JP7147380B2 - Type estimation method, information processing device and type estimation program - Google Patents

Description

The present invention relates to a type estimation method and the like.

A knowledge utilization system is known that generates a table by estimating the intention of the creator of the table from the schema information of the table and automatically generating a query for the knowledge graph. Such systems infer the type (class) of a table from within the knowledge graph in order to automatically generate queries that can fill the table as much as possible.

For example, as a table type estimation method, there is a method using a tf-idf (term frequency-inverse document frequency) method (see, for example, Non-Patent Document 1). In such a method, for example, the appearance frequency is calculated for each type instance obtained from the schema information of the table, the appearance frequency (tf) is weighted by the tf-idf method, and the score of each type is calculated. Then, the type with the highest score is estimated as the table type. Specifically, the type is regarded as a document, and the predicate of an instance to which this type is assigned is regarded as a word, the word score (tf-idf) is calculated, and the score of each type is calculated. It should be noted that in Non-Patent Document 1, DBpedia and Freebase (registered trademark) are used as information included in the knowledge graph.

Japanese Patent Publication No. 2017-531240 JP 2013-254421 A

Duan S., Fokoue A., Hassanzadeh O., Kementsietsidis A., Srinivas K., Ward M.J. (2012) Instance-Based Matching of Large Ontologies Using Locality-Sensitive Hashing. In: Cudre-Mauroux P. et al. (eds ) The Semantic Web - ISWC 2012. ISWC 2012. Lecture Notes in Computer Science, vol 7649. Springer, Berlin, Heidelberg

However, the conventional technique has the problem that it is difficult to properly estimate the table type. That is, in this technique, the type is regarded as a document, and the predicate possessed by an instance given this type is regarded as a word, the score of the word (tf-idf) is calculated, and the score of each type is calculated. Type scores tend to be calculated higher. In other words, the general type is inferred as the table type. Therefore, it is difficult to properly estimate the table type with the conventional technique.

In one aspect, the present invention aims at properly estimating the type of target table.

In one aspect, when estimating the type of a table in which a column name and a column corresponding to the column name are associated, the type estimation method uses types included in graph data and the frequency of occurrence of predicates of instances of each type. and based on the information on the ontology that defines the conceptual relationship between the words included in the graph data, from among the identified candidates of the table type A computer performs the process of estimating the type.

According to one embodiment, the type of table of interest can be reasonably inferred.

FIG. 1 is a block diagram showing the functional configuration of the type estimation device according to the embodiment. FIG. 2 is a diagram showing an image of the type estimation method according to the embodiment. FIG. 3 is a diagram showing an example of a query used for statistical score calculation. FIG. 4 is a diagram illustrating an example of the data structure of a statistical information DB; FIG. 5 is a diagram illustrating an example of a deduction rule DB schema according to the embodiment; FIG. 6 is a diagram illustrating an example of type estimation according to the embodiment. FIG. 7 is a diagram illustrating an example of a flowchart of score calculation processing according to the embodiment. FIG. 8 is a diagram illustrating an example of a flowchart of deduction rule extraction processing according to the embodiment. FIG. 9 is a diagram illustrating an example of a flowchart of type estimation processing according to the embodiment. FIG. 10A is a diagram (1) illustrating an example of a type estimation application according to the embodiment; FIG. 10B is a diagram (2) illustrating an example of a type estimation application according to the embodiment; FIG. 11 is a diagram showing an example of a computer that executes a type estimation program. FIG. 12 is a diagram showing a reference example of the type estimation method.

Embodiments of the type estimation method, the information processing apparatus, and the type estimation program disclosed in the present application will be described in detail below with reference to the drawings. In addition, this invention is not limited by an Example.

First, a reference example of the type estimation method will be described with reference to FIG. Here, in order to automatically generate a query that can fill the target table T100 as much as possible, a case will be described in which the schema information of the table T100 is input and the table type is estimated from within the knowledge graph. DBpedia, for example, is used as information included in the knowledge graph.

FIG. 12 is a diagram showing a reference example of the type estimation method. As shown in FIG. 12, it is assumed that the schema information of the target table T100 is "name, location, capital, business description, related organization".

An application that performs type estimation (hereinafter abbreviated as application) inputs schema information (S100), and calculates a statistical score corresponding to the schema information from the knowledge graph (S110). Statistical scores are calculated using the tf-idf method, as an example.

As an example of score calculation, the application regards the type as a document, and the predicate (here, each attribute name included in the schema information) of an instance to which the type is assigned as a word. The application calculates the appearance frequency of each word in the document from the knowledge graph, and weights the appearance frequency (tf) by tf-idf. The idf indicates how many documents each word is commonly used in, and words that are used in several documents are not important and have a low score. The app then totals the scores for each word by type to calculate a score for each type. Here, if the type is "Company", the number of instances is 214,734. An example of an instance (entity) includes an actual company name. Examples of predicates that an instance has include "name", "location", "capital", "business description", and "affiliated organization". The score when the predicate is "name" is 5.85, the score when it is "location" is 3.08, and the score when it is "capital" is 44.68. The score for "business content" is 44.1, and the score for "related organization" is 13.44. Then, the application calculates a score of 111.15 when the type is "company". Similarly, the app calculates a score of 146.91 when the type is "Organization".

The application then estimates the type with the highest score as the type of table T100 (S120). Here, the score 146.91 for the type "organization" is higher than the score 111.15 for the type "company". Therefore, the type of table T100 is presumed to be "organization".

After this, the app automatically generates a query that can fill the target table T100 as much as possible using the inferred type. Then, the application uses the generated query to set information from the knowledge graph to the column of the column name in which information is not set in the target table 100 (S130). Here, since the type is "organization", the formal name of the organization name "FLE" is set as the fictitious organization "Farm Leader Education Organization" (c100). Furthermore, the capital is set at "10 billion yen" (c110), and related organizations are not set (c120).

In other words, if the entered schema information is "name, location, capital, business description, affiliated organization", "company" is more likely than "organization" as the type, so "company" should be used as the type. is desirable. However, if the app uses only statistical scores, it will calculate a high score for the general (higher) type “organization”. Therefore, the application may not be able to properly estimate the type of the target table T100.

Therefore, a type estimating device for appropriately estimating the type of a target table will be described below.

[Functional configuration of type estimation device]
FIG. 1 is a block diagram showing the functional configuration of the type estimation device according to the embodiment. As shown in FIG. 1, when estimating the type of a table in which column names and columns corresponding to column names are associated from a knowledge graph, the type estimation device 1 adds scores statistically obtained from the entire knowledge graph to Then, a more appropriate type is estimated using the ontology attached to the knowledge graph. An ontology is a definition of relationships between concepts included in a knowledge graph. DBpedia is an example of information included in the knowledge graph. The type estimation device 1 is an example of an information processing device.

Here, an image of the type estimation method executed by the type estimation device 1 according to the embodiment will be described with reference to FIG. FIG. 2 is a diagram showing an image of the type estimation method according to the embodiment. As shown in FIG. 2, the type estimation device 1 estimates the type of the target table from the knowledge graph 21 by combining the statistical method and the deductive method. It is assumed that the schema information of the target table is "name, location, capital, business description, related organization".

The statistical method is the same method as the type estimation method shown in FIG. That is, the statistical method calculates statistical scores corresponding to schema information from knowledge graph 21 . The calculated statistical scores are shown in the left figure. A statistical score is a score for each predicate of an instance to which a type is assigned, and is set for each type. Note that the statistical score corresponds to a statistical information DB (DataBase) 22, which will be described later. The statistical method then sums the scores of each predicate by type to compute a statistical score for each type.

The deductive method uses an ontology 23' attached to the knowledge graph 21 to extract deductive rules. The deduction rule here means a rule indicating the hierarchical relationship between concepts (types) defined in the ontology 23'. The ontology 23' is shown in the right figure. This ontology 23' shows the hierarchical relationship between types from the concept (type) located at the root downward. The deduction rule corresponds to the deduction rule DB 23, which will be described later.

The type estimation device 1 estimates the type of the target table by applying the deduction rule using the ontology 23' in addition to the statistical score for each type calculated by the statistical method. As a result, the type estimation device 1 can appropriately estimate the type of the target table by using not only the statistical method but also the knowledge maintained in the knowledge graph 21 .

Returning to FIG. 1 , the type estimation device 1 has a statistical score calculator 11 , a deduction rule extraction unit 12 , a statistical estimation unit 13 , a deduction rule application unit 14 and a type determination unit 15 . These functional units are included in a control unit (not shown). The control unit corresponds to an electronic circuit such as a CPU (Central Processing Unit). The control unit has an internal memory for storing programs defining various processing procedures and control data, and executes various processing using these. Note that the statistical estimation unit 13 is an example of an identification unit. The deduction rule application unit 14 and the type determination unit 15 are examples of an estimation unit.

The type estimation device 1 also has a knowledge graph 21 , a statistical information DB 22 and a deduction rule DB 23 . These functional units are included in a storage unit (not shown). The storage unit is, for example, a semiconductor memory device such as a RAM or a flash memory, or a storage device such as a hard disk or an optical disk.

The knowledge graph 21 is data obtained by converting information such as DBpedia into LOD (Linked Open Data) format. Note that the knowledge graph 21 is generally open to the public using, for example, a cloud system, and can be used arbitrarily.

The statistical information DB 22 stores statistical information calculated by statistical methods. The statistical information DB 22 is generated by the statistical score calculator 11, which will be described later.

The deduction rule DB 23 stores deduction rules generated by deductive techniques. The deduction rule DB 23 is generated by the deduction rule extraction unit 12, which will be described later.

The statistical score calculator 11 constructs a statistical information DB 22 for statistically estimating the type from the knowledge graph 21 .

For example, when estimating the type of the target table, the statistical score calculation unit 11 regards the types included in the knowledge graph 21 as documents, and the predicates of instances to which types are assigned as words. Then, the statistical score calculator 11 calculates the appearance frequency (tf) of each word in the document from the knowledge graph 21 . Inquiries about types and frequencies of appearance can be made using queries, respectively.

Then, the statistical score calculation unit 11 calculates a score by weighting the appearance frequency of words by tf-idf. That is, the statistical score calculation unit 11 weights the frequency of appearance of words with a value indicating the uniqueness of each word within a plurality of documents to calculate a score. In other words, the statistical score calculator 11 weights low scores for words (predicates) that are used in many documents without being important. Then, the statistical score calculation unit 11 stores the calculated score in the statistical information DB 22 for each type (document) and for each predicate (word) of the instance to which the type is assigned.

An example of a query used by the statistical score calculator 11 will now be described with reference to FIG. FIG. 3 is a diagram showing an example of a query used for statistical score calculation. The query shown in the upper diagram of FIG. 3 is a query for all types from the knowledge graph 21 . The query shown in the lower diagram of FIG. 3 is a query for inquiring the frequency of occurrence of predicates possessed by instances to which types are assigned for all types in the knowledge graph 21 .

Also, an example of the data structure of the statistical information DB 22 will be described with reference to FIG. FIG. 4 is a diagram illustrating an example of the data structure of a statistical information DB; As shown in FIG. 4, the statistical information DB 22 stores types 22a, predicates 22b, and scores 22c in association with each other. A type 22a refers to a concept included in the knowledge graph 21 . The predicate 22b refers to a word possessed by an instance (entity) to which a type is assigned. Score 22c refers to a statistical score calculated by tf-idf.

As an example, when the type 22a is "company", "name" is stored as the predicate 22b and "5.85" is stored as the score 22c. "Location" is stored as the predicate 22b, and "3.08" is stored as the score 22c. "Capital" is stored as the predicate 22b, and "44.68" is stored as the score 22c. "Business content" is stored as the predicate 22b, and "44.1" is stored as the score 22c. "Related organization" is stored as the predicate 22b, and "13.44" is stored as the score 22c.

Returning to FIG. 1 , the deduction rule extraction unit 12 extracts deduction rules from the knowledge graph 21 . For example, the deduction rule extraction unit 12 extracts deduction rules from an ontology that defines conceptual relationships between words included in the knowledge graph 21 . The deduction rule extraction unit 12 stores the extracted deduction rule in the deduction rule DB 23 . Queries for extracting deduction rules from an ontology can be made using queries. As an example, the query includes a query for extracting sibling deduction rules and a query for extracting parent-child deduction rules.

The statistical estimation unit 13 estimates type candidates corresponding to a set of column names from statistical scores.

For example, when the statistical estimation unit 13 receives a set of column names from the type determination unit 15 described later, the statistical information DB 22 extracts the type 22a, the predicate 22b, and the score 22c corresponding to the set of column names. The statistical estimation unit 13 totals the scores 22c associated with the predicate 22b for each type 22a, and sets the total value as a statistical score. The statistical estimation unit 13 estimates a plurality of types with high statistical scores as type candidates. In other words, when estimating the type of the target table, the statistical estimation unit 13 estimates type candidates such that words highly correlated with each of the column names in the table are prioritized as the type. The statistical estimation unit 13 then outputs the statistical score for each estimated type candidate to the type determination unit 15 .

The deduction rule applying unit 14 applies the deduction rule to obtain a correction score corresponding to the type candidate.

For example, when the deduction rule application unit 14 receives a type candidate from the type determination unit 15 described later, the deduction rule application unit 14 acquires a correction score corresponding to the type candidate based on the deduction rule DB 23 . That is, the deduction rule application unit 14 acquires the depth of the ontology corresponding to the type candidate as a correction score based on the deduction rule DB 23 . The deeper the depth, the more specific the concepts (types) shown. Then, the deduction rule applying unit 14 outputs the corrected score for each acquired type candidate to the type determining unit 15 .

The type determination unit 15 uses the statistical estimation unit 13 and the deduction rule application unit 14 to determine the type of the target table.

For example, the type determination unit 15 inputs a set of column names of the target table. The set of column names of the table referred to here indicates, for example, schema information. The type determination unit 15 outputs the input set of column names to the statistical estimation unit 13 in order to estimate type candidates corresponding to the set of column names based on the statistical scores. Further, the type determination unit 15 outputs the estimated type candidate to the deduction rule application unit 14 in order to adjust the statistical score of the type candidate estimated by the statistical estimation unit 13 . Then, the type determination unit 15 calculates a score for each type candidate based on the statistical score for each type candidate and the correction score for each type candidate obtained from the deduction rule by the deduction rule application unit 14 . The type determination unit 15 determines the type from the type candidates so that the higher the calculated score, the higher the type. In other words, the type determination unit 15 determines the type of the target table from the type candidates estimated by the statistical estimation unit 13 so that words (predicates) whose type candidates are relatively specific are prioritized as types. Estimate (judgment). Then, the type determination unit 15 outputs the estimated (determined) type.

[An example of a deduction rule DB schema]
FIG. 5 is a diagram illustrating an example of a deduction rule DB schema according to the embodiment; Note that the left diagram of FIG. 5 shows the ontology given to the knowledge graph 21 . The ontology indicates a hierarchical relationship between types existing below "Thing" as a concept (type) located at the root. For example, if the type is "Organization", "Agent" is a subclass of "Thing", "Organization" and "Person" are subclasses of "Agent", and "Company" and "Legislative" are subclasses of "Organization". ” exists. As an example, the relationship between "Thing" and "agent" is a parent-child relationship. The relationship between "organization" and "company" is a parent-child relationship. As an example, the relationship between "organization" and "person" is a sibling relationship. The relationship between the "company" and the "legislature" is that of a brother. Then, the depth of the ontology increases by 1, with "Thing" being "1".

The deduction rule DB 23 generated from such an ontology stores information that associates sources 23a, relationships 23b, and targets 23c. The source 23a indicates a higher class (type) of the target 23c, which will be described later. The target 23c indicates a class (type) that is subordinate to the source 23a. A relationship 23b indicates the relationship between the higher order and the lower order. As an example, the relation 23b "subClassOf" indicates that the lower class is a subclass of the higher class.

As an example, when the source 23a is "Thing", "subClassOf" is stored as the relation 23b and "agent" is stored as the target 23c. When the source 23a is "agent", "subClassOf" is stored as the relationship 23b, and "organization" and "person" are stored as the targets 23c. When the source 23a is "organization", "subClassOf" is stored as the relation 23b, and "company" and "legislature" are stored as the targets 23c.

[An example of type estimation]
FIG. 6 is a diagram illustrating an example of type estimation according to the embodiment. In addition, in FIG. 6, it is assumed that the statistical information DB 22 is generated by the statistical score calculator 11 . Also, it is assumed that the deduction rule DB 23 is generated by the deduction rule extraction unit 12, and an ontology 23' having the same content as the deduction rule DB 23 is shown.

Under such circumstances, when the type determination unit 15 receives a set of column names of the target table, “name, location, capital, business description, related organization,” the set of column names is sent to the statistical estimation unit 13. Output.

Upon receiving the set of column names from the type determination unit 15, the statistical estimation unit 13 extracts the type 22a, predicate 22b, and score 22c corresponding to the set of column names from the statistical information DB 22. FIG. Here, the predicate 22b and the score 22c when the type 22a is "company" and the predicate 22b and the score 22c when the type 22a is "organization" are extracted.

Then, the statistical estimation unit 13 totals the scores 22c associated with the predicate 22b for each type 22a to calculate a statistical score. Here, if the type 22a is "Company", the statistical score is calculated as 111.15. If the type 22a is "Organization", the statistical score is calculated as 146.91. The statistical estimation unit 13 estimates a plurality of types with high total statistical scores as type candidates. Here, it is assumed that "company" and "organization" are estimated as type candidates. The statistical estimation unit 13 outputs the statistical score for each estimated type candidate to the type determination unit 15 . Here, the statistical score “111.15” for the type candidate “company” and the statistical score “146.91” for the type candidate “organization” are output to the type determination unit 15 .

The type determination unit 15 then outputs the type candidates estimated by the statistical estimation unit 13 to the deduction rule application unit 14 .

Upon receiving a type candidate from the type determination unit 15, the deduction rule application unit 14 acquires the depth of the ontology 23' as a correction score corresponding to the type candidate based on the deduction rule DB 23 (ontology 23'). Here, the deduction rule application unit 14 acquires the depth "4" of the ontology 23' as the correction score corresponding to the type candidate "company". The deduction rule application unit 14 acquires the depth "3" of the ontology 23' as the correction score corresponding to the type candidate "organization". The greater the depth of the ontology 23', the more specific the concepts (types) shown. In other words, "company" is relatively more specific than "organization".

The deduction rule application unit 14 then outputs the corrected score for each type candidate to the type determination unit 15 . Here, the corrected score “4” for the type candidate “company” and the corrected score “3” for the type candidate “organization” are output to the type determination unit 15 .

Then, the type determination unit 15 calculates the score for each type candidate based on the statistical score for each type candidate estimated by the statistical estimation unit 13 and the correction score for each type candidate obtained by the deduction rule application unit 14. to calculate Here, the score calculation method is an example in which the statistical score and the correction score have the same weight, and is a method of multiplying the statistical score and the depth. Specifically, if the type candidate is "Company", the score is calculated as "444.60" obtained by multiplying the statistical score "111.15" and the depth "4". Also, if the type candidate is "tissue", the score is calculated as "440.73" obtained by multiplying the statistical score "146.91" and the depth "3". Then, the type determination unit 15 estimates (determines) the type from the type candidates so that the higher the calculated score, the higher the type. Here, since the score of the type candidate "company" is "444.60" and the score of the type candidate "organization" is "440.73", the type candidate "company" is higher than the type candidate "organization". High score. Therefore, the type candidate "company" is estimated (determined) as the type.

Then, the type determination unit 15 outputs the determined type. Here, "Company" is output as the type.

After this, the type estimation device 1 automatically generates a query that can fill the target table as much as possible using the determined type. Then, using the generated query, the type estimating device 1 sets information from the knowledge graph 21 to columns of column names in which information is not set in the target table. Here, since the type is "company", the official name of the organization name is "FLE" and "F Laboratories of Europe" is set (c10). Further, the capital is set as "" (c11), and the affiliated organization is set as "F" (c12).

That is, if the set of entered column names is "name, location, capital, business description, related organization", the more specific (lower) type "company" than "organization" is more plausible. By estimating (determining) the type as "company", the type estimation device 1 can fill the target table with correct information.

Although it has been described that the type estimation device 1 uses the type with the highest score in order to automatically generate the relevant query, it is not limited to this. The type estimating device 1 may output type candidates whose priority is determined, and use a type selected by the user from the output type candidates.

[Flowchart of Score Calculation Processing] The deduction rule extraction unit 12
FIG. 7 is a diagram illustrating an example of a flowchart of score calculation processing according to the embodiment. Assume that the statistical score calculation unit 11 receives a statistical score calculation request, for example.

As shown in FIG. 7, the statistical score calculator 11 that has received the statistical score calculation request acquires all types in the knowledge graph 21 (step S11). For example, the statistical score calculator 11 acquires all types in the knowledge graph 21 using a query (query for all types from the knowledge graph 21) as shown in the upper diagram of FIG.

The statistical score calculator 11 stores all types in the knowledge graph 21 in the type 22a of the statistical information DB 22 (step S12). The statistical score calculation unit 11 regards the type as a document, regards the predicate of the instance to which the type is assigned as a word, and calculates the appearance frequency of the word in the document (step S13). For example, the statistical score calculation unit 11 uses a query (query for querying the appearance frequency of a predicate possessed by an instance to which a type is assigned) as shown in the lower diagram of FIG. Calculate the predicate frequency.

The statistical score calculation unit 11 calculates tf-idf using the type and the appearance frequency of the predicate (step S14). That is, the statistical score calculation unit 11 calculates a score by weighting the appearance frequency of words (predicates) in documents (types) with a value indicating the uniqueness of each word in a plurality of documents.

The statistical score calculator 11 stores the calculation result in the statistical information DB 22 (step S15). For example, the statistical score calculation unit 11 stores the calculated score in the statistical information DB 22 for each type (document) and for each predicate (word) of the instance to which the type is assigned. Then, the statistical score calculation unit 11 terminates the statistical score calculation process.

[Flowchart of Deduction Rule Extraction Processing]
FIG. 8 is a diagram illustrating an example of a flowchart of deduction rule extraction processing according to the embodiment. Assume that the deduction rule extraction unit 12 receives a deduction rule extraction request, for example.

As shown in FIG. 8, the deduction rule extraction unit 12 that receives the deduction rule extraction request initializes the queue (Q) (step S21). Then, the deduction rule extraction unit 12 puts root (Thing), which is the leading node of the ontology assigned to the knowledge graph 21, into Q (step S22).

The deduction rule extraction unit 12 determines whether or not Q is empty (step S23). If it is determined that Q is not empty (step S23; No), the deduction rule extraction unit 12 extracts node (v) from Q (step S24). The deduction rule extraction unit 12 acquires a sibling node list (Sibling) of the extracted node (v) (step S25). For example, the deduction rule extraction unit 12 may use a query for extracting sibling deduction rules.

The deduction rule extraction unit 12 registers all sibling relationships in Sibling in the deduction rule DB 23 (step S26). As an example, when the selected node is "organization" shown in FIG. 5, "organization" and "person" are siblings. If the selected node is "Company", then "Company" and "Legislation" are siblings.

The deduction rule extraction unit 12 determines whether or not all nodes (v') in Sibling are stored in Q (step S27). If it is determined that all of them are not stored in Q (step S27; No), the deduction rule extraction unit 12 extracts sibling nodes (v') that are not stored in Q in Sibling, and are stored in the deduction rule DB 23 as a parent-child relationship (step S28).

Then, the deduction rule extraction unit 12 stores v' in Q (step S29). Then, the deduction rule extraction unit 12 shifts to step S27 in order to store all of the nodes (v') in Sibling in Q.

On the other hand, if it is determined that all are stored in Q (step S27; Yes), the deduction rule extraction unit 12 proceeds to step S23 to determine whether or not Q is empty.

Regarding step S23, when it is determined that Q is empty (step S23; Yes), the deduction rule extraction unit 12 terminates the deduction rule extraction process.

[Type estimation processing flow chart]
FIG. 9 is a diagram illustrating an example of a flowchart of type estimation processing according to the embodiment. It is assumed that the type determination unit 15 receives a type determination request, for example.

As shown in FIG. 9, the type determination unit 15 that has received the type determination request receives a set of column names of the target table (step S31). The type determination unit 15 outputs the received set of column names to the statistical estimation unit 13 .

Subsequently, the statistical estimation unit 13 reads the type corresponding to the set of column names from the statistical information DB 22 (step S32). The statistical estimation unit 13 determines whether or not the read number is 0 (step S33). If it is determined that the number of readouts is 0 (step S33; Yes), the statistical estimation unit 13 terminates the type estimation process.

On the other hand, when it is determined that the number of readouts is not 0 (step S33; No), the statistical estimation unit 13 calculates a statistical score for the readout type and estimates the type from the statistical score (step S34). For example, the statistical estimation unit 13 totals the scores 22c associated with the predicate 22b for each type 22a read from the statistical information DB 22 to calculate the statistical score. The statistical estimation unit 13 estimates a plurality of types with high statistical scores as type candidates. Then, the statistical estimation unit 13 associates the estimated type with the statistical score (step S35).

Subsequently, the deduction rule application unit 14 applies the deduction rule to the estimated type to obtain a corrected score (step S36). For example, the deduction rule application unit 14 acquires a correction score corresponding to the type candidate estimated by the statistical estimation unit 13 based on the deduction rule DB 23 . As an example, the deduction rule application unit 14 acquires the depth of the ontology corresponding to the type candidate as a correction score.

The type determination unit 15 then calculates a final score for the estimated type using the statistical score and the correction score (step S37). For example, for each type candidate estimated by the statistical estimation unit 13, the type determination unit 15 determines the final score to calculate

Then, the type determination unit 15 outputs the type with the highest score to the screen (step S38). The type determination unit 15 then terminates the type estimation process.

[Application example of type estimation]
10A and 10B are diagrams illustrating an example application of type estimation according to an embodiment.

As shown in FIG. 10A, the set of column names of the target table is "company name, company name, location, capital, main business, office, president, head office, shareholder". In such a case, the type determination unit 15 outputs "company" as the type.

As shown in FIG. 10B, this is the case where the set of column names of the target table is "store name, address, access, business hours, closed days, fee, budget, telephone number, parking lot, homepage". In such a case, the type determination unit 15 outputs "restaurant" as the type.

In the embodiment, the type determination unit 15 determines the type based on the statistical score for each type candidate estimated by the statistical estimation unit 13 and the correction score for each type candidate obtained by the deduction rule application unit 14. Compute a score for each candidate. The score calculation method is described as a method of multiplying the statistical score and the corrected score when the statistical score and the corrected score have the same weight. However, the score calculation method is not limited to this, and the statistical score and the correction score may have different weights. For example, the final score f may be calculated by the following formula (1). Note that α in the expression (1) is a real number greater than 0.0 and less than 1.0.
f = α × statistical score × (1-α) × corrected score formula (1)

In addition, in the embodiment, the type determination unit 15 uses the statistical score and the correction score obtained by the deduction rule application unit 14 for the type candidate estimated using the statistical score by the statistical estimation unit 13, The final type candidates were scored. However, the type determination unit 15 is not limited to this, and may selectively exclude (filter) type candidates estimated using the statistical scores by the statistical estimation unit 13 using deduction rules. For example, the type determination unit 15 may filter type candidates whose depth is less than a predetermined value using a deductive rule. Such a predetermined value may be determined, for example, by a set of input column names of the target table.

Moreover, in the embodiment, the hierarchical relationship of concepts (types) is used as the deduction rule stored in the deduction rule DB 23 . However, the deduction rule is not limited to this, and ontology information other than the concept (type) hierarchy may be used. Further, the information of the ontology to be used is not limited to the information of the ontology 23' used in the embodiment, and may be information of any ontology.

[Effect of Example]
According to the above embodiment, the type estimation device 1 performs the following processing when estimating the type of the table in which the column name and the column corresponding to the column name are associated. The type estimation device 1 identifies table type candidates based on the types included in the knowledge graph 21 and the appearance frequency of the predicate of the instance of each type. The type estimation device 1 estimates the table type from among the specified type candidates based on the information on the depth of the ontology that defines the relationship between the types included in the knowledge graph 21 . According to such a configuration, the type estimation device 1 estimates the optimum type of the table by using information on the depth of the ontology that defines the relationship between types, in addition to the appearance frequency of the instance predicate for each type. can do.

Further, according to the above-described embodiment, the type estimation device 1 estimates the table type from among the identified type candidates so that the deeper the depth of the ontology, the higher the priority. According to such a configuration, the type estimation device 1 can give priority to a specific (lower-order) type of concept to be indicated while taking into account the appearance frequency of the predicate of the instance of each type.

Further, according to the above-described embodiment, the type estimation device 1 uses the estimated type to set information from the knowledge graph 21 in the columns of column names in which information is not entered in the table. According to such a configuration, the type estimation device 1 uses the type estimated using the information on the depth of the ontology that defines the relationship between the types, in addition to the appearance frequency of the instance predicate for each type. Improve the accuracy of the information set in the column of column name.

[others]
It should be noted that each component of the illustrated type estimation device 1 does not necessarily have to be physically configured as illustrated. That is, the specific mode of distribution/integration of the type estimation device 1 is not limited to the illustrated one, and all or part of it can be functionally or physically implemented in arbitrary units according to various loads and usage conditions. It can be distributed and integrated. For example, the statistical estimation unit 13, the deduction rule application unit 14, and the type determination unit 15 may be integrated as one unit. Further, the type determination unit 15 may be separated into an input/output unit for input/output and a determination unit for determining the type. Also, a storage unit (not shown) may be connected to the type estimation device 1 as an external device via a network.

Moreover, various processes described in the above embodiments can be realized by executing a prepared program on a computer such as a personal computer or a workstation. Therefore, an example of a computer that executes a type estimation program that implements the same functions as those of the type estimation device 1 shown in FIG. 1 will be described below. FIG. 11 is a diagram showing an example of a computer that executes a type estimation program.

As shown in FIG. 11, the computer 200 has a CPU 203 that executes various arithmetic processes, an input device 215 that receives data input from the user, and a display control section 207 that controls the display device 209 . The computer 200 also has a drive device 213 that reads programs and the like from a storage medium, and a communication control unit 217 that exchanges data with other computers via a network. The computer 200 also has a memory 201 that temporarily stores various information and a HDD (Hard Disk Drive) 205 . The memory 201 , CPU 203 , HDD 205 , display control section 207 , drive device 213 , input device 215 and communication control section 217 are connected via a bus 219 .

The drive device 213 is a device for the removable disk 210, for example. HDD 205 stores type estimation program 205a and type estimation processing related information 205b.

The CPU 203 reads the type estimation program 205a, develops it in the memory 201, and executes it as a process. Such a process corresponds to each functional unit of the type estimating device 1 . The type estimation processing related information 205b corresponds to the knowledge graph 21, the statistical information DB 22 and the deduction rule DB 23. Then, for example, the removable disk 210 stores information such as the type estimation program 205a.

Note that the type estimation program 205a does not necessarily have to be stored in the HDD 205 from the beginning. For example, a flexible disk (FD), a CD-ROM (Compact Disk Read Only Memory), a DVD (Digital Versatile Disk), a magneto-optical disk, an IC (Integrated Circuit) card inserted into the computer 200, or other "portable physical medium ” to store the program. Then, the computer 200 may read and execute the type estimation program 205a from these.

1 type estimation device 11 statistical score calculation unit 12 deduction rule extraction unit 13 statistical estimation unit 14 deduction rule application unit 15 type determination unit 21 knowledge graph 22 statistical information DB
23 Deduction rule DB

Claims (8)

When estimating the type of the table in which the column name and the column corresponding to the column name are associated, based on the type included in the graph data and the appearance frequency of the instance predicate for each type, the type of the table identify candidates,
A type estimation method in which a computer executes a process of estimating the type of the table from among the identified type candidates based on information about the depth of an ontology that defines the relationship between types included in the graph data. 2. The type estimation method according to claim 1, wherein a computer executes the process of estimating the type of the table from among the identified type candidates so that the deeper the depth of the ontology, the higher the priority. . 3. The computer according to claim 1 or 2, wherein the estimated type is used to set information in the columns of the column names in the table in which information is not entered from the graph data. type estimation method. When estimating the type of the table in which the column name and the column corresponding to the column name are associated, based on the type included in the graph data and the appearance frequency of the instance predicate for each type, the type of the table an identification part that identifies a candidate;
an estimating unit for estimating the type of the table from among the identified type candidates based on information on the depth of an ontology that defines the relationship between types included in the graph data;
An information processing device comprising: When estimating the type of the table in which the column name and the column corresponding to the column name are associated, based on the type included in the graph data and the appearance frequency of the instance predicate for each type, the type of the table identify candidates,
A type estimation program that causes a computer to execute a process of estimating the type of the table from among the specified type candidates based on information about the depth of an ontology that defines the relationship between types included in the graph data. When estimating a type of a table in which a column name and a column corresponding to the column name are associated, specifying the candidates for the type such that words highly correlated with each of the column names in the table are prioritized as the type. death,
A type estimation method in which a computer executes a process of estimating the type of the table so that, from among the identified type candidates, the word having a relatively specific concept is prioritized as the type. When estimating a type of a table in which a column name and a column corresponding to the column name are associated, specifying the candidates for the type such that words highly correlated with each of the column names in the table are prioritized as the type. a specific part to
an estimating unit for estimating the type of the table so that a word with a relatively more specific concept is prioritized as the type from among the type candidates specified by the specifying unit;
An information processing device comprising: When estimating a type of a table in which a column name and a column corresponding to the column name are associated, specifying the candidates for the type such that words highly correlated with each of the column names in the table are prioritized as the type. death,
A type estimation program that causes a computer to execute a process of estimating the type of the table so that, from among the specified type candidates, words that indicate relatively specific concepts are prioritized as the type.

Images