Computer-implemented method for keyword search in a knowledge graph

Abstract

Computer-implemented method for keyword search in a knowledge graph. Computer-implemented method (100) for keyword search in a dataset (Q), wherein data (q) of the dataset (Q) is represented by a knowledge graph (KG), wherein the knowledge graph (KG) comprises vertices (v) representing entities (v) of the dataset (Q) and edges (e) representing relations between said entities (v), the method (100) comprising the steps of: providing (110) a set of subgraphs (G) of the knowledge graph (KG) connecting at least two entities (v) of a keyword query; computing (120) a relevance (R) of each subgraph (G) based on both a relevance (R) of each subgraph (G) to the dataset (Q), referred to as data relevance, and a relevance (R) of the entities (v) of each subgraph (G) to each other, referred to as internal relevance, and ranking (130) the subgraphs (G) according to their relevance (R) among the set of subgraphs (G).

Description

Background Technology

[0001] This disclosure relates to a computer-implemented method for performing keyword searches in a dataset, wherein the data of the dataset is represented by a knowledge graph.

[0002] Furthermore, this disclosure relates to a system for keyword searching in a dataset and a computer program for keyword searching in a dataset.

[0003] A knowledge graph (KG) is a graph where vertices are entities interconnected by binary relations and annotated with types and data values. Knowledge graphs are increasingly being developed for various domains, supporting a wide range of tasks from enterprise data integration to intelligent search.

[0004] A key task in knowledge graph-based applications is to extract subgraphs, also known as relational subgraphs (RSGs), that relate to a given set of vertices. For example, during a keyword search, a set of vertices in a knowledge graph can be selected, where a user queries the knowledge graph by proposing a set of keywords, and the system matches these keywords with entities in the knowledge graph, finds relational subgraphs that associate these entities, and returns them as the answer.

[0005] To be effectively enriched, subgraphs should be relevant to their context, i.e., the content of the dataset. Therefore, the task of this disclosure is to provide a measure of subgraph relevance. Summary of the Invention

[0006] This is achieved through the systems and methods according to the independent claims.

[0007] The embodiment relates to a computer-implemented method for performing keyword search in a dataset, wherein the data of the dataset is represented by a knowledge graph, wherein the knowledge graph includes vertices representing entities in the dataset and edges representing relationships between the entities, the method comprising the following steps: Provides a set of subgraphs of a knowledge graph that connects at least two entities for a keyword query; The relevance of each subgraph is calculated based on both the relevance of each subgraph to the dataset (called data relevance) and the relevance of entities within each subgraph to each other (called internal relevance). The subgraphs are sorted according to their relevance within the subgraph set.

[0008] Advantageously, the relevance of a subgraph can be considered from two perspectives: the relevance of the subgraph to the dataset and the relevance of the elements (i.e. entities) of the subgraph to each other.

[0009] According to a preferred embodiment of the method, the method further includes the steps of calculating data correlation and / or calculating internal correlation.

[0010] According to a preferred embodiment of the method, the data relevance between the subgraph and the dataset is calculated based on the relevance between entities in the dataset and entity pairs in the subgraph.

[0011] According to a preferred embodiment of the method, the correlation between entities in the dataset and entities in the subgraph is calculated based on the similarity of the entities' categories, particularly the category type.

[0012] Each entity in the dataset and each entity in the subgraph is annotated with its category type, which is determined by... The markings, among which It is, for example, a collection of all categories given in a certain directory.

[0013] To calculate the two entities v i v j The correlation between them r(v) i , v j ), entity v i v j Represented as by A dimensional vector space gives two vectors in the vector space. , , where the l-th dimension corresponds to the category And calculate two vectors in this space. , Cosine similarity: .

[0014] For vectors , Consider the category type as well as the category types of their neighbors in the subgraph. By considering the category types of the neighbors, implicit relevance can be taken into account.

[0015] Specifically, for ,Right now For the l-th dimension, define the following formula: in Indicator type frequency - t l Relative to v i The reverse entity frequency. It adapts the well-known term frequency—the reverse document frequency in information retrieval—to the knowledge graph: , in , and in Calculate t lInstances in v i The proportion among the neighbors, and Calculate t l The normalized information content. The standard definition of information content is given by the following formula. .

[0016] Information content and t l The number of instances is inversely proportional to the logarithm of the number of instances. In this case, TYPS(v) is a label derived from type(v) and its... The category hierarchy consists of a set of categories composed of ancestors. In other words, subsumption-based reasoning is performed to obtain implicit entity types from the hierarchy.

[0017] According to a preferred embodiment of the method, data relevance is calculated by computing the average relevance between entities in the dataset and entities in the subgraph, wherein the average relevance is given by the following formula. .

[0018] Formally, Q denotes the set of entities mentioned in the dataset in the knowledge graph, and Identify the entities in the subgraph. Calculate data relevance by evaluating the average relevance between pairs of entities in the dataset and the subgraph.

[0019] According to a preferred embodiment of the method, the internal correlation between entities in a subgraph is calculated based on the similarity between entity pairs in the subgraph.

[0020] According to a preferred embodiment of the method, wherein sim(v) i v j The similarity between entity pairs given is calculated based on the similarity of the entities' categories, specifically category types. One method for calculating the similarity sim(v) is... i v j Possible ways of doing this are disclosed, for example, by Ganggao Zhu and Carlos Angel Iglesias in Computing semantic similarity of concepts in knowledge graphs (IEEE Trans. Knowl. Data Eng., 29(1):72–85,2017). Specifically, it is the similarity between the category type of one entity in a pair of entities and the category type of the other entity in the pair of entities, based on the relative position of the categories in the category hierarchy. Two categories are similar if they are close together and generally fall under a specific category in the hierarchy. The similarity sim(v) is calculated as follows: i vj It can be calculated using the following formula: Where len(type(v) i ), type(v j )) is entity v in the category hierarchy i Category type and entity v j The length of the shortest path between category types, These are the parameters to be tuned, and ic(lcs(type(v) i ), type(v j ))) is the category type (v) in the category hierarchy. i ) and category type (v j The information content of the least common subsumer (lcs) of two categories. The least common subsumer of two categories is their shared ancestor at the lowest level of the category hierarchy.

[0021] According to a preferred embodiment of the method, internal correlation is calculated by calculating the average similarity between entity pairs in the subgraph, wherein the average similarity is given by the following formula. .

[0022] According to a preferred embodiment of the method, the correlation of each subgraph is calculated by the following formula. , in Indicate the relevance of subplots. Indicate the data correlation of subplots, Indicate the internal correlation of the subplot, and It is the correlation of tuning data. The parameter relating to the relative importance of internal correlations.

[0023] Another embodiment relates to a system for keyword searching in a dataset. The system is configured to implement a computer-based method according to any embodiment.

[0024] Another embodiment relates to a computer program that includes computer-readable instructions that, when executed by a computer, particularly a computer of a system according to an embodiment, cause the computer to perform a method according to any embodiment.

[0025] Other embodiments involve using the methods and / or systems according to embodiments in knowledge graph-based text enrichment, wherein words or phrases in documents such as news articles or reports are annotated with types from a given background knowledge graph and linked to vertices of the knowledge graph. For effective document enrichment, subgraphs should be relevant to the document content.

[0026] Previous methods used simple word-based metrics to compute the relevance of subgraphs to documents. This is disclosed, for example, in Viswanathan and Krish namurthi Ilango’s Ranking semantic relationships between two entities using personalization in context specification (Inf. Sci., 207:35–49, 2012).

[0027] According to the disclosed method, document enrichment is based on entity-based relevance and can therefore leverage ontology knowledge.

[0028] Further advantageous embodiments are derived from the following description and accompanying drawings. In the drawings... Figure 1 A schematic view depicts a computer-implemented method for keyword search according to an embodiment; Figure 2 A schematic view of a knowledge graph according to an embodiment is depicted, and Figure 3 A schematic view of a system for keyword search according to an embodiment is depicted.

[0029] Figure 1 A computer-implemented method 100 is provided for keyword search in a dataset Q. The dataset Q is represented by a knowledge graph KG, wherein the knowledge graph KG includes vertices v representing entities in the dataset Q and edges e representing relationships between said entities.

[0030] Figure 2 An exemplary representation of a knowledge graph (KG) is given. A knowledge graph KG can be used for graph-based knowledge representation by describing entities (e.g., real-world entities) and their relationships. Data q of dataset Q is represented in the knowledge graph KG. The knowledge graph can include additional information besides the data q of dataset Q. Dataset Q itself is, for example, a text document containing words or phrases. For example, a document is a news article or report, annotated with types from a given background knowledge graph, linked to vertices of the knowledge graph. Each entity v is annotated with its type, which is a category defined in the ontology and assigned by... Marking. The categories in the text constitute a classification hierarchy.

[0031] According to the disclosed embodiments, method 100 includes the following steps: A subgraph of a knowledge graph KG that provides at least two entities v for 110-connection keyword queries. A set; Based on each subgraph Correlation with dataset Q (referred to as data relevance) The correlation between entities v in each subgraph and each other. (referred to as internal correlation) Both are used to calculate 120 for each subgraph. correlation ,as well as According to the subgraph Subgraphs in a set correlation Pair diagram Sort 130.

[0032] According to a preferred embodiment of method 100, method 100 further includes computing a subgraph. Data correlation with dataset Q Step 140.

[0033] According to a preferred embodiment of method 100, based on the entity v and subgraph of dataset Q The correlation between entity v pairs is calculated using 140 subgraphs. Data correlation with dataset Q .

[0034] According to a preferred embodiment of the method, based on the similarity of the categories of the entity v, particularly the category type, the similarity between entity v and subgraph in dataset Q is calculated. The degree of correlation between entities v.

[0035] Each entity v and subgraph in dataset Q Each entity v is annotated with its category type, which is determined by... The markings, among which It is, for example, a collection of all categories given in a certain directory.

[0036] To calculate the two entities v i v j The correlation between them r(v) i , v j ), entity v i v j Represented as by A dimensional vector space gives two vectors in the vector space. , , where the l-th dimension corresponds to the category And calculate two vectors in this space. , Cosine similarity: .

[0037] For vectors , Considering category type and also in subgraphs Consider the class types of their neighbors. By taking into account the class types of neighbors, implicit relevance can be taken into account. A vertex v's neighbor is adjacent to v by an edge. Let dist(u, v) represent all neighbors of entity v in the knowledge graph KG. The distance between two vertices u and v, labeled by dist(u, v), is the length of the shortest path connecting u and v in G, or, if no such path exists, the distance is 0. .

[0038] Specifically, for ,Right now For the l-th dimension, define the following formula: in Indicator type frequency - t l Relative to v i The reverse entity frequency. It adapts the well-known term frequency—the reverse document frequency in information retrieval—to the knowledge graph: , in , and in Calculate t l Instances in v i The proportion among the neighbors, and Calculate t l The normalized information content. The standard definition of information content is given by the following formula. .

[0039] Information content and t l The number of instances is inversely proportional to the logarithm of the number of instances. In this case, TYPS(v) is a label derived from type(v) and its... The category hierarchy consists of a set of categories composed of ancestors. In other words, classification-based reasoning is performed to obtain implicit entity types from the hierarchy.

[0040] According to a preferred embodiment of the method, the entity v and subgraph of dataset Q are calculated. Data relevance is calculated by the average relevance between pairs of entities v. The average correlation is given by the following formula. .

[0041] Formally, Q denotes the set of entities v mentioned in dataset Q in the knowledge graph, and Subgraph The entities. By calculating the dataset Q and the subgraph Data relevance is calculated by the average relevance between pairs of entities v. .

[0042] According to a preferred embodiment of method 100, method 100 further includes calculating each subgraph. The internal correlation between entities v Step 150.

[0043] According to a preferred embodiment of method 100, based on subgraph The similarity between pairs of entities v is used to calculate 150 subgraphs. The internal correlation between entities v .

[0044] According to a preferred embodiment of the method, wherein sim(v) i v j The similarity between pairs of entities v is based on the entities v. i v j The similarity is calculated based on the category similarity, particularly the category type. One method for calculating similarity is sim(v). i v j Possible methods for this are disclosed, for example, in Ganggao Zhu and Carlos Angel Iglesias's *Computing semantic similarity of concepts in knowledge graphs* (IEEE Trans. Knowl. Data Eng., 29(1):72–85, 2017). Specifically, based on the relative position of the category in the category hierarchy, an entity v in the pair of entities... i The category type and another entity v in the pair of entities j The similarity between category types. Two categories are considered similar if they are close together and typically contained under a specific category in a hierarchy. The similarity is expressed as sim(v). i v j It can be calculated using the following formula: Where len(type(v) i ), type(v j )) is entity v in the category hierarchy i Category type and entity v j The length of the shortest path between category types, These are the parameters to be tuned, and ic(lcs(type(v) i ), type(v j ))) is the category type (v) in the category hierarchy. i ) and category type (v j The information content of the least common classifier (lcs) of two categories. The least common classifier of two categories is their shared ancestor at the lowest level of the category hierarchy.

[0045] According to a preferred embodiment of method 100, by calculating the subgraph Internal correlation is calculated by averaging the similarity between pairs of entities v. The average similarity is given by the following formula. .

[0046] According to a preferred embodiment of method 100, each sub-graph The correlation is calculated using the following formula. , in Subgraph The correlation, Subgraph Data relevance, Subgraph The internal correlation, and It is the correlation of tuning data. and internal correlation The parameter indicating the relative importance of the number of elements.

[0047] Figure 3 A schematic view of a system 200 for keyword searching in dataset Q, according to an embodiment, is depicted.

[0048] System 200 is configured to perform at least one of steps 110, 120, 130, 140, and 150 of method 100.

[0049] System 200 includes computing unit 210, such as microprocessor and / or microcontroller and / or programmable logic device, particularly FPGA and / or application-specific integrated circuit ASIC and / or digital signal processor DSP and / or combinations thereof.

[0050] System 200 includes at least one storage unit 220. Storage unit 220 may further include volatile memory, particularly random access memory (RAM), and / or non-volatile memory, such as flash EEPROM. Storage unit 220 contains at least one computer program PRG1 for computing unit 210, which controls the execution of method 100 for keyword search in dataset Q.

[0051] System 200 may further include interface unit 230 for receiving data constituting the knowledge graph KG and / or dataset Q and / or search query from at least one external data source. The data constituting the knowledge graph and / or dataset Q and / or search query may be stored in the system's storage unit 220, or in another (e.g., external) storage unit.

[0052] Another embodiment relates to a computer program PRG 1. The computer program PRG 1 includes computer-readable instructions that, when executed by a computer, particularly computer 210 of system 200 according to the embodiment, cause the computer to perform the steps of method 100, particularly method 100 according to the embodiment.

Claims

1. A computer-implemented method (100) for keyword search in a dataset (Q), wherein data (q) of the dataset (Q) is represented by a knowledge graph (KG), wherein the knowledge graph (KG) includes vertices representing entities (v) of the dataset (Q) and edges (e) representing relationships between said entities (v), the method (100) comprising the following steps: Provides a subgraph of the knowledge graph (KG) of at least two entities (v) that are connected to a keyword query (110). A set of ) Based on each subgraph ( The data relevance between the dataset (Q) and the dataset (Q) ) and each subgraph ( The internal correlations between entities (v) of each other () Both are used to calculate (120) each subgraph ( The correlation of ) ),and According to the subgraph ( Subgraphs in the set of ) The correlation of ) Pair graphs Sort (130), where the subgraph ( The data relevance between the dataset (Q) and the dataset (Q) ) is based on the dataset (Q) and the entities (v) and subgraphs ( The correlation (r) between pairs of entities (v) is calculated. Among them, based on the type of the category of the entity (v), the entities (v) and subgraphs (v) of the dataset (Q) are calculated. The relevance (r) between entities (v) of ) This involves calculating the entities (v) and subgraphs (v) of the dataset (Q). The data relevance is calculated by the average relevance (r) between pairs of entities (v) of the data. The average relevance (r) is given by the following formula: ， Where Q represents the set of entities (v) mentioned in the dataset (Q) in the knowledge graph, and The subgraph represents the subgraph ( ) entities, and in which Indicates the correlation of the data, r(v) i , v j ) represents two entities v i v j The correlation between them Among them, based on subgraph ( The similarity between pairs of entities (v) in a subgraph is used to calculate the similarity between them. The internal correlations between entities (v) of each other () ), Among them, by calculating the subgraph ( Internal correlation is calculated by the average similarity between pairs of entities (v) in ) The average similarity is given by the following formula: ， in This indicates the internal correlation. The subgraph represents the subgraph ( The entity is , and sim(v) i , v j ) represents entity v i v j The similarity between pairs.

2. The computer-implemented method (100) according to claim 1, wherein, Based on the similarity of the categories of the entities (v), the similarity between entities (v) and subgraphs (v) in the dataset (Q) is calculated. The relevance (r) between entities (v) of ).

3. The computer-implemented method (100) according to claim 1, wherein sim(v i v j The similarity between pairs of entities (v) is calculated based on the similarity of the categories of the entities (v).

4. The computer-implemented method (100) according to claim 1, wherein sim(v i v j The similarity between pairs of entities (v) is calculated based on the type of the category of the entity (v).

5. The computer-implemented method (100) according to any one of claims 1-4, wherein each sub-graph ( The correlation of ) ) Calculated using the following formula ， in It is about tuning data correlation and internal correlation. The parameter indicating the relative importance of [the parameter].

6. A system (200) for keyword search in a dataset (Q), wherein the data (q) of the dataset (Q) is represented by a knowledge graph (KG), wherein the knowledge graph (KG) includes vertices representing entities (v) of the dataset (Q) and edges (e) representing relationships between said entities (v), wherein the system (200) is configured to perform a computer-implemented method (100) according to any one of claims 1 to 5.

7. A computer program product comprising a computer program (PRG 1), wherein the computer program (PRG 1) includes computer-readable instructions that, when executed by a computer, cause the computer to perform the method (100) according to any one of claims 1 to 5.

Images