In particular, a device, data structure, and computer-implemented method for determining a knowledge graph for performing knowledge graph inference.
The method enriches knowledge graph embeddings with numerical aggregates and permutations to improve link prediction accuracy by incorporating literal values, addressing the limitations of existing methods.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- ROBERT BOSCH GMBH
- Filing Date
- 2025-09-30
- Publication Date
- 2026-04-13
AI Technical Summary
Existing knowledge graph inference methods lack the ability to effectively incorporate numerical information from literal values into relationship embeddings, leading to suboptimal link prediction accuracy in various environments.
A method for determining a knowledge graph that aggregates literal values separately for entities and relations, generating distinct numerical aggregates which are combined with relationship embeddings to enrich correlation information, and uses permutations to enhance inference accuracy.
Enhances the accuracy of link prediction in knowledge graphs by leveraging numerical correlations between attributes, improving performance in manufacturing and other environments.
Smart Images
Figure 2026064233000001_ABST
Abstract
Description
Technical Field
[0001] Background The present invention relates to an apparatus, a data structure, and a computer-implemented method for determining a knowledge graph for performing knowledge graph inference, in particular.
Summary of the Invention
Problems to be Solved by the Invention
[0002] Disclosure of the Invention The computer-implemented method and apparatus provide relationship-centric enhancement for knowledge graph inference.
Means for Solving the Problems
[0003] A method for determining a knowledge graph for performing knowledge graph inference includes providing a set of entities, a set of relations, and a set of attributes in an embedding space; providing a first set of triples, where each triple in the first set includes a first entity from the set of entities, a relations from the set of relations, and a last entity from the set of entities; providing a second set of triples, where each triple in the second set includes an entity from the set of entities, an attribute from the set of attributes, and a literal value; and the method, with respect to an entity from the set of entities, determines, for each entity from the set of entities, an association including a literal value from the triple in the second set of triples that contains each entity; and the association from the set of relations. For each relation, this includes determining a first aggregation of literal values for each literal value, depending on the literal values associated within the relation, for the entity that contains as the leading entity in the first set of triples containing each relation; determining a second aggregation of literal values for each relation from the set of relations, depending on the literal values associated within the relation, for the entity that contains as the trailing entity in the first set of triples containing each relation; determining a permutation for each relation in the embedding space, depending on the relation, the first aggregation determined for each relation, and the second aggregation determined for each relation; and permuting each relation within the triples in the first set of triples by the permutation determined for each relation.
[0004] This means that the method incorporates numerical information from literal values into the embedding of connection relationships in the knowledge graph. The literal values of attributes are aggregated separately with respect to the domain of the relationship to generate two distinct numerical aggregates, namely the first aggregate and the second aggregate. The aggregated numerical information is then combined with the relationship embedding to form the relationship substitution. Such embedding enriches the relationship embedding with information about possible correlations between attributes.
[0005] The method may involve determining a matrix that defines associations, the matrix containing entries for each entity in the set of entities and each attribute in the set of attributes, the entries for each entity contain the literal value of the attribute for each entity if the second set of triples contains the triple for each entity and each attribute, or, if not, indicate in particular by the value of zero that the second set of triples does not contain the triple for each entity and each attribute.
[0006] The method may include determining a first vector for each relation that defines a first aggregate of literal values, wherein the first vector includes an aggregate of row entries in a matrix with respect to entities included as the first entity by at least one triple containing each relation from a first set of triples, and determining a second vector for each relation that defines a second aggregate of literal values, wherein the second vector includes an aggregate of row entries in a matrix included as the last entity by at least one triple containing each relation from a first set of triples.
[0007] Multiple aggregation methods can be used. According to this method, for example, determining the aggregation of literal values includes determining the mean, median, mode, minimum, maximum, sum, total, range, interquartile range, variance, or standard deviation of the distribution of literal values in the same column of a row vector.
[0008] Aggregation can be based on a linear combination of several different aggregation methods. Determining the aggregation of literal values involves, for example, determining a linear combination of at least two of the following: mean, median, mode, minimum, maximum, sum, total, range, interquartile range, variance, or standard deviation of the distribution of the literal values.
[0009] Linear combinations can be treated as hyperparameters for machine learning. This method involves learning at least one parameter that defines a linear combination, depending on training data containing, for example, a first set of triples and a second set of triples.
[0010] This method may involve performing knowledge graph inference on triples containing permutations from a first set of triples, in order to predict the validity of triples containing a given first entity, a given last entity, and a given relationship corresponding to an existing permutation. The permutation embedding contains information about possible correlations between attributes. This means that inference on triples containing permutation embeddings is enhanced by information about possible correlations between attributes.
[0011] Knowledge graphs can be adapted for use in manufacturing environments. An example of a manufacturing environment is a production line, particularly for welding, for example, for automotive welding. For example, a set of entities includes entities representing sensors on a production line and entities representing workstations on a production line, a set of relationships includes the following relationships, namely, relationships indicating that a workstation represented by an entity linked to an entity representing a sensor has that sensor, the method includes providing a set of range-indicating literal values, and a set of attributes includes the following attributes, namely, attributes indicating that a sensor represented by an entity linked to one of the literal values in the set of literal values has the range indicated by that literal value. In the case of automotive welding, the link prediction task in knowledge graph inference may include finding which sensors, in particular which sensors providing measurements of welding spots, belong to which workstations, in particular which car bodies being welded within the workstations. In this context, there is a large amount of numerical data from multiple sensors, each measuring a different physical quantity. These sensors may each have a different range, i.e., a measurement range. Attributes and literal values representing the range are linked to the entity representing the sensor. This allows the knowledge graph to leverage these ranges to improve the accuracy of link predictions.
[0012] Knowledge graphs can also be adapted for use in other environments. For example, a set of entities may include entities representing people and entities representing houses, a set of relationships may include the following relationships, namely, relationships that indicate that a person represented by an entity linked to a house rents that house, the method includes providing a first set of literal values indicating monthly rent, and a set of attributes may include the following first attribute, namely, the first attribute may be linked to a first literal value in the first set of literal values via the first attribute A first attribute to indicate that the house represented by the linked entity is subject to a monthly rent indicated by the first literal value, and / or the method includes providing a second set of literal values indicating monthly income, the set of attributes including the following second attribute, namely, the second attribute to indicate that the person represented by the entity linked to the second literal value in the second set of literal values is responsible for the monthly rent indicated by the second literal value.
[0013] In particular, the device for determining a knowledge graph for performing knowledge graph inference comprises at least one processor and at least one memory, the at least one memory storing instructions for causing the device to perform the method when executed by at least one processor.
[0014] We can provide a computer program that, when executed by a computer, includes computer-readable instructions for causing the computer to perform this method.
[0015] A data structure comprising: at least one data field for embedding in an embedding space of a set of entities, a set of relationships, a set of attributes, and at least one set of literal values; a first set of triples, each of which triples in the first set comprises at least one data field for the first set of triples, containing a leading entity from the set of entities, a relationship from the set of relationships, and a trailing entity from the set of entities; a second set of triples, each of which triples in the second set comprises at least one data field for the second set of triples, containing an entity from the set of entities, an attribute from the set of attributes, and a literal value; and, with respect to an entity from the set of entities, for each entity from the set of entities, at least one data field for relationships from the second set of triples, determined to contain a literal value from the triple in which each entity is contained. A data structure can be provided that includes one more data field, and for each relation from the set of relations, at least one data field for each first aggregation of literal values determined depending on the literal values associated within the relation, for the entity that contains as the leading entity from the first set of triples containing each relation, and for each relation from the set of relations, at least one data field for each second aggregation of literal values determined depending on the literal values associated within the relation, for the entity that contains as the trailing entity from the first set of triples containing each relation, and for each relation from the set of relations, at least one data field for each second aggregation of literal values determined depending on the literal values associated within the relation, and at least one data field for each relation in the embedding space for each relation, determined depending on the relation, the first aggregation (lh) determined for each relation, and the second aggregation (lt) determined for each relation.
[0016] Further examples can be derived from the following explanation and diagrams. [Brief explanation of the drawing]
[0017] [Figure 1] This diagram schematically shows a device for determining a knowledge graph, particularly for performing knowledge graph inference. [Figure 2] This flowchart specifically includes each step of the method for determining the knowledge graph necessary for performing knowledge graph inference. [Figure 3] This diagram provides a schematic representation of the data structure. [Modes for carrying out the invention]
[0018] Figure 1 schematically shows the device 100 for determining the knowledge graph.
[0019] The device 100 may be configured to perform knowledge graph inference.
[0020] The device 100 includes at least one processor 102 and at least one memory 104.
[0021] The device 100 may include at least one interface 106.
[0022] At least one memory 104 stores instructions that cause the device 100 to perform a method for determining a knowledge graph, in particular for performing knowledge graph inference, when executed by at least one processor 102.
[0023] A knowledge graph is based on a set of entities e E, a set of relations r R, a set of attributes A, and a set of literal values V within the embedding space of the knowledge graph.
[0024] The knowledge graph contains a first set of triples. Each triple in the first set contains a first entity h from the set of entities E, a relation r from the set of relations R, and a last entity e from the set of entities E.
[0025] The knowledge graph contains a second set of triples. Each triple in the second set contains an entity e from the set of entities E, an attribute a from the set of attributes A, and a literal value v from the set of literal values V.
[0026] According to the first example, the set of entities E includes entities e representing sensors on a manufacturing line and entities e representing workstations on a manufacturing line.
[0027] According to the first example, the set of relations R includes the following relation r, namely, the relation r is a relation that indicates that a workstation represented by an entity linked to an entity representing a sensor has that sensor.
[0028] According to the first example, the set of literal values V represents a range.
[0029] According to the first example, attribute set A includes attribute a, namely attribute a which indicates that a sensor represented by entity e linked via attribute a to one literal value v from a set of literal values V has the range indicated by that literal value v.
[0030] According to the second example, the set of entities E includes entities e representing people and entities e representing houses.
[0031] According to the second example, the set of relations R includes the following relation r, namely, relation r is a relation that indicates that a person represented by an entity linked to an entity representing a house rents that house.
[0032] In the second example, two sets V of literal values are provided.
[0033] In other words, a first set V1 of literal values representing monthly rent and a second set V2 of literal values representing monthly income are provided.
[0034] According to the second example, attribute set A includes the following first attribute a1, which is a first attribute a1 that indicates that a house represented by an entity linked via the first attribute a1 to a first literal value v1 in a first set of literal values V1 has a monthly rent indicated by the first literal value v1.
[0035] According to the second example, attribute set A includes the following second attribute a2, namely, the second attribute a2 is a second attribute a2 that indicates that a person represented by an entity linked via the second attribute a2 to a second literal value v2 in a second set of literal values V2 is responsible for the monthly rent indicated by the second literal value v2.
[0036] Figure 2 shows a flowchart that includes the multiple steps of this method.
[0037] This method includes step 200.
[0038] Step 200 is performed within the embedded space, Set E of entities e, Set R of relation r, Set A of attribute a This includes providing.
[0039] Step 200 includes providing a first set of triples and a second set of triples.
[0040] This method includes step 202.
[0041] Step 202 includes determining, for each entity e from the set E of entities, a relationship L that includes a literal value v from the triples of the second set of triples in which each entity e is included, for each entity e from the set E of entities.
[0042] The relationship L may be defined as being by the matrix
Number
[0043] This method may include determining a matrix L that defines the relationship L.
[0044] The matrix is determined to include, for example, an entry L ik The entry L i for entity e k and attribute a ik is determined to include, for example, the literal value v of attribute a i for entity e k if the second set of triples includes a triple with entity e i and attribute a k or, if not, indicates that the second set of triples does not include a triple with entity e i and attribute a k [[ID= (39)]]using a value of zero in particular.
[0045] That is, the entry L i for each entity e in the set E of entities k and each attribute a in the set A of attributes ik includes the literal value v of attribute a i )]]for each entity e k if the second set of triples includes a triple with each entity e i and each attribute a k or, if not, indicates that the second set of triples does not include a triple with each entity ei and each attribute a k This means that the absence of a triple containing both elements is specifically indicated using the value of zero.
[0046] This method includes step 204.
[0047] Step 204 involves determining, for each relation r from the set of relations R, a first aggregate lh of literal values for each entity e from a first set of triples, where at least one triple containing each relation contains the entity e as the leading entity h, depending on the literal values v associated within relation L.
[0048] Step 204 involves determining a second aggregate lt for each relation r from the set of relations R, depending on the literal value v associated within relation L, for each entity e from the first set of triples, where at least one triple containing each relation r contains the entity e as a trailing entity t.
[0049] Aggregation can be considered as a vector with the same dimensions as the relationship within the embedding space.
[0050] Determining the first aggregation involves defining a first vector that defines the first aggregation lh of literal values for each relation r.
number
[0051] The first vector l h This includes, for example, an aggregation of row entries in matrix L such that, of the first set of triples, each relation is contained in matrix L with respect to entity e which contains as the leading entity h.
[0052] Determining the second aggregation involves defining a second vector for each relation r that defines the second aggregation lt of literal values.
number
[0053] The second vector l t This includes, for example, an aggregation of row entries in matrix L such that, of the first set of triples, each relation is contained in matrix L with respect to entity e which contains as a trailing entity t.
[0054] The first vector l h or the second vector l t This can be considered determined as a row vector.
[0055] Determining the aggregation of each literal value is done by the first vector l h or the second vector l t This may include determining the average of the literal values in the same column of the row vector as the aggregated value in each column.
[0056] Determining the aggregation of each literal value is done by the first vector l h or the second vector l t This may include determining the median of the literal values in the same column of the row vector as the aggregated value in each column.
[0057] Determining the aggregation of each literal value is done by the first vector l h or the second vector l t This may include determining the mode of the literal values in the same column of the row vector as the aggregated value in each column.
[0058] Determining the aggregation of each literal value is done by the first vector l h or the second vector l tThe aggregated value in each column may include determining the minimum value of the literal values in the same column of the row vector.
[0059] Determining the aggregation of each literal value is done by the first vector l h or the second vector l t The aggregated value in each column may include determining the maximum value of the literal values in the same column of the row vector.
[0060] Determining the aggregation of each literal value is done by the first vector l h or the second vector l t This may include determining the sum of the literal values in the same column of the row vector as the aggregated value in each column.
[0061] Determining the aggregation of each literal value is done by the first vector l h or the second vector l t This may include determining the total number of literal values in the same column of the row vector as the aggregated value in each column.
[0062] Determining the aggregation of each literal value is done by the first vector l h or the second vector l t This may include determining the range of literal values in the same column of a row vector as an aggregated value in each column, for example, as a two-dimensional aggregated value that includes two boundaries of the range.
[0063] Determining the aggregation of each literal value is done by the first vector l h or the second vector l t This may include determining the interquartile range of the literal values in the same column of the row vector as the aggregated value in each column.
[0064] Determining the aggregation of each literal value is done by the first vector l h or the second vector l tThis may include determining the variance of the literal values in the same column of the row vector as the aggregated value in each column.
[0065] Determining the aggregation of each literal value is done by the first vector l h or the second vector l t This may include determining the standard deviation of the distribution of literal values in the same column of the row vector as an aggregated value in each column.
[0066] Determining the aggregation of literal values may involve determining a linear combination of at least two of the following: mean, median, mode, minimum, maximum, sum, total, range, interquartile range, variance, or standard deviation of the distribution of the literal values.
[0067] Linear combinations are, y=σ(UW a +b1) It is often defined as follows, where,
number
number
number
number
[0068] Rows u and aggregate functions for N attributes
number
number
[0069] This method depends on the training data, which includes a first set of triples and a second set of triples, and defines at least one parameter, e.g., W, that defines a linear combination. a This may include learning b1.
[0070] This method includes step 206.
[0071] Step 206 includes determining the substitution for each relation r in the embedding space, depending on the relation r, a first aggregate lh determined for the relation r, and a second aggregate lt determined for the relation r.
[0072] For example, each relationship
number
number
[0073] function g lin teeth,
number
number
number
[0074] This method depends on the training data which contains a first set of triples and a second set of triples, and uses the function g lin At least one parameter that defines it, for example, W r This may include learning b2.
[0075] Substitution is a gated function
number
number
number
number
[0076] For example, each relationship
number
number
[0077] This method depends on the training data which contains a first set of triples and a second set of triples, and uses the function g gated At least one parameter that defines it, for example, W zr ,W zlh ,W zlt , and / or, may include learning b3.
[0078] Learning may involve learning parameters that result in the substitution of relationships from the set of relationships R that improve the quality of link prediction, with respect to the quality of link prediction achieved using triples in the knowledge graph that contain relationships from the set of relationships R.
[0079] This method includes step 208.
[0080] Step 208 is to determine each relation r within the triples in the first set of triples by the permutation r determined for each relation r. lit This includes replacing with.
[0081] function g lin or function g gated The output is a vector with the same dimensions as relation r.
[0082] function g lin or function g gated The resulting vector is a literal-enhanced embedding vector that can be used in place of the original embedding vector of relation r within the scoring function for link prediction.
[0083] For example, relation r i That is, each embedding vector is r r,i By, especially r r,i =g lin (l h ,r i ,l t ) or r r,i =g gated (l h ,r i ,l t) is replaced by.
[0084] This method may include step 210.
[0085] Step 210 involves performing knowledge graph inference on the triples containing permutations from a first set of triples in order to predict the validity of triples that include a given leading entity, a given trailing entity, and a given relationship corresponding to an existing permutation.
[0086] According to the first example, the presence of a weld spot on a vehicle body within a workstation is detected when a link is predicted between an entity representing a sensor that provides a measurement of a weld spot and an entity representing a workstation where the vehicle body is located, using knowledge graph inference on a triple that includes a substitution determined with respect to the first set of triples according to the first example, among the first set of triples.
[0087] In the second example, the person who rents the house is determined when a link between an entity representing a person and an entity representing a house is predicted using knowledge graph inference on a triple that contains the permutations determined with respect to the first set of triples according to the second example, among the first set of triples.
[0088] Figure 3 shows a schematic representation of data structure 300.
[0089] The data structure 300 includes at least one data field 302 for embedding in the embedding space a set of entities, a set of relationships, a set of attributes, and at least one set of literal values.
[0090] The data structure 300 includes at least one data field 302 for a first set of triples and at least one data field 302 for a second set of triples.
[0091] The data structure 300 includes at least one data field 302 for association.
[0092] The data structure 300 includes at least one data field 302 for each first aggregation and at least one data field 302 for each second aggregation.
[0093] The data structure 300 includes at least one data field 302 for substitution for each relationship in the embedding space.
Claims
1. In particular, a computer-implemented method for determining a knowledge graph for performing knowledge graph inference, The aforementioned method, (200) Provide a set of entities, a set of relationships, and a set of attributes within the embedded space. (200) to provide a first set of triples, wherein each triple in the first set includes a first entity from the set of entities, a relation from the set of relations, and a last entity from the set of entities. (200) provides a second set of triples, wherein each triple in the second set includes an entity from the set of entities, an attribute from the set of attributes, and a literal value. Includes, The aforementioned method, With respect to the entities from the set of entities, for each entity from the set of entities, determine an association that includes a literal value from the triple in the second set of triples that contains each entity (202), For each relationship from the set of relationships, the first aggregation of the literal values for each of the literal values is determined for the entity that includes as the leading entity in the first set of triples, where at least one triple in the first set of triples contains each of the relationships, depending on the literal values associated within the relationship (204). For each relationship from the set of relationships, the second aggregation of each literal value is determined for each entity that is included as the trailing entity by at least one triple in the first set of triples that contains each relationship, depending on the literal value associated within the relationship (204), The substitution for each of the relationships in the embedded space is determined depending on the relationship, the first aggregation determined for each of the relationships, and the second aggregation determined for each of the relationships (206), (208) Substituting each of the relationships within the triples in the first set of triples with the permutation determined for each of the relationships, A method characterized by including
2. The method includes determining the matrix that defines the association (202), The aforementioned matrix includes entries, The entries relating to each entity from the set of entities and each attribute from the set of attributes include, if the second set of triples includes triples comprising each entity and each attribute, the literal value of the attribute relating to each entity, or, if not, the fact that the second set of triples does not include triples comprising each entity and each attribute, is indicated, in particular by the value of zero. The method according to claim 1.
3. The aforementioned method, (204) For each of the aforementioned relationships, a first vector is determined that defines a first aggregate of the literal values, wherein the first vector includes an aggregate of row entries of the matrix, wherein the matrix is included with respect to entities that are included as the leading entity by at least one triple in the first set of triples that contains each of the aforementioned relationships (204), (204) Determining a second vector that defines a second aggregation of the literal values for each of the aforementioned relationships, wherein the second vector includes an aggregation of row entries of the matrix, wherein the matrix is included with respect to entities that are included as the trailing entities by at least one triple in the first set of triples that contains each of the aforementioned relationships (204), The method according to claim 2, including the method described in claim 2.
4. Determining the aggregation of the literal values (204) includes determining the mean, median, mode, minimum, maximum, sum, total, range, interquartile range, variance, or standard deviation of the distribution of the literal values in the same column of the row vector. The method according to claim 3.
5. Determining the aggregation of the literal values (204) includes determining a linear combination of at least two of the mean, median, mode, minimum, maximum, sum, total, range, interquartile range, variance, or standard deviation of the distribution of the literal values. The method according to claim 4.
6. The method includes learning at least one parameter that defines the linear combination, depending on the training data which includes a first set of triples and a second set of triples (204), The method according to claim 3 or 4.
7. The method includes performing knowledge graph inference on the triples containing the permutation from a first set of triples in order to predict the validity of triples containing a given leading entity, a given trailing entity, and a given relationship corresponding to the existing permutation (210), The method according to any one of claims 1 to 6.
8. The set of entities includes entities representing sensors on a manufacturing line and entities representing workstations on a manufacturing line, The set of relationships includes the following relationships, namely, relationships that indicate that a workstation represented by an entity linked to an entity representing a sensor has that sensor: The method includes providing a set of literal values that indicate a range, The set of attributes includes the following attributes, namely, attributes that indicate that a sensor represented by an entity linked via that attribute to one of the literal values in the set of literal values has the range indicated by that literal value. The method according to any one of claims 1 to 7.
9. The aforementioned set of entities includes entities representing people and entities representing houses, The set of relationships includes the following relationships, namely, relationships that indicate that a person represented by an entity linked to an entity representing a house rents that house: The method includes providing a first set of literal values representing monthly rent, The set of attributes includes the following first attribute, namely, the first attribute is a first attribute that indicates that a house represented by an entity linked to a first literal value in the first set of literal values via the first attribute is subject to the monthly rent indicated by the first literal value. and / or, The method includes providing a second set of literal values representing monthly income, The set of attributes includes the following second attribute, namely, the second attribute is a second attribute that indicates that a person represented by an entity linked to a second literal value in the second set of literal values via the second attribute is responsible for the monthly rent indicated by the second literal value. The method according to any one of claims 1 to 7.
10. In particular, a device (100) for determining a knowledge graph for performing knowledge graph inference, The aforementioned device (100) At least one processor (102), At least one memory (104), Equipped with, The at least one memory stores instructions, when executed by the at least one processor, that cause the device to carry out the method according to any one of claims 1 to 9. A device (100) characterized by the following.
11. It is a computer program, The computer program, when executed by a computer, includes computer-readable instructions to cause the computer to perform the method according to any one of claims 1 to 9. A computer program characterized by the following features.
12. A data structure (300), wherein the data structure is A data field (302) for embedding in the embedding space of a set of entities, a set of relationships, a set of attributes, and at least one set of literal values, A first set of triples, each of which triples in the first set comprises at least one data field (302) for the first set of triples, including a first entity from the set of entities, a relation from the set of relations, and a last entity from the set of entities. A second set of triples, each of which triples in the second set comprises at least one data field (302) for the second set of triples, which includes entities from the set of entities, attributes from the set of attributes, and literal values. With respect to the entities from the set of entities, for each entity from the set of entities, there is at least one data field (302) for association which is determined to contain a literal value from the triple in which each entity is included, For each relationship from the set of relationships, the entity that includes as the leading entity the first set of triples that contains each of the relationships has at least one data field (302) for each first aggregation of the literal values, determined depending on the literal values associated within the relationship, For each relationship from the set of relationships, the entity that includes as the trailing entity at least one triple in the first set of triples that contains each of the relationships has at least one data field (302) for a second aggregation of the literal values, determined depending on the literal values associated within the relationship, A data field (302) for substitution for each of the aforementioned relationships in the embedding space, determined depending on each of the aforementioned relationships, the first aggregation determined for each of the aforementioned relationships, and the second aggregation determined for each of the aforementioned relationships, A data structure (300) characterized by including