge representation learning method of integrating market subject dishonesty information knowledge graph auxiliary information according to claim 5 is characterized in that: The step 5 is specifically as follows: Through the embedding E of the triple s , embedding of entity description information d , Description of entity type information E l , we get the embedding model integrating the auxiliary information of the knowledge graph of the market subject’s dishonesty information. The calculation method is as follows: in, represents the concatenation operation; The stochastic gradient descent method is used to optimize the training model, and the loss function is: Among them, h, r, t are the head entity, relation and tail entity respectively, γ represents the margin boundary parameter, E(h, r, t) is the overall energy function of the model, S represents the set of positive triples, S' represents the set of negative triples, which is obtained by randomly replacing the head entity or tail entity of the positive triple, that is: S'={(h',r,t)|h'∈E}∪{(h,r',t)|r'∈R}∪{(h,r,t')|t'∈E}.

Images