JP2011039838A - Pattern classification device and pattern classification method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a pattern classification device and a pattern classification method for classifying a plurality of patterns satisfactorily. <P>SOLUTION: A query issuing part 14 generates the number of N×M-pieces of retrieval patterns by making a first node in the number of N-pieces of patterns for retrieving subgraphs in a graph G include the number of M-pieces of keywords different from one another (S11) and retrieves the subgraph agreeing with each retrieval pattern from the graph G (S15). A pattern classification part 15 classifies the number of N-pieces of patterns based on an instance in a second node in the retrieved subgraph (S19). <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a pattern classification device and a pattern classification method.

  In recent years, there are a large number of data sources on computer networks, and technologies such as data warehouses that combine multiple data sources to extract information that cannot be extracted from a single data source are attracting attention. Collecting.

  On the other hand, as a framework for handling information obtained from a plurality of different data sources in a unified manner, a semantic Web technology using RDF (Resource Description Framework), which is a data model that can be represented in a graph, has attracted attention.

  The Semantic Web proposes RDF search technology that searches for necessary information by matching search patterns (hereinafter simply referred to as patterns) using an RDF query language such as SPARQL.

  Non-Patent Document 1 is a system that searches a natural sentence including a keyword character string and uses a similar query as a keyword having similar search results, and provides a mechanism for classifying each similar query and proposing a similar query to the user. Is disclosed. However, there were the following problems.

  Queries referred to in these techniques correspond to the search keyword variables of the pattern, and no reference is made to the portion corresponding to the pattern. Therefore, pattern classification cannot be performed.

  For example, Non-Patent Document 2 proposes a structural similarity that reflects the characteristics of the target graph set, where the structural similarity is defined using a characteristic partial structure, and the similarity of the subgraphs is defined. Make a decision.

  However, this technology targets unlabeled undirected graphs, and cannot be applied directly to labeled directed graphs such as RDF.

  When selecting a pattern to be used as a query for searching graph structure data with labels on nodes and arcs such as RDF, if the structure of the graph is complex, find a pattern that can perform the intended search Therefore, it is necessary to select a pattern efficiently.

  In particular, in the case of a graph combined from a plurality of different data sources, semantic duplication is included in the graph, and there are many patterns having different structures that are semantically similar, so that the selection operation becomes complicated.

Toru Onoda, Takayuki Yumoto, Kazutoshi Kakutani, “Topic clustering based on partial similarity of search tendency”, Transactions of the Database Society of Japan Vol.7, No.3, pp.49-54, December 2008 Takahisa Wada, Hiroyuki Ohno, Hiromasa Inazumi, “Proposal of Structural Similarity Reflecting Characteristics of Object Graph Sets”, Database Society of Japan Letters Vol.6, No.1, pp.185-188, June 2007

  The present invention has been made in view of the above, and an object thereof is to provide a pattern classification apparatus and a pattern classification method capable of classifying a plurality of patterns.

  In order to solve the above problems, a pattern classification apparatus according to the present invention includes a graph database in which a graph in which nodes having instances are connected by arcs is stored, and N graphs for searching subgraphs in the graph. Graph search means for generating N × M search patterns by including M different keywords for the first node in the pattern, and searching the graph for subgraphs matching the search patterns; Using the instance in the second node in the searched subgraph, obtain a degree of similarity for each of the combinations of the two patterns in the N patterns, and determine the N patterns based on the degree of similarity. Pattern classification means for classifying.

  The pattern classification method according to the present invention is a pattern classification method performed by a pattern classification apparatus including a graph database in which a graph in which nodes having instances are connected by arcs is stored. Means generates N × M search patterns by including M keywords different from each other for the first node in the N patterns for searching the subgraph in the graph, and generates the N × M search patterns from the graph; Sub-graphs that match each search pattern are searched, and the pattern classification unit of the pattern classification device uses a combination of two patterns in the N patterns by using an instance in a second node in the searched sub-graph. Find the degree of similarity for each and based on the degree of similarity Characterized by classifying the N patterns.

The pattern classification means classifies the N patterns into a pattern cluster including one or more patterns based on the degree of similarity, and each of the keywords is classified into one of two patterns. A set of instances in the second node in one or more subgraphs that match the search pattern including the keyword is A, and one or more subgraphs that match the search pattern that includes the keyword in the other of the two patterns Let B be the set of instances in the second node at
Similarity judgment value = | A∩B | ÷ | A∪B |
However,
A and B are not empty sets,
| A∩B | is the number of instances in the intersection of A and B,
| A∪B | is the number of instances in the union of A and B,
When calculating
The two patterns in which k or more (0 <k ≦ K: K is the number of the keywords) among the similarity determination values corresponding to the keywords are equal to or greater than a predetermined threshold value are similar to each other. These may be included in the same pattern cluster.

Alternatively, the pattern classification unit may perform processing as follows.
First, from the N × M search patterns, the search patterns used in the search that could not obtain the corresponding subgraph are excluded.
Next, when a subgraph is obtained from a search pattern obtained by including a common keyword in each of the two patterns, the two patterns are associated with each other, obtained from a plurality of patterns associated with each other, and excluded. One or more keywords included in a search pattern that is derived from a single pattern that is defined as a keyword cluster and that is not associated with other patterns and that is not excluded Is defined as a keyword cluster, and the plurality of keywords are classified into one or more keyword clusters.
Next, one keyword is selected from each of the keyword clusters, and the number of search patterns including the selected keyword and not excluded is maximized.
Next, for each of the keyword clusters, one or more patterns used to generate a search pattern that includes the selected keyword and is not excluded are selected.
Next, a plurality of patterns included in the selected one or more patterns and similar to each other are defined as a pattern cluster, and a single pattern included in the selected one or more patterns and not similar to other patterns The pattern is defined as a pattern cluster, the selected one or more patterns are classified into one or more pattern clusters, and any two patterns included in the former pattern cluster are included in one of the two patterns. A set of instances in the second node in one or more subgraphs matching the search pattern including the selected keyword is A, and the search pattern including the selected keyword is included in the other of the two patterns. If the set of instances in the second node in one or more subgraphs is B,
Similarity judgment value = | A∩B | ÷ | A∪B |
However,
A and B are not empty sets,
| A∩B | is the number of instances in the intersection of A and B,
| A∪B | is the number of instances in the union of A and B,
Is greater than or equal to a predetermined threshold.

  According to the present invention, it is possible to search for a subgraph that matches a search pattern obtained by including a keyword in the first node of the pattern, and to classify patterns when obtaining an instance in the second node in the subgraph as a search result, The number of patterns can be substantially reduced. In addition, when the user selects a pattern, the user can easily select the pattern.

It is a lineblock diagram of a graph search device concerning this embodiment. 6 is a diagram illustrating a part of a graph G. FIG. It is the figure which illustrated the subgraph by the data of the RDF / XML format, its original data, and the data of this format. It is the figure which illustrated the pattern as a graph. It is a sequence diagram which shows the operation | movement which classifies a pattern. It is a figure which shows the screen displayed when selecting a class by the classification of a pattern and specifying a threshold value. It is a figure which shows the search pattern produced | generated when classifying a pattern. It is a figure which shows the subgraph searched when classifying a pattern. It is a sequence diagram which shows the operation | movement which searches the instance of a node. It is a figure which shows the screen displayed when selecting a class by the search of an instance. It is a figure which shows the screen displayed when selecting a pattern by searching an instance and inputting a keyword. It is a figure which shows the screen which displays the searched instance. It is a figure which shows the pattern and keyword used by description of another method of classifying a pattern. It is a flowchart which shows operation | movement of step S19 in the method. It is a figure which shows the mode of the exclusion of a pattern in the method, an association, and the selection of a keyword. It is a figure which shows a mode that the set of instances is comprised in the method. It is a figure which shows the result of having classified 330 patterns in the method. It is a figure which shows the pattern contained in one pattern cluster obtained by the classification.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In addition, the same code | symbol is provided to the same or similar thing, and duplication description is abbreviate | omitted.

FIG. 1 is a configuration diagram of a graph search apparatus according to the present embodiment.
The graph search device 1 is connected to a user terminal 2, and a display device 3 is connected to the user terminal 2.

The graph search device 1 stores a user interface 11 that generates an input interface and an output interface displayed on the display device 3 and transmits them to the user terminal 2, and a graph in which nodes having instances are connected by arcs. There are a graph database 12, a pattern database 13 for storing patterns used for subgraph search, a query issuing unit 14 for searching for subgraphs and patterns, a pattern classifying unit 15 for classifying patterns, and a pattern cluster obtained by classification. And a pattern cluster database 16 to be stored.
As can be seen from the fact that the graph search device 1 includes the pattern classification unit 15, it also functions as a pattern classification device.

  FIG. 2 is a diagram illustrating a part of the graph G that can be displayed using all the data groups stored in the graph database 12.

  Using all the data groups stored in the graph database 12, the graph G illustrated in part in FIG. 2, that is, nodes having different instances are connected by arcs having labels, and classes of the instances are defined. A graph G can be displayed. Conversely, a data group for displaying the graph G is stored in the graph database 12. Hereinafter, the data group is referred to as a graph G for convenience. Also, for convenience, graphs and subgraphs can be used to display any graphs, subgraphs (some graphs themselves or graphs included in them), paths (graphs without branches and closed loops), etc. , Path and so on.

  The label is an identifier for identifying the type of arc, the class is a node indicating a concept to which each instance belongs, and the instance is a node indicating individual things other than the class.

  In the graph G, for example, nodes having instances such as “politics” and “Sachiko Yamamoto” are connected by an arc having a label such as “theme”. In the graph G, a class “theme”, which is a concept such as an instance “politics”, is defined in the node.

  As shown in FIG. 3, the original data is “graph F”, the author of the original data is Taro Yamada, the title is “Introduction to B Technology”, the keyword is B technology, and the original data is the graph database 12. Then, it becomes RDF / XML format data and is stored in the graph database 12, which forms a sub-graph of the graph G. What is entitled “RDF graph representation” is a graphical representation of this subgraph. RDF is described in the following documents.

"Resource Description Framework (RDF) Model and Syntax Specification", Ora Lassia, Ralph R. Swick, [online], Internet <URL: http://www.w3.org/TR/1999/REC-rdf-syntax- 19990222 />
"RDF Vocabulary Description Language 1.0: RDF Schema", Dan Brickley, RVGuha, [online], Internet <URL: http://www.w3.org/TR/rdf-schema/>
Returning to FIG. 1, the pattern database 13 stores patterns used for subgraph search.

  FIG. 4 is a diagram illustrating three patterns of the patterns stored in the pattern database 13 as graphs.

  The pattern is the same as the data group forming a part of the data group (graph G) stored in the graph database 12 and can be graphed as shown in FIG. The pattern is not displayed but is used for searching the displayed graph. Since it is redundant to explain the actual pattern as a data group one by one, it will be explained for convenience with a graphed pattern.

  In general, in a pattern, some nodes and arcs have instances and labels, and the rest do not. Variables are set for nodes and arcs that do not have instances or labels. Variables as shown in the figure? Followed by a word.

  Here, a node with class “theme” is defined at one end position, a node with class “organization” is defined at the other end position, each node has no instance, and each arc It is assumed that patterns P1, P2, and P3 having a label are stored in the pattern database 13. These are all patterns for knowing the organization from the theme. The pattern P1 has the meaning of “organization to which the theme belongs”, and the pattern P2 has the meaning of “organization to which the person in charge of the theme belongs”. The pattern P3 means “organization to which the person in charge of the theme belongs”.

  A subgraph retrieved from a certain graph by such a pattern has the following conditions.

  In other words, what is searched is (1) the graph or its subgraph, (2) having a pattern structure without excess or deficiency, (3) having instances or labels within the pattern without deficiency, That is, a node or arc at a position equal to the position of a node or arc having an instance or label in the pattern has an instance or label equal to the instance.

  If the condition of (3) is supplemented, for example, if the instance of the node at one end of the pattern is “A”, the instance of the node at the one end of the subgraph to be searched must also be “A”. Also, if the label of the only arc connected to the node at one end of the pattern is “B”, the label of the only arc connected to the node at one end of the subgraph is not “B”. Rather, such instance and label matching is required for all nodes and arcs that have instances and labels in the pattern.

  Note that searching for a subgraph by a pattern in this way is called searching for a subgraph that matches (also matches) the pattern.

  Returning to FIG. 1, the query issuing unit 14 searches for a pattern from the pattern database 13. In addition, the query issuing unit 14 searches for a subgraph of the graph G.

  The graph search device 1 only needs to be able to transmit and receive (deliver) data in each unit (including a database). That is, each unit may be arranged on the same computer or may be distributed on a plurality of computers. Further, a computer program that causes these computers to operate as a graph search device or a pattern classification device may be transmitted / received via a communication line. The computer program may be recorded on a recording medium such as a semiconductor memory, a magnetic disk, an optical disk, a magneto-optical disk, or a magnetic tape, and the recording medium may be distributed.

(Operation of this embodiment)
FIG. 5 is a sequence diagram showing an operation of classifying patterns in the graph search device 1.

  In the graph search device 1, the user interface 11 generates an input interface, transmits it to the user terminal 2 (S1), and displays it as shown in FIG. 6 (S3).

  Here, it is assumed that the user wants to know where an organization such as a company related to the theme of “politics” is located in a newspaper company, for example. In addition, the user wants to make the similarity determination between patterns stricter, and when the degree is set to 1 as the maximum, it is assumed that the user thinks that “0.7”.

  In this example, the information “theme” (class “theme”) and the information “organization” (class “organization”) are selected from the information included in the input interface. To do.

  In the user terminal 2, it is assumed that a value “0.7” (threshold value “0.7”) is designated by a user operation.

  The user terminal 2 transmits these parameters to the graph search device 1 (S5).

  In the graph search device 1, the query issuing unit 14 transmits a query, which is a search syntax including the class “theme”, to the graph database 12, whereby an instance in the node where the class “theme” is defined (here, The instance “politics”, “history”, and “science” (hereinafter referred to as keywords K1, K2, and K3) are searched from the graph database 12 (S7).

  The query issuing unit 14 transmits a query having a search syntax including the class “theme” and the class “organization” to the pattern database 13, thereby searching the pattern database 13 for patterns P 1, P 2 and P 3 (S 9).

  Next, the query issuing unit 14 includes the keywords K1, K2, and K3 in the patterns P1, P2, and P3 in which the class “theme” is defined (referred to as the first node) and includes 3 (pattern Number) × 3 (number of keywords) (total of 9) patterns (hereinafter referred to as search patterns P11 to P33) are generated (S11).

  FIG. 7 is a diagram showing these search patterns.

In this figure, the class is not shown. For example, the search pattern P11 includes the keyword K1 “politics” as an instance in the first node of the pattern P1.
The search pattern P12 includes the keyword K1 “history” as an instance in the first node of the pattern P1.
The search pattern P13 includes the keyword K1 “science” as an instance in the first node of the pattern P1.
The search pattern P21 includes the keyword K1 “politics” as an instance in the first node of the pattern P2.
The search pattern P22 includes the keyword K1 “history” as an instance in the first node of the pattern P2.
The search pattern P23 includes the keyword K1 “science” as an instance in the first node of the pattern P2.
The search pattern P31 includes the keyword K1 “politics” as an instance in the first node of the pattern P3.
The search pattern P32 includes the keyword K1 “history” as an instance in the first node of the pattern P3.
The search pattern P33 includes the keyword K1 “science” as an instance in the first node of the pattern P3.

  Returning to FIG. 5, the query issuing unit 14 converts the search patterns P11 to P33 into queries, and transmits them to the graph database 12, thereby acquiring a subgraph matching the search pattern from the graph G (S15).

  Here, it is assumed that subgraphs SG (P11) to SG (P33) as shown in FIG. 8 are obtained from the search patterns P11 to P33, respectively.

  Returning to FIG. 5, the query issuing unit 14 gives the patterns P1, P2, and P3, the subgraphs SG (P11) to SG (P33), the threshold value “0.7”, and the class “organization” to the pattern classification unit 15 ( S17).

  The pattern classification unit 15 classifies the pattern based on the subgraph, threshold value, and class (S19), and stores the classification result in the pattern cluster database 16 (S21).

Here, step S19 will be described in detail.
The pattern classification unit 15 classifies the patterns P1 to P3 into one or more pattern clusters. Here, a set including a plurality of similar patterns or a single pattern not similar to other patterns is referred to as a pattern cluster. Accordingly, the pattern classification unit 15 generates, for example, a pattern cluster including the patterns P1 and P2 and a pattern cluster including the pattern P3.

  At this time, the pattern classification unit 15 performs the following processing for all combinations of patterns in the patterns P1 to P3.

An example of a set of patterns P1 and P2 will be described.
The pattern classification unit 15 is a node (referred to as a second node) in which a class “organization” is defined in one or more subgraphs (here, subgraph SG (P11)) matching the search pattern P11 in which the keyword P1 is included in the pattern P1. ) Is a set of instances (here, “XX group only”) in R11, and a class in one or more subgraphs (here, subgraph SG (P21)) that matches the search pattern P21 including the keyword K1 in the pattern P2. A set of instances (here “XX group”) in a node (second node) in which “organization” is defined is R21,
If R11 and R21 are not empty sets, the patterns P1, P2 and the keyword K1 are
Similarity determination value T (P1, P2, K1) = | R11∩R21 | ÷ | R11∪R21 |
However,
| R11∩R21 | is the number of instances in the product set of R11 and R21,
| R11∪R21 | is the number of instances in the union of R11 and R21,
Is calculated, and it is determined whether or not the similarity determination value T (P1, P2, K1) is equal to or greater than the threshold value “0.7”.

  Here, P1, P2, and K1 in the similarity determination value T (P1, P2, and K1) mean that they relate to the patterns P1 and P2, and the keyword K1. The same applies to the similarity determination value described below.

  Here, the similarity determination value T (P1, P2, K1) is calculated for the patterns P1, P2 and the keyword K1, and it is determined whether or not the threshold is “0.7” or more.

Since R11∩R21 includes “XX group”, | R11∩R21 | = 1.
Since R11∪R21 includes “XX group”, | R11∪R21 | = 1.
Therefore, the similarity determination value T (P1, P2, K1) = 1 ÷ 1 = 1, and it is determined that the threshold value is “0.7” or more.

Similarly, the pattern classification unit 15 calculates a similarity determination value T (P1, P2, K2) for the patterns P1, P2, and the keyword K2, and determines whether or not the threshold is “0.7” or more.
Since the intersection set includes “XX group”, the number of instances in the intersection set = 1.
Since the union includes “XX group”, the number of instances in the union = 1.
Therefore, the similarity determination value T (P1, P2, K2) = 1 ÷ 1 = 1, and it is determined that the threshold is “0.7” or more.

Similarly, the pattern classification unit 15 calculates the similarity determination value T (P1, P2, K3) for the patterns P1, P2, and the keyword K3, and determines whether or not the threshold is “0.7” or more.
Since the product set is an empty set, the number of instances in the product set = 0.
Since the union includes “XX group” and “ΔΔ group”, the number of instances in the union = 2.
Therefore, the similarity determination value T (P1, P2, K3) = 0, and is determined to be less than the threshold value “0.7”.

  Next, for example, the pattern classification unit 15 has a threshold equal to or greater than k (0 <k ≦ K: where K is the number of keywords (here, “3”)) among the three similarity determination values. It is determined whether or not “0.7” or more. If k or more is equal to or greater than the threshold value, it is determined that the patterns P1 and P2 are similar. The value of k may be input as a parameter or may be a default value.

For example, if k = 1, it is determined that the patterns P1 and P2 are similar if at least one of the three similarity determination values is greater than or equal to the threshold value.
For example, if k = 3, if all of the three similarity determination values are equal to or greater than the threshold value, the patterns P1 and P2 are determined to be similar.

  When calculating the similarity determination value, if one or both of the two sets to be calculated are empty sets, the calculation is skipped. That is, not only the case where three similarity determination values are calculated without skipping, but also one or two similarity determination values smaller than that may be calculated.

  In this way, the pattern classification unit 15 generates, for example, a pattern cluster including the patterns P1 and P2 and a pattern cluster including the pattern P3 (S19).

  The pattern classification unit 15 stores each pattern cluster (classification result) in the pattern cluster database 16 (S21). Here, not only a pattern cluster including two or more patterns but also a pattern cluster including only one pattern is stored in the pattern cluster database 16.

  FIG. 9 is a sequence diagram showing an operation of searching for an instance of a node from the graph G in the graph search device 1.

  The user interface 11 generates an input interface, transmits it to the user terminal 2 (S51), and displays it as shown in FIG. 10 (S53).

  Here, the user is, for example, a user who classifies patterns, and wants to know where the organization such as the company related to the theme “politics” is.

  Here, it is assumed that the information “theme” (class “theme”) and the information “organization” (class “organization”) are selected from the information included in the input interface by the user's operation. .

  The user terminal 2 transmits these parameters to the graph search device 1 (S55).

  In the graph search device 1, the query issuing unit 14 transmits a query having a search syntax including the class “theme” and the class “organization” to the pattern database 13, and thereby the patterns P 1, P 2, and P 3 are transmitted from the pattern database 13. Search is performed (S59).

  Next, the query issuing unit 14 refers to the pattern cluster database 16, and since there is a pattern cluster including the patterns P1 and P2, excludes, for example, the pattern P2, which is one of the searched patterns P1 and P2 (S60). The two patterns P1 and P3 excluding the pattern P2 are given to the user interface 11.

  In step S60, a pattern cluster including a plurality of patterns is searched from the pattern database 13, and a plurality of patterns matching the plurality of patterns in the pattern cluster are selected from the plurality of patterns searched in step S59. Any one or more of the plurality of patterns are left and the rest are excluded.

Specifically, for example, the longest pattern may be left. This is because it is a simple and easy-to-understand pattern.
Alternatively, a pattern in which the search result is not empty and has many keywords may be left. This is because the search results are easy to obtain.
Alternatively, a pattern in which the search result is not empty and the number of keywords is small may be left. This is because an exact search result can be easily obtained.
In addition, the operator may determine such selection.

  The user interface 11 generates an input interface, transmits it to the user terminal 2 (S61), and displays it as shown in FIG. 11 (S63). Here, the patterns P1 and P3 are displayed, but in FIG. 11, the class and arc labels are not shown and are simply shown.

  In response to this, the user inputs the selection instruction of the pattern P1 and the keyword “politics” to the user terminal 2, and the user terminal 2 transmits these parameters to the graph search device 1 (S65).

  In the graph search device 1, the query issuing unit 14 includes the keyword “politics” for the node in which the class “theme” is defined in the selected pattern P <b> 1, and the search pattern (that is, the search pattern in FIG. 7). P11) is generated (S67).

  The query issuing unit 14 converts the search pattern P11 into a query, and transmits it to the graph database 12, thereby acquiring a subgraph that matches the search pattern (that is, the subgraph SG (P11) in FIG. 8) from the graph G. (S71).

  Next, the query issuing unit 14 takes out an instance (that is, “XX group”) in the node in which the class “organization” in the subgraph SG (P11) is defined, and gives it to the user interface 11 (S73).

  The user interface 11 generates an output interface, transmits it to the user terminal 2 (S75), and displays it as shown in FIG. 12 (S77).

  Therefore, the user does not need to select one from the three patterns P1, P2, and P3, but selects one from the two patterns P1 and P3, which improves convenience.

  Although one pattern is selected here, two or more patterns may be selected. Even in this case, for example, the number of selection candidates can be reduced, and two or more patterns may be selected from the few selection candidates, so that convenience is improved.

  The pattern classification unit 15 may use another method when classifying a plurality of patterns into one or more pattern clusters.

  An example will be described below. The points not described are the same as in the above embodiment.

  Here, for the sake of convenience, as shown in FIG. 13, five patterns P101 to P105 are used instead of the above-described pattern P1, and five keywords K11 to K15 are used instead of the above-described keyword K1. And

  In step S11 of FIG. 5, the query issuing unit 14 includes 5 keywords (number of patterns) × 5 (number of keywords) (25 in total) including the keywords K11 to K15 for the first nodes in the patterns P101 to P105. ) Search patterns P1011 to P1055) are generated (S11).

  The query issuing unit 14 converts the search patterns P1011 to P1055 into queries, and transmits them to the graph database 12, thereby acquiring a subgraph that matches the search pattern from the graph G (S15).

  The query issuing unit 14 gives the patterns P101 to P105, the subgraph, the threshold value “0.7”, and the class “organization” to the pattern classification unit 15 (S17).

  The pattern classification unit 15 classifies the pattern based on the subgraph, threshold value, and class (S19), and stores the classification result in the pattern cluster database 16 (S21).

  Here, step S19 will be described in detail. In step 19, the pattern classification unit 15 classifies the patterns P101 to P105 into one or more pattern clusters.

  FIG. 14 is a flowchart showing the operation in step S19.

  The pattern classification unit 15 first excludes the search pattern used in the search for which the corresponding subgraph was not obtained from the search patterns P1011 to P1055 (S191).

  As shown in FIG. 15, the pattern classification unit 15 excludes search patterns corresponding to cells marked with “x”.

  Next, the pattern classification unit 15 is obtained by including a keyword (for example, keyword K11) common to the two patterns (for example, patterns P101 and P102), as indicated by the frame Y1 in FIG. When subgraphs are obtained from the search patterns (for example, search patterns P1011 and P1021), the two patterns (patterns P101 and P102) are associated with each other (S193).

  In FIG. 15, the four patterns P101 to P104 are associated with each other as shown by being surrounded by the frame line Y2.

  Further, the pattern classification unit 15 defines one or more keywords included in the search pattern obtained from a plurality of patterns associated with each other and not excluded as a keyword cluster (S195). According to FIG. 15, the keywords K11 to K14 are defined as keyword clusters (referred to as keyword clusters C1).

  Further, the pattern classification unit 15 defines one or more keywords included in a search pattern obtained from a single pattern not associated with another pattern and not excluded as a keyword cluster (S195). According to FIG. 15, the keyword K15 is defined as a keyword cluster (referred to as keyword cluster C2).

  Next, the pattern classification unit 15 selects one keyword from each of the keyword clusters C1 and C2 (S197). At that time, the pattern classification unit 15 maximizes the number of search patterns that include the selected keyword and are not excluded. According to FIG. 15, keywords K11 and K14 are selected from the keyword clusters C1 and C2, respectively.

  Next, the pattern classification unit 15 includes, for the keyword cluster C1, one or more patterns P101, P102, P1021, P1021, which are used to generate the search patterns P1011, P1021, P1031 that include the selected keyword K11 and are not excluded. P103 is selected (S199).

  Further, the pattern classification unit 15 selects one or more patterns P105 used to generate the search pattern P1055 that includes the selected keyword K14 and is not excluded for the keyword cluster C2 (S199).

  For the keyword cluster C1, patterns P101, P102, and P103 are selected (S199), and the pattern P104 is not selected. If there are a plurality of patterns that are not selected in this way, The pattern may be classified into one or more pattern clusters in the same manner as in step S19.

  Next, the pattern classification unit 15 defines a plurality of patterns included in the selected patterns P101, P102, and P103 and similar to each other as a pattern cluster, and is included in the patterns P101, P102, and P103 and other patterns. Is defined as a pattern cluster, and the patterns P101, P102 and P103 are classified into one or more pattern clusters (S1911).

  Here, the pattern classification unit 15 determines whether the similarity determination value obtained for the two patterns is predetermined for any two patterns included in the pattern cluster including a plurality of patterns such as the patterns P101, P102, and P103. To satisfy.

  An example of the patterns P101 and P102 will be described.

The pattern classification unit 15 sets a set of instances in a node (referred to as a second node) in which a class “organization” is defined in one or more subgraphs that match the search pattern P1011 including the keyword K11 in the pattern P101, as R101, A set of instances in a node (second node) in which the class “organization” in one or more subgraphs matching the search pattern P1021 including the keyword K11 in the pattern P102 is defined as R102.
Similarity determination value T (P101, P102) = | R101∩R102 | ÷ | R101∪R102 |
However,
| R101∩R102 | is the number of instances in the product set of R101 and R102,
| R101∪R102 | is the number of instances in the union of R101 and R102,
And the similarity determination value is set to be equal to or greater than the threshold value “0.7”.

  Here, P101 and P102 in the similarity determination values T (P101 and P102) mean that they relate to the patterns P1 and P2.

  If the similarity determination value is less than the threshold value, the patterns P101 and P102 are not included in one pattern cluster.

  For example, as shown in FIG. 16, when R101 includes instances “A department” and “B department”, and R102 includes instances “A department”, “B department”, and “C department”, R101∩R102: Since “A department” and “B department” are included, | R101∩R102 | = 2.

  Since R101∪R102 includes “A department”, “B department”, and “C department”, | R101∪R102 | = 3.

  Therefore, the similarity determination value T (P101, P102) = 2 ÷ 3≈0.67. For example, if the threshold value is “0.6”, it is determined that the threshold value is equal to or greater than the threshold value. As a result, the patterns P101 and P102 are included in one pattern cluster. For example, if the threshold is “0.9”, it is determined that the threshold is less than the threshold. As a result, the patterns P101 and P102 are not included in one pattern cluster.

  In this way, the pattern classification unit 15 generates, for example, a pattern cluster including two patterns P101 and P102 and a pattern cluster including one pattern P103.

  In addition, the pattern classification unit 15 performs the same for the pattern P105. In this case, the pattern P105 is a pattern cluster.

  In the description so far, for the sake of convenience, an example in which three or four patterns are classified has been shown. However, in practice, for example, about 300 patterns are often classified.

  When the above-mentioned another method is used and the threshold value at that time is set to “0.8” and 330 patterns are classified, for example, a result as shown in FIG. 17 is obtained.

  First, 172 pattern clusters were obtained from 330 patterns. The number of pattern clusters including one pattern was 116. The number of pattern clusters including two patterns was 35. Hereinafter, the relationship between the number of patterns in the pattern cluster and the number of pattern clusters is as shown in the figure.

  FIG. 18 is a diagram showing nine patterns in the pattern cluster indicated by the arrow Y3 in FIG.

  Since these patterns are similar to each other, if the query issuing unit 14 searches for these 330 patterns in step S59, the query issuing unit 14 selects one of the nine patterns in the figure included in the 330 patterns. . Similarly, the query issuing unit 14 narrows 330 pieces to 172 pieces.

  When 172 are displayed in step S63, the user only has to select a desired pattern from 172 patterns, that is, it is not necessary to select from 330 patterns, so that convenience is improved.

  As described above, according to the present embodiment, the graph database 12 storing the graph G in which nodes having instances are connected by arcs, and N patterns for searching for subgraphs in the graph G N × M search patterns are generated for the first node in FIG. 5 by including different M keywords (S11), and subgraphs matching each search pattern are searched from the graph G (S15). Graph search Using the means (query issuing unit 14) and the instance in the second node in the searched subgraph, the degree of similarity is obtained for each of the combinations of the two patterns in the N patterns, and based on the degree of similarity And pattern classification means (pattern classification unit 15) for classifying N patterns (S19). By searching for a subgraph that matches the search pattern obtained by including a keyword in the first node of the pattern, and by obtaining an instance in the second node in the subgraph as a search result, the patterns can be classified. Can be substantially reduced. In addition, when the user selects a pattern, the user can easily select the pattern.

The pattern classification means classifies the N patterns into pattern clusters including one or more patterns based on the degree of similarity, and for each keyword, assigns the keyword to one of the two patterns. The set of instances in the second node in one or more subgraphs that match the included search pattern is A, and the first set in one or more subgraphs that match the search pattern that includes the keyword in the other of the two patterns. Let B be the set of instances in two nodes.
Similarity judgment value = | A∩B | ÷ | A∪B |
However,
A and B are not empty sets,
| A∩B | is the number of instances in the intersection of A and B,
| A∪B | is the number of instances in the union of A and B,
When calculating
The same pattern cluster includes the two patterns in which k or more (0 <k ≦ K: K is the number of the keywords) among the similarity determination values corresponding to the keywords are equal to or greater than a predetermined threshold value. Therefore, the number of patterns can be substantially reduced. In addition, when the user selects a pattern, the user can easily select the pattern.

The pattern classification means excludes the search pattern used in the search for which the corresponding subgraph could not be obtained from the N × M search patterns (S191), and includes a common keyword in each of the two patterns. If both subgraphs are obtained from the obtained search patterns, the two patterns are associated (S193), and one or more keywords included in the search patterns that are obtained from a plurality of associated patterns and are not excluded Is defined as a keyword cluster (S195), and one or more keywords that are obtained from a single pattern that is not associated with other patterns and are not excluded are defined as keyword clusters (S195). Categorize multiple keywords into one or more keyword clusters, 1 from each keyword cluster Search keywords including the selected keyword and not excluded (S197), and for each keyword cluster, a search including the selected keyword and not excluded. One or more patterns used to generate the pattern are selected (S199), a plurality of patterns included in the selected one or more patterns and similar to each other are defined as a pattern cluster, and the selected 1 A single pattern that is included in one or more patterns and is not similar to another pattern is defined as a pattern cluster, and the selected one or more patterns are classified into one or more pattern clusters (S1911). Any two patterns included in the pattern cluster are the keywords selected as one of the two patterns. A set of instances in the second node in one or more subgraphs that match the search pattern including the search field A, and one or more that matches the search pattern including the selected keyword in the other of the two patterns If the set of instances in the second node in the subgraph is B,
Similarity judgment value = | A∩B | ÷ | A∪B |
However,
A and B are not empty sets,
| A∩B | is the number of instances in the intersection of A and B,
| A∪B | is the number of instances in the union of A and B,
Is larger than a predetermined threshold value, so that the number of patterns can be substantially reduced. In addition, when the user selects a pattern, the user can easily select the pattern.

DESCRIPTION OF SYMBOLS 1 ... Graph search apparatus 2 ... User terminal 3 ... Display apparatus 11 ... User interface 12 ... Graph database 13 ... Pattern database 14 ... Query issuing part 15 ... Pattern classification part 16 ... Pattern cluster database

Claims (5)

A graph database storing a graph in which nodes having instances are connected by arcs;
N × M search patterns are generated by including M keywords different from each other for the first node in the N patterns for searching the subgraphs in the graph, and each search pattern is generated from the graph. A graph search means for searching for a subgraph that matches
Using the instance in the second node in the searched subgraph, obtain a degree of similarity for each of the combinations of the two patterns in the N patterns, and determine the N patterns based on the degree of similarity. A pattern classification device comprising: pattern classification means for classifying. The pattern classification means includes
Classifying the N patterns into pattern clusters including one or more patterns based on the degree of similarity;
For each keyword, A is a set of instances in the second node in one or more subgraphs that match the search pattern that includes the keyword in one of the two patterns, and the keyword is included in the other of the two patterns. Let B be a set of instances in the second node in one or more subgraphs that match the search pattern
Similarity judgment value = | A∩B | ÷ | A∪B |
However,
A and B are not empty sets,
| A∩B | is the number of instances in the intersection of A and B,
| A∪B | is the number of instances in the union of A and B,
When calculating
The same pattern cluster includes the two patterns in which k or more (0 <k ≦ K: K is the number of the keywords) among the similarity determination values corresponding to the keywords are equal to or greater than a predetermined threshold value. The pattern classification apparatus according to claim 1, wherein: A pattern classification method performed by a pattern classification apparatus including a graph database in which a graph in which nodes having instances are connected by arcs is stored,
The graph search means of the pattern classifier includes N different M keywords for the first node in the N patterns for searching the subgraphs in the graph, and sets N × M search patterns. Generate and search the graph for subgraphs that match each search pattern,
The pattern classification unit of the pattern classification device obtains a similarity degree for each of the combinations of two patterns in the N patterns using the instance in the second node in the searched subgraph, A pattern classification method, wherein the N patterns are classified based on a degree. The pattern classification means includes
Classifying the N patterns into pattern clusters including one or more patterns based on the degree of similarity;
For each keyword, A is a set of instances in the second node in one or more subgraphs that match the search pattern that includes the keyword in one of the two patterns, and the keyword is included in the other of the two patterns. Let B be a set of instances in the second node in one or more subgraphs that match the search pattern
Similarity judgment value = | A∩B | ÷ | A∪B |
However,
A and B are not empty sets,
| A∩B | is the number of instances in the intersection of A and B,
| A∪B | is the number of instances in the union of A and B,
When calculating
The same pattern cluster includes the two patterns in which k or more (0 <k ≦ K: K is the number of the keywords) among the similarity determination values corresponding to the keywords are equal to or greater than a predetermined threshold value. The pattern classification method according to claim 3, wherein:   A computer program for operating a computer as the pattern classification apparatus according to claim 1.

Images