Devices, computer-implemented methods, and computer programs for automated data analysis.

Abstract

To provide a computer-implemented method for automatic analysis of a resource description framework (RDF) dataset (D), a device, and a computer program.SOLUTION: An RDF dataset (D) contains a sets of triples. The RDF dataset (D) is provided as a non-directed graph (KG) containing nodes (N) and edges (E). A node (N) indicates an entity (e), and an edge (e) indicates a link between entities (e).SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] Background of the Invention The present invention relates to an apparatus, computer program, and computer-implemented method for automated data analysis. [Background technology]

[0002] Knowledge graphs (KGs) can be used to automatically analyze data. The results of the data analysis can be determined automatically. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Shi, Y., Cheng, G., Kharlamov, E., “Keyword search over knowledge graphs via static and dynamic hub labellings”, WWW 2020.pp.235-245(2020) [Non-Patent Document 2] Cheng, G., Jin, C., Ding, W., Xu, D., Qu, Y., “Generating illustrative snippets for open data on the web”, WSDM2017.pp.151-159 (2017) [Non-Patent Document 3] Wang, X., Cheng, G., Kharlamov, E., “Towards multi-facet snippets for dataset search”, PROFLILES&SemEx2019.pp.1-6(2019) [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] Summary of the Invention The apparatus, method, and computer program described in the independent claim further improve automated analysis. [Means for solving the problem]

[0005] This disclosure relates to a computer-implemented method for automated analysis of Resource Description Framework (RDF) datasets, where the RDF dataset comprises a set of triples, and the RDF dataset is provided as an undirected graph (D) containing nodes and edges, where nodes represent entities and edges represent links between entities. The method describes - For each instance-level entity, at least one entity description pattern including at least one triple describing at least one class and / or at least one property of the entity, - For each edge representing an entity link between two instance-level entities, at least one link description pattern containing at least one triple that describes the link between the two instance-level entities. The steps to generate, The steps include generating an entity link graph by converting the edges of an undirected graph into nodes, and Steps to generate subgraphs of the entity link graph, The steps include generating an extended subgraph by adding missing links to each node of the subgraph representing an entity link so that each node of the extended subgraph representing an entity link is connected to both entities it links to, -For each node in the extended subgraph representing an entity from the entity description pattern, at least one triple describing the class of the entity, and at least one triple describing the properties of the entity for each property of the entity description pattern, - For each node in the extended subgraph representing an entity link, at least one triple from the link description pattern The step of generating a representative subset of the RDF dataset from the extended subgraph by adding the following: Includes.

[0006] A typical subset of an RDF dataset is a pattern coverage snippet. Understanding the contents of an RDF dataset is a prerequisite for reusing it. Existing methods generate a summarized version of an RDF dataset primarily by extracting representative data patterns as summaries, in order to help understand its large and complex structure. As a supplement, recent attempts have extracted representative subsets of specific data as snippets. Snippets and summaries provide a complementary view of the RDF dataset, namely, snippets containing representative triples at the instance level and summaries containing representative patterns at the schema level. According to the present invention, the strength of the summary is injected into the snippet by generating a pattern coverage snippet that best exemplifies the patterns of entity descriptions and links within the RDF dataset.

[0007] Preferably, when converting the edges of an undirected graph into nodes to generate an entity link graph, all nodes containing the same entity description pattern or the same link description pattern form a group.

[0008] Preferably, the subgraphs of the entity link graph are generated to connect at least one node from each group.

[0009] In a preferred embodiment, when generating a representative subset of RDF from an extended subgraph, the method includes, for each node in the extended subgraph representing an entity from an entity description pattern, adding all triples describing the class of the entity and a single triple describing the property of the entity for each property of the entity description pattern.

[0010] An RDF dataset contains a set of triples, i.e., <subject, predicate, object> triples. The subject or object of such a triple is referred to as an entity. The predicate is referred to as a relation. The set of triples can be naturally represented as a directed graph with nodes and edges labeled. The elements of an RDF dataset and a knowledge graph can be distinguished between instance-level elements and schema-level elements, where the schema is a formal description of the elements and the instance is the specific information actually stored in the dataset.

[0011] The schema-level elements of these triples form an entity description pattern (EDP) of e, which is a set of classes (C), forward properties (FP), and backward properties (BP), i.e., [Number] composed of.

[0012] Triples where the object is an entity are particularly important for representing a link between two entities. The predicate within such a triple <e i , p, e j > and the EDPs of the two entities form the link pattern (LP) of this triple, i.e., lp(<e i , p, e j , D) = <edp(e i , D), p, edp(e j , D)> to form.

[0013] According to a preferred embodiment, the method further includes the step of labeling each instance-level entity with its entity description pattern, and / or the step of labeling each edge representing an entity link between two instance-level entities with its link description pattern.

[0014] According to a preferred embodiment, the step of converting an edge of an undirected graph into a node includes the step of subdividing each edge.

[0015] According to a preferred embodiment, generating a subgraph of an entity link graph is based on solving a group Steiner tree problem.

[0016] To process the non-connectivity within a dataset, preferably, the union of all entity description patterns and all link description patterns is regarded as the universal set, and for each component D of the RDF dataset j EDP(D j ) ∪ LP(D j ) ⊆ EDP(D) ∪ LP(D) is a set, and the method includes the step of finding the minimum number of sets whose union is equal to the universal set.

[0017] Preferably, the method includes the step of determining the frequency of entity description patterns and / or the frequency of link description patterns. The frequency is defined as the occurrence of the pattern within the RDF dataset.

[0018] Preferably, the method includes the step of ranking entity description patterns and / or link description patterns according to their frequencies, for example, based on descending order.

[0019] According to a preferred embodiment, the method includes the step of restricting a subgraph of an entity link graph based on the frequency of entity description patterns and / or the frequency of link description patterns such that the subgraph connects at least one node from each group, and the group refers to entity description patterns and / or link description patterns having a frequency exceeding a defined threshold.

[0020] Further advantageous embodiments can be derived from the following description and drawings.

Brief Description of the Drawings

[0021] [Figure 1]This is a diagram illustrating an exemplary knowledge graph. [Figure 2] This figure shows an embodiment of a method for automated analysis of a Resource Description Framework (RDF) dataset D. [Figure 3] This figure shows a further aspect of the method for automated analysis of the Resource Description Framework (RDF) dataset D. [Figure 4] This figure shows a further aspect of the method for automated analysis of the Resource Description Framework (RDF) dataset D. [Figure 5] This figure shows the method for evaluating space-saving performance. [Figure 6] This figure shows a method for evaluating the cumulative distribution of space savings. [Figure 7] This figure shows the method for evaluating the cumulative distribution of snippet size. [Figure 8] This figure shows the methods for evaluating execution time. [Figure 9] This figure shows a method for evaluating the cumulative distribution of execution time. [Modes for carrying out the invention]

[0022] Figure 1 shows an exemplary Knowledge Graph (KG) 100. The Knowledge Graph represents the data from the RDF dataset D, which is a collection of triples, i.e.,<subject,predicate,object> It is composed of triples. The subject or object of such a triple is called an entity. The predicate is called a relation. A set of triples in KG can naturally be represented as a directed graph with labeled nodes and edges. Elements of RDF datasets and knowledge graphs can be distinguished between instance-level elements and schema-level elements, where the schema is a formal description of the element and the instance is the specific information actually stored in the dataset.

[0023] Knowledge graph (KG) nodes N1.1, N1.2, N2, N3.1, N3.2, N4.1, N4.2, N4.3, N5.1, N5.2, and N5.3 are, for example, N1.1: Germany, abbreviated as DE N1.2: The United Kingdom, abbreviated as UK N2: Europe N3.1: Berlin N3.2: London N4.1: Munich N4.2: Augsburg N4.3: Oxford N5.1: Specific number 1 N5.2: Specific number 2 N5.3: Specific number 3 This refers to instance-level entities such as those mentioned above.

[0024] Nodes N6 and N7 of the Knowledge Graph (KG) are, for example, N6: Capital N7: City This refers to schema-level entities such as those shown.

[0025] The Knowledge Graph (KG) edges E1, E2, E3.1, and E3.2 are, for example, E1: Part of ~ E2: Capital of ~ E3.1: Located at ~ E3.2: Located at ~ This refers to instance-level relationships like the one described.

[0026] Edges E4 and E5 of the Knowledge Graph (KG) are, for example, E4: Type E5: Area This refers to schema-level relationships like the one shown.

[0027] In the triples of the RDF dataset D, an instance-level entity e is described by a subset of triples, where e is the subject or object. The schema-level elements of these triples form an entity description pattern (EDP) for e, consisting of a set of classes (C), forward properties (FP), and backward properties (BP), i.e.,

number

[0028] Triples whose object is an entity are particularly important because they represent a link between two entities. <e i ,p,e j The predicates within > and the EDP of the two entities form this triple link pattern (LP), i.e., lp( <e i ,p,e j >,D)= <edp(e i ,D),p,edp(e j ,D)> It forms.

[0029] Referring to Figure 1, an example pattern is shown, namely,

number

[0030] The sets of all EDPs and all LPs in D, represented by EDP(D) and LP(D), respectively, are obtained by iterating through all entities and links in D.

[0031] Preferably, each node of KG in Figure 1 can be labeled with its EDP, and each edge can be labeled with its LP.

[0032] Figure 2 shows the entity-linked graph (ELG) representation of the RDF dataset D. The ELG is obtained by subdividing each edge, thereby converting the labeled edges of the KG into labeled nodes. The subdivision is referred to as the entity-linked graph representation of the RDF dataset D. The entity-linked graph ELG is reduced to include only instance-level elements.

[0033] Figure 3 shows the extended subgraph eSG of the entity link graph ELG. The extended subgraph eSG is generated as follows:

[0034] First, a subgraph SG of the entity link graph ELG is generated, connecting at least one node from each group. Subgraph SG is generated by solving the group Steiner tree problem. Essentially, this means generating a minimally connected subgraph of the entity link graph ELG where node labels cover entity description pattern EDP and link description pattern LP. According to one embodiment, the subgraph is generated as an unweighted version of the group Steiner tree problem, where all nodes with the same label form a group. Solving the group Steiner tree problem requires finding the smallest tree that connects at least one node from each group and covers all different labels. The group Steiner tree problem can be solved using state-of-the-art approximation algorithms such as KeyKG+, disclosed in Shi, Y., Cheng, G., Khallamov, E., “Keyword search over knowledge graphs via static and dynamic hub labellings”, WWW 2020, pp.235-245 (2020).

[0035] Next, referring to Figure 2, the subgraph includes the following nodes N1.1, N2, N3.1 and N4.1 representing entities, and the following E1, E2, E3.1 and E3.2 representing links between entities.

[0036] Secondly, the extended subgraph eSG is generated by adding missing links to each node in the subgraph representing the entity link, so that each node in the extended subgraph representing the entity link connects both entities that are linked. This is shown by the dotted line between node E3.1 and node N3.1 in Figure 3. According to one embodiment, for each leaf in the computed subgraph representing the entity link, the subgraph is extended to include both entities that the node links.

[0037] Finally, Figure 4 shows a representative subset S of the RDF dataset D generated from the extended subgraph eSG.

[0038] A typical subset S is the pattern coverage snippet S of the RDF dataset D.

[0039] As shown in the example in Figure 4, a typical subset S can be derived from the extended subgraph eSG as follows: For each node in the extended subgraph eSG representing an entity from the entity description pattern, all triples describing the class of that entity and a unique triple describing the property of that entity for each property of the entity description pattern are added. Furthermore, for each node in the extended subgraph SG representing an entity link, the corresponding triple from the link description pattern LP is added.

[0040] The embodiments described with respect to Figures 1 to 4 are based on the assumption of connectivity of the RDF dataset D. The non-connectivity of the RDF dataset D can be handled as follows.

[0041] Non-associativity can be viewed as an instance of the well-known set cover problem, where the union of all entity description patterns EDP(D) and all link description patterns LP(D) is considered the total set, and each component D of the RDF dataset D is considered a component of the RDF dataset D. j In contrast, EDP(D j )∪LP(Dj )⊆EDP(D)∪LP(D) is considered a set. The set cover problem is solved by finding the smallest number of sets whose union is equal to the entire set.

[0042] This solution involves iteratively generating representative subsets S until the entire set is completely covered. Each component D contains the most uncovered pattern. i For this, a representative subset S is generated as described above.

[0043] According to a favorable embodiment, the generation of a representative subset S is performed by component D i It can be modified to generate a potentially smaller subsnippet. The subsnippet is an entity description pattern EDP(D i ) and description pattern LP(D i It is only necessary to cover a part of the union with ), and the complete union EDP(D i )∪LP(D i ) intersects with the entire set, rather than the entire set. This means that when generating the subgraph SG of ELG, the union EDP(D) does not have a universal set. i )∪LP(D i The groups corresponding to the pattern can be ignored.

[0044] This approach aims to find the smallest subset of components that cover all patterns in the RDF dataset D, thereby improving the compactness and generation efficiency of a representative subset S.

[0045] If the RDF dataset D is highly heterogeneous and contains many different patterns, the representative subset S will inevitably be very large. In the next step, we handle the high degree of heterogeneity by making a trade-off between pattern coverage and snippet size. Patterns within the RDF dataset D may not be that important. The relative frequency of an entity description pattern EDP is defined as the proportion of entities in the RDF dataset D that have this entity description pattern EDP. The relative frequency of a link description pattern LP is defined similarly. Patterns with higher frequencies are considered more important. The generated patterns EDP and LP can be ranked in descending order of their relative frequency.

[0046] When generating a subgraph SG of an entity link graph ELG, the subgraph SG can be restricted based on the frequency of entity description patterns and / or link description patterns such that the subgraph joins to at least one node from each group, where a group refers to entity description patterns and / or link description patterns that have a frequency exceeding a defined threshold. The threshold is, for example, a percentage.

[0047] Based on this, the generated representative subset S will likely be a smaller snippet, covering only the most important patterns in the RDF dataset D.

[0048] The following embodiments refer to generating a query-biased representative subset S.

[0049] A keyword query Q contains at least one or more keywords. All keywords are considered keyword patterns. Each entity or entity link in the RDF dataset D is extended to have a set of patterns consisting of its entity description pattern EDP or link description pattern LP, and all keyword patterns it matches. This calculation can be performed by an off-the-shelf matcher. Entity e matches keyword q∈Q if q appears in any triple describing entity e in the RDF dataset D. Entity Link <e i ,p,e j > matches keyword q when it appears in the text form of p. Therefore, for each keyword pattern of keyword q ∈ Q, a group consisting of all entities and entity links that match keyword q is added. Thus, when generating the subgraph SG of the entity link graph ELG that connects at least one node from each group, keywords are considered by adding a group consisting of all entities and entity links that match each keyword q for each keyword q.

[0050] Based on this, the generated representative subset S matches all the keywords in Q.

[0051] Prior art, "Generating humorous snippets for open data on the web" by Cheng, G., Jin, C., Ding, W., Xu, D., and Qu, Y., WSDM2017, pp.151-159 (2017), also discloses a method for generating snippets called IlluSnip. To compactly illustrate the contents of large RDF datasets, IlluSnip generates snippets by formulating a combined graph problem of maximum weight and coverage. This aims to extract the optimal subset of k triples, represented as a combined RDF graph that covers the most frequent classes, properties, and most central entities in the RDF dataset. Unlike IlluSnip, KSD, disclosed in "Towards multi-facet snippets for dataset search" by Wang, X., Cheng, G., and Khallamov, E., PROFLILES&SemEx2019, pp.1-6 (2019), formulates a weighted maximum coverage problem that removes connectivity constraints. The goal of optimization is to cover the most keywords for keyword queries that are more suitable for an RDF dataset search engine. Compared to IlluSnip and KSD, this approach also aims to cover schema-level elements, focusing on patterns of entity descriptions and links, which are combinations of classes and properties. Patterns can provide a "higher-order" data preview than individual classes and properties.

[0052] Below, we compare the space savings and execution time of this approach with those of IlluSnip and KSD. This approach is referred to as PSCG.

[0053] The space savings achieved by the RDF dataset approach are as follows: Space saving = 1 - (Number of triples in the generated snippets = Number of triples in the RDF dataset) / (Number of triples in the RDF dataset) It is defined as follows.

[0054] The snippet size is reported in triples. Additionally, the execution time for each method on the RDF dataset is reported.

[0055] For evaluation purposes, we obtained datasets using RDF dumps from two data portals, DataHub.io and Data.gov, and parsed 9544 RDF datasets using Apache Jena 3.9.0. It should be noted that many entities within the Data.gov dataset are not typed and are likely described by uniform patterns converted from tabular data.

[0056] The space savings of the PSCG approach were calculated for each of the 9544 RDF datasets. The results are summarized in the table in Figure 5. The PSCG approach reduced the size of the RDF datasets by an average of approximately 90%. The space savings for PCSG, PCSG-90%, and PCSG-80% were above 95% for 57%, 69%, and 72% of all RDF datasets, respectively, as shown in the cumulative distribution in Figure 6. As shown in the cumulative distribution in Figure 7, the median number of triples in the generated snippets was 41, 20, and 17, respectively. The results support the compactness of the snippets generated by the PSCG approach.

[0057] For each approach, the execution time for each of the 9544 RDF datasets was recorded. The results are summarized in the table shown in Figure 8. PCSG(-τ) was more than two orders of magnitude faster than IlluSnip. The execution times for PCSG, PCSG-90%, and PCSG-80% were less than 1 second for 98%, 98%, and 99% of all RDF datasets, respectively, as shown by the cumulative distribution in Figure 9. This result supports the computational efficiency of the PSCG approach. However, for some highly heterogeneous datasets, including thousands of EDPs and LPs, PCSG(-τ) took more than 1 hour. The PSCG approach is still faster than IlluSnip and acceptable for offline computation, but this suggests there is room for further performance improvement.

[0058] This method is preferably applied when processing RDF datasets using neural networks. The datasets include, for example, data from the field of manufacturing or production, such as information on materials and substances and their properties. The RDF datasets are processed, for example, in relation to automation, diagnostics, and optimization. A representative subset S of the RDF dataset can be processed first for rapid estimation or evaluation of the RDF dataset.

[0059] Further embodiments refer to the use of computer-implemented methods for processing RDF datasets, particularly using neural networks. Using this method means automatically generating a representative subset S of the RDF dataset, thereby extracting information from the text data, for example, information about entities, particularly people, places, organizations, etc., and / or concepts, particularly information about proteins, chemicals, materials, automated processes, diagnostics, and optimizations.

[0060] Further embodiments refer to the use of multiple RDF datasets, at least two RDF datasets, and in particular a computer-implemented method for processing neural networks, the method comprising receiving a keyword query Q, generating query-biased representative subsets S of at least two RDF datasets, and determining the degree of similarity and / or similarity between the keyword query Q and each query-biased representative subset S. Based on the degree of similarity and / or similarity, a suitable representative subset, and therefore a suitable RDF dataset with respect to the keyword query Q, can be determined. The determination of the degree of similarity and / or similarity can be based, for example, on vector embeddings of words in a high-dimensional vector space, i.e., the keyword and the elements of the representative subsets, where words are similar if they are close to each other in the vector space.

[0061] Further embodiments relate to the use of computer-implemented methods for creating databases, in particular structured knowledge databases, and especially knowledge graphs, which, according to embodiments, are applied to extract information, for example, to generate a representative subset S, and the information, for example, the representative subset S, is used to create a database, in particular structured knowledge database, and especially knowledge graph.

[0062] The method according to this embodiment can be applied to RDF datasets from different domains.

Claims

1. A computer-implemented method for automated analysis of a Resource Description Framework (RDF) dataset (D), wherein the RDF dataset (D) comprises a set of triples, the RDF dataset (D) is provided as an undirected graph (KG) comprising nodes (N) and edges (E), where the nodes (N) represent entities (e), and the edges (E) represent links between entities (e), and the method is: - For each instance-level entity, at least one entity description pattern (EDP) comprising at least one triple describing at least one class and / or at least one property of the entity, - For each edge representing an entity link between two instance-level entities, at least one link description pattern (LP) comprising at least one triple describing the link between the two instance-level entities. The steps to generate, A step of generating an entity link graph (ELG) by converting the edges of the undirected graph (KG) into nodes, wherein all nodes containing the same entity description pattern (EDP) or the same link description pattern (LP) form a group. A step of generating a subgraph (SG) of the entity link graph (ELG), wherein the subgraph (SG) connects at least one node from each group, The steps include generating an extended subgraph (eSG) by adding missing links to each node of the subgraph (SG) representing entity links so that each node of the extended subgraph (eSG) representing entity links is connected to both entities it links to, and - For each node in the extended subgraph (eSG) representing an entity (e) from the entity description pattern (EDP), there is at least one triple describing the class of the entity, and at least one triple describing the properties of the entity for each property of the entity description pattern, and - At least one triple from the link description pattern (LP) for each node of the extended subgraph (eSG) representing an entity link The steps include: adding a representative subset (S) of the RDF dataset from the extended subgraph (eSG), A method that includes this.

2. The method according to claim 1, further comprising the steps of labeling each instance-level entity with its entity description pattern (EDP), and / or labeling each edge representing an entity link between two instance-level entities with its link description pattern (LP).

3. The method according to claim 1, wherein the step of converting the edges of the undirected graph (KG) into nodes includes the step of subdividing each edge.

4. The method according to claim 1, wherein generating subgraphs of the entity link graph (ELG) is based on solving a group Steiner tree problem.

5. The union of all entity description patterns (EDPs) and all link description patterns (LPs) is considered the total set, and each component D of the RDF dataset (D) ｊ In contrast, EDP (D ｊ )∪LP(D ｊ The method according to claim 1, wherein ) ⊆ EDP(D) ∪ LP(D) is a set, and the method comprises the step of finding the smallest number of sets whose union is equal to the entire set.

6. The method according to claim 1, comprising the step of restricting the subgraph (SG) of the entity link graph (ELG) based on the frequency of entity description patterns (EDPs) and / or link description patterns (LPs) such that the subgraph (SG) connects at least one node from each group, wherein the group refers to entity description patterns (EDPs) and / or link description patterns (LPs) having a frequency exceeding a defined threshold.

7. The method according to claim 1, wherein when generating the subgraph (SG) of the entity link graph (ELG), the method includes receiving a keyword query (Q), generating a keyword pattern for each keyword in the keyword query, and adding to each keyword pattern a group consisting of all entities and entity links that match the keyword.

8. An apparatus for automated analysis of a Resource Description Framework (RDF) dataset (D), wherein the RDF dataset (D) comprises a set of triples, the RDF dataset (D) is provided as an undirected graph (KG) comprising nodes (N) and edges (E), where the nodes (N) represent entities (e), and the edges (E) represent links between entities (e), and the apparatus comprises at least one processor, the processor being, - For each instance-level entity, at least one entity description pattern (EDP) comprising at least one triple describing at least one class and / or at least one property of the entity, - For each edge representing an entity link between two instance-level entities, at least one link description pattern (LP) comprising at least one triple describing the link between the two instance-level entities. To generate, The process involves generating an entity link graph (ELG) by converting the edges of the aforementioned undirected graph (KG) into nodes, wherein all nodes containing the same entity description pattern (EDP) or the same link description pattern (LP) form a group. The process involves generating a subgraph (SG) of the entity link graph (ELG), wherein the subgraph (SG) connects at least one node from each group, The process involves generating an extended subgraph (eSG) by adding missing links to each node of the subgraph (SG) representing entity links, so that each node of the extended subgraph (eSG) representing entity links is connected to both entities it links to, and - For each node in the extended subgraph (eSG) representing an entity (e) from the entity description pattern (EDP), there is at least one triple describing the class of the entity, and at least one triple describing the properties of the entity for each property of the entity description pattern, and - For each node of the extended subgraph (eSG) representing an entity link, at least one triple from the link description pattern (LP) By adding this, a representative subset (S) of the RDF dataset is generated from the extended subgraph (eSG), A device configured to perform the following actions.

9. The apparatus according to claim 8, wherein the apparatus is configured to label each instance-level entity with its entity description pattern (EDP) and / or label each edge representing an entity link between two instance-level entities with its link description pattern (LP).

10. The apparatus according to claim 8, wherein the apparatus is configured to perform a transformation of the edges of the undirected graph (KG) into nodes, which includes subdividing each edge.

11. The apparatus according to claim 8, wherein the apparatus is configured to generate subgraphs of the entity link graph (ELG) based on solving a group Steiner tree problem.

12. The union of all entity description patterns (EDPs) and all link description patterns (LPs) is considered the total set, and each component D of the RDF dataset (D) ｊ In contrast, EDP (D ｊ )∪LP(D ｊ The apparatus according to claim 8, wherein ) ⊆ EDP(D) ∪ LP(D) is a set, and the apparatus is configured to find the smallest number of sets whose union is equal to the entire set.

13. The apparatus according to claim 8, wherein the apparatus is configured to restrict the subgraphs (SG) of the entity link graph (ELG) based on the frequency of entity description patterns (EDPs) and / or link description patterns (LPs) such that the subgraphs (SG) connect at least one node from each group, and the group refers to entity description patterns (EDPs) and / or link description patterns (LPs) having a frequency exceeding a defined threshold.

14. The apparatus according to claim 8, wherein, when generating the subgraph (SG) of the entity link graph (ELG), the apparatus receives a keyword query (Q), generates a keyword pattern for each keyword in the keyword query, and adds to each keyword pattern a group consisting of all entities and entity links that match the keyword.

15. A computer program for automatically analyzing keywords, comprising a computer-readable instruction for causing the computer to perform the method according to any one of claims 1 to 7 when executed by the computer.

Images