JP5149768B2 - Taxonomy completion method, taxonomy completion program, and storage medium - Google Patents

Description

  The present invention relates to a taxonomy complement method, a know-how analysis method, a know-how analysis method, a taxonomy complement program, and a storage medium storing the program, and is particularly suitable for automatically complementing a taxonomy class that has been pre-constructed. It is.

  When developing systems such as systems that provide network services and financial services, many engineers belonging to multiple departments carry out development work from specification design to coding and testing of system configuration products that they are responsible for. It is often implemented in cooperation with engineers of other products. Users (know-who) who are familiar with system development work for each product and know-how of development work that are obtained through interaction on the social network of engineers involved in development work, This is important knowledge that is often required not only for system development but also for future system updates and operations.

  On the other hand, expressing and constructing know-how and know-how as formal knowledge is a troublesome task for an unfamiliar user, and is often postponed. In such a case, the know-how and know-how will be held as tacit knowledge by the engineers involved in the development, but the same engineer may not be involved in the same department or development for a long time, resulting in know-how and know-how. Is lost, and a new engineer needs to learn the knowledge from scratch again. Therefore, it is conceivable that a small taxonomy that can be constructed in a short time by an engineer is prepared first, and this taxonomy is complemented and used for analysis of know-how and know-how. A taxonomy is defined as a concept (class) hierarchy constructed by experts for each domain and comprehensively classifying information of the domain.

As a research related to such taxonomy complementation, as a conventional technique described in Non-Patent Document 1 below, using LSA (Latent Semantec Analysis), a key related to an instance (entity) in an ontology is obtained from a document set. The phrase is extracted, and a new class having the key phrase as an instance is generated in the ontology. Moreover, as a prior art described in Non-Patent Document 2 below, a user's interest is extracted according to a taxonomy by classifying a blog article with respect to a taxonomy of content constructed by a service provider for a blog. The accuracy rate of 95% or more has been confirmed. There are a relatively large number of content taxonomies, such as music domain content taxonomies.
Gliozzo, A. ~ M., Gangemi, A., Presutti, V., Cardillo, E., Dage, E., Salvati, A. and Troiani, G .: A Collaborative Semantec Web Layer to Enhance Legacy Systems, ISWC / ASWC, pp. 764--777 (2007) Nakatsuji, M., Miyoshi, Y. and Otsuka, Y.,: Innovation Detection Based on User-Interest Ontology of Blog Community., ISWC, pp 515--528 (2006)

  However, the prior art described in Non-Patent Document 1 does not attempt to extract related phrases in consideration of taxonomy, and does not attempt to extract know-how or know-how. In the prior art described in Non-Patent Document 2, there is no taxonomy related to tacit knowledge. Therefore, in tacit knowledge analysis, the trouble of constructing a taxonomy is a big problem. Further, the taxonomy can be supplemented by the conventional technique described in Non-Patent Document 1, and the user interest and user knowledge can be extracted along the taxonomy by the conventional technique described in Non-Patent Document 2. However, since the conventional technique described in Non-Patent Document 1 is not taxonomy completion using taxonomy, it is not possible to generate a completion class that inherits the properties of the parent class, and as a result, extracted user knowledge The accuracy of know-how obtained from articles classified into classes is low.

  The present invention has been developed by focusing on these problems. First, a taxonomy that can be created by an expert is first constructed, and the taxonomy is automatically complemented based on an e-mail. A taxonomy complement method, know-how analysis method, know-how analysis method, taxonomy complement program, and storage medium storing the program, which can create a complement class that can inherit the attributes of the class and enable high-precision know-how and know-how confirmation Is intended to provide.

In order to solve the above problems, the taxonomy completion method of the present invention is a taxonomy completion method executed by a computer based on software, and a child class that inherits the properties of a parent class is defined as a child class belonging to the parent class. In addition to setting the name attribute common to the parent class to the child class, a taxonomy is constructed, and a set of e-mails related to the same thread for a predetermined period is set as a mail set, and the description in the article of a specific mail set Categorize all mail sets into taxonomies, so that the particular mail set is classified into the specific class, and the specific phrase in the specific class Determine whether the name attribute of a specific class can be inherited, and the specific phrase is the name attribute of the specific class If that can be inherited, performs processing to complement the specific phrase name attributes and child class as a child class of the particular class, the processing of determining and said complementing includes a number of articles the particular class Ci Another class Cj having the same number of articles and belonging to the parent class of the specific class Ci is selected, the appearance frequency Ki of the specific phrase K in the mail set classified into the specific class Ci, and the other class Cj An appearance frequency score U (K) is calculated based on the appearance frequency Kj of a specific phrase K in the mail set classified as, and the appearance frequency score U (K) is calculated repeatedly a predetermined number of times while changing other classes Cj. If the average value of is equal to or greater than the threshold, to complement the related class name attribute the specific phrase K as a child class of the particular class Ci It is an feature.

The taxonomy completion method of the present invention is a taxonomy completion method executed by a computer based on software, and sets a child class that inherits the properties of the parent class as a child class belonging to the parent class, The taxonomy is constructed by assigning a name attribute common to the parent class, a set of e-mails related to the same thread for a predetermined period is set as a mail set, and a description in a specific mail set article is specified in a specific taxonomy If there is a class name attribute, classify the specific mail set into the specific class, classify all mail sets into taxonomies, and determine whether there is a relationship between the two specific classes If there is a relationship, a process for generating a child class common to the two classes is performed, and the two specific classes are the same. A different child class belongs to the class, the relevance of the mail set of each of the two child class is calculated based on the simpson coefficient, its relevance is, among all the comparison class belonging to the one parent class In the case of a predetermined upper rank, the fusion class composed of the common mail set of the two different child classes is complemented as a common child class belonging to the two different child classes.

A taxonomy complementing program of the present invention is for causing a computer to execute the taxonomy complementing method .

The storage medium of the present invention is a computer-readable medium storing the taxonomy complementing program .

  According to the present invention, it is determined whether a specific phrase of a specific class can inherit the name attribute of the specific class, and the specific phrase can inherit the name attribute of the specific class In addition, by complementing a child class with the specific phrase as a name attribute as a child class of the specific class, a taxonomy that can be created by an expert is first constructed and used to create a class attribute. It is possible to create a complementary class that can inherit from, which makes it possible to check know-how and know-how with high accuracy. In particular, when the average frequency of occurrence of a specific phrase of a specific class is equal to or greater than a threshold, a related class having the specific phrase as a name attribute is complemented as a child class of the specific class, or two different children When the relevance level of each mail set of a class is within a predetermined upper level among all comparison target classes belonging to the parent class, a fusion class composed of a common mail set of the two different child classes is designated as the two different child classes. By complementing as a common child class belonging to a class, it is possible to accurately create a completion class that inherits the attributes of the class. By browsing this exact completion class, high-precision know-how and know-how are confirmed. can do.

Hereinafter, an embodiment in which the taxonomy complementing method, the know-how analysis method, and the know-how analysis method of the present invention are applied to a social network for network system development will be described with reference to the drawings.
FIG. 1 is a block diagram of a schematic configuration of a system showing a social network of the present embodiment. In FIG. 1, terminals 2 to 4 and a server 5 are connected via a communication network 1. Also, terminals 7 to 9 are connected to the server 5, respectively. As the communication network 1, for example, a public communication network that performs IP communication can be used, and the Internet may be used. In addition, it may be a private communication network between companies or a public communication network, but it is like IP-VPN (Internet Protocol-Virtual Private Network) that can provide dedicated communication with high reliability and security. It may be a net. Further, the terminals 2 to 4 and 7 to 9 may be notebook personal computers or desktop personal computers, and a mobile phone terminal or a PDA (Personal Data Assistant) can be used. The server 5 can be installed on various service providers. The server 5 is also constructed by various computer systems, like the terminals 2 to 4.

  FIG. 2 is a schematic configuration diagram of the hardware of the computer system in each of the terminals 2 to 4 and 7 to 9 or the server 5, and the taxonomy complementing method and know-how analysis of the present invention based on the program stored in the storage medium. It is a system for executing methods and know-how analysis methods. The hardware of this computer system includes a central processing unit (CPU) 60 that is a central processing unit that performs various types of control and arithmetic processing, a random access memory (RAM) 62 that constitutes a main storage, and a read-out. A dedicated storage device ROM (Read Only Memory) 64 is connected by various internal / external buses 68 such as PCI (Peripheral Component Interconnect) bus, ISA (Industrial Standard Architecture) bus, etc. Via an interface (I / F) 66, an external storage device (Secondary Storage) 70 such as an HDD, an output device 72 such as a printing unit, a CRT, or an LCD monitor, an input device 74 such as an operation panel, a mouse, a keyboard, or a scanner. And a network cable L or the like for connecting to the communication network 1 is connected.

  When the power is turned on, a system program such as BIOS stored in the ROM 64 or the like loads various dedicated computer programs stored in advance in the ROM 64 into the RAM 62, and the CPU 60 follows the instructions described in the program loaded in the RAM 62. However, by performing predetermined control and arithmetic processing by making full use of various resources, each function to be described later is realized on software.

  Next, a schematic configuration of a taxonomy complementing method, a know-how analysis method, and a know-how analysis method that are actually constructed in any computer system by arithmetic processing according to a program will be described with reference to the block diagram of FIG. . In this taxonomy complement method, know-how analysis method, and know-how analysis method, an expert of a network system development engineer constructs a small taxonomy of an article that will be described in an e-mail. The article classification section built with software acquires a part of this small taxonomy and the relationship between the topics on the mailing list (ML) used for information exchange between engineers and the topics on the acquired taxonomy. While calculating the sex and complementing the taxonomy with the related topics, the mailing list articles are classified into taxonomy classes, and a user ID is given. From the user's articles and user IDs classified into classes on the taxonomy, the knowledge extracting unit 12 extracts the user's knowledge along the taxonomy, and from the user's knowledge, the knowledge similarity calculating unit 13 , And the inter-user relationship analysis unit 14 analyzes the relationship between the users, which is meaningful by the taxonomy.

Next, the flowchart of FIG. 4 will be described as a first embodiment of the arithmetic processing for achieving the taxonomy complement method, know-how analysis method, and know-how analysis method of FIG.
First, in step S1010, the calculation process performs the calculation process of FIG. 5 described later, thereby inputting articles on the mailing list and a part of the built taxonomy, classifying the articles into classes on the taxonomy, The classified user article and user ID are output.

Next, in step S1020, the user knowledge and user ID classified into the taxonomy class are used, and user knowledge is extracted along the taxonomy as described below.
Next, in step S1030, the user knowledge extracted along the taxonomy is used, and the similarity of the user knowledge is calculated in consideration of the taxonomy as described later.
Next, in step S1040, the user's knowledge along the taxonomy and the similarity between the users are used to refer to the taxonomy as described later, and the relationship between the users is defined, and the user defined by the taxonomy Output the relationship.

Next, the calculation process of FIG. 5 performed in step S1010 will be described. In this calculation process, first, in step S2010, as will be described later, articles on the mailing list and a part of the built taxonomy are input, the articles are classified into classes on the taxonomy, and the classified user articles are classified. And the user ID.
Next, in step S2020, by performing the arithmetic processing of FIG. 6 and FIG. 7 described later, a child class that inherits the properties of the class on the taxonomy is automatically generated, or a class on another taxonomy is Analyzing the relationship, generating a fusion class that combines the properties of both, and automatically generating it as a child class of the class on the taxonomy, to complement the taxonomy.

Next, in step S2030, the articles on the mailing list and a part of the built taxonomy are input as described later, including the complemented taxonomy classes, and the articles are classified into the classes on the taxonomy. Then, the classified user article and user ID are output.
Next, the calculation process of FIG. 6 performed in step S2020 will be described. In S <b> 3010, this calculation process automatically generates a child class that inherits the properties of the class on the taxonomy by performing the calculation process of FIG. 8 described later.

Next, the calculation process of FIG. 7 performed in step S2020 will be described. In S3020, this arithmetic processing is performed in FIG. 10 to be described later, thereby analyzing the relationship between classes on different taxonomies, generating a fusion class having both properties, and classifying the class on each taxonomy. Automatically generated as a child class.
Next, the calculation process of FIG. 8 performed in step S3010 will be described. In this calculation process, first, in step S4010, the user sets the number of hops (the number of Hops in the figure). The number of hops means the number of classes existing on the taxonomy between the filter target class (class for which a complementary class is to be set) and the adjacent class. The fact that a specific class Ci is close to another class Cj as a class hierarchy means that it is semantically close to the nature of the taxonomy, and the number of hops is used as a parameter for controlling this semantic proximity.

  In step S4020, an important phrase set K is extracted from the mailing list within the set number of hops. For this extraction, termEX (http://gensen.dl.itc.u-tokyo.ac) was developed in a study to automatically extract phrases that could become technical terms from corpus (electronic text) prepared for each specialized field. .jp / win.html). termEX extracts technical terms using the frequency of adjacent nouns when a single noun forms a code noun. Thereby, as will be described later, the engineer of the development project extracts a phrase set K that can be an important topic regarding the development project from the extracted phrases.

Next, in step S4030, one important phrase Kl is randomly selected from the important phrase set K.
Next, in step S4040, a class on the taxonomy is held in the list C holding the parent-child relationship, and a certain (specific) class Ci is randomly selected from the list C. Each class has one parent class. The root class clearly indicates that fact.

  Next, in step S4050, it is checked whether the important phrase Kl can inherit the attribute (property) of the specific class Ci by performing the arithmetic processing of FIG. 9 described later, and the important phrase Kl is the attribute of the specific class Ci. Can be inherited, a class Cr having the important phrase Kl as a class name (also an attribute) is created, and the class Cr is set as a child class (also referred to as a related class) of a specific class Ci.

Next, in step S4060, it is determined whether or not all the phrases Kl in the phrase set K are checked. If all the phrases Kl in the phrase set K are checked, the process proceeds to step S4070. Then, the process proceeds to step S4030.
In step S4070, it is determined whether or not all the classes Ci in the list C have been checked. If all the classes Ci in the list C have been checked, the process returns. If not, the process proceeds to step S4040. The parent class is excluded. Also, if the granularity of the concept (class) is thought to be too coarse, exclude the child classes of those hierarchies so as not to add the complementary classes to the shallow classes such as the first hierarchy and the second hierarchy. May be.

Next, the calculation process of FIG. 9 performed in step S4050 will be described. In this calculation process, first, in step S5010, a class set G having the same number of articles as that of the specific class Ci is listed with a class exceeding the number of hops set by the user as an adjacent class of the specific class Ci. To do.
Next, in step S5020, one class Cj different from the specific class Ci in the class set G is selected. The reason why the number of articles is the same is that the class with more articles is expected to have a coarser semantic growth, and the appearance frequency of the important phrase Kl tends to be higher than the specific class Ci. This is because it is not appropriate.

  Next, in step S5030, the occurrence frequency | KlεS (Ci) | of the important phrase Kl in the mail set S (Ci) classified into the specific class Ci is calculated, and the mail set classified into a different class Cj. The appearance frequency | KlεS (Cj) | of the important phrase Kl in S (Cj) is calculated, and the appearance frequency score U (Kl) is calculated according to the following equation 1 using the appearance frequencies.

Next, in step S5040, while changing different classes Cj in the class set G, steps S5020 and S5030 are repeated for a predetermined heuristic (experimental) number of trials X times, and the appearance frequency score U (Kl) is set. calculate.
Next, in step S5050, an average value of the appearance frequency score U (Kl) for the number of trials X times is calculated as U (Kl) a.

Next, in step S5060, it is determined whether or not the calculated appearance frequency score average value U (Kl) a is greater than or equal to the heuristic threshold value α, and the appearance frequency score average value U (Kl) a is greater than or equal to the threshold value α. If there is, the process proceeds to step S5070, and if not, the process proceeds to step S5080.
In step S5070, it is assumed that the important phrase Kl can inherit the attribute (property) of the specific class Ci.

In step S5080, “False” is assumed that the important phrase Kl cannot inherit the attribute (property) of the specific class Ci.
As the different classes Cj are closer to a specific class Ci as a class hierarchy, that is, closer to the semantics, characteristic phrases are extracted from each class, but the number of phrases is reduced. Therefore, the number of hops becomes a parameter for adjusting the distance on the taxonomy of the class Cj different from the specific class Ci. As the number of trials X increases, phrases characteristic of each class remain in the mail set, but the number of remaining phrases decreases.

  Next, the calculation process of FIG. 10 performed in step S3020 will be described. First, in step S6010, two different classes on taxonomies are held in two lists C1 and C2 each having a parent-child relationship in step S6010, and the list C1, that is, from a certain (specific) taxonomy (specific) The class Ci is selected at random. Each class has one parent class. The root class clearly indicates that.

Next, in step S6020, a class C2, that is, a different class (also a specific class) Cj from different taxonomies is randomly selected.
Next, in step S6030, the simpson coefficient (Rasmussen, E. m.,: Clustering Algorithms, Information) is applied to the mail sets S (Ci) and S (Cj) classified into the specific class Ci and the different class Cj. Retieval: Data Structures & Algorithms, pp. 419--442 (1992)) is calculated according to the following two equations. In the simpson coefficient, there is a problem that the degree of relevance R is particularly large when the number of articles in the mail sets S (Ci) and S (Cj) is extremely small. By setting the degree R to 0, this problem is avoided.

Next, in step S6040, if the relevance R is within the pre-set upper Y among all the comparison target classes C, a fusion class Cij composed of a common mail set | S (Ci) ∩S (Cj) | And a child class common to a specific class Ci and a different class Cj.
Next, the flowchart of FIG. 11 will be described as a second embodiment of the arithmetic processing that achieves the taxonomy complement method, know-how analysis method, and know-how analysis method of FIG.
In this calculation process, first, in step S7010, as in the case of step S1010, by performing the calculation process of FIG. 12 described later, articles on the mailing list and a part of the built taxonomy are input, and these articles are converted into taxonomies. Classify into the upper class, and output the classified user article and user ID.

Next, in step S7020, a user knowledge is extracted along the taxonomy using the user article and user ID classified into the taxonomic class by performing the arithmetic processing of FIG. 13 described later.
Next, in step S7030, similar to step S1030, the user knowledge extracted along the taxonomy is used, and the similarity of the user knowledge is calculated while considering the taxonomy.

Next, in step S7040, as in step S1040, the user's knowledge along the taxonomy and the similarity between the users are used, the taxonomy is referred to, meaning the relationship between the users, and the user implied by the taxonomy. Output the relationship between.
Next, the calculation process of FIG. 12 performed in step S7010 will be described. In this calculation process, first, in step S8010, as in step S2010, articles on the mailing list and a part of the built taxonomy are input, the articles are classified into classes on the taxonomy, and the classified users are classified. The article and user ID are output.

Next, in step S8020, in the same manner as in step S2020, a child class inheriting the property is automatically generated for the class on the taxonomy, or the relationship with the class on another taxonomy is analyzed, and the properties of both are analyzed. By creating a fusion class that has both and automatically generating it as a child class of the class on the taxonomy, the taxonomy is complemented.
Next, in step S8030, as in step S2030, the articles on the mailing list are classified into classes on the taxonomy, including the complemented taxonomy classes, and the classified user articles and classification results are output. . In particular, by classifying the mail set into the complemented taxonomy, knowledge for each user can be extracted along the supplementary taxonomy. For the related class Cr and the fusion class Cij of the specific class Ci, the mail set having the description of the name attribute of the specific class Ci and the name attribute of the related class Cr or the fusion class Cij is classified.

  Next, the calculation process of FIG. 13 performed in step S7020 will be described. In this arithmetic processing, first, in step S9010, knowledge analysis is performed on the complemented class, and the user's knowledge on the complemented class is detected. At that time, the knowledge score N indicating the depth of the knowledge of the user for each class constituting the taxonomy is introduced to analyze the user having knowledge for each class. The knowledge score N is defined below.

1) The user's knowledge score N per article is 1.
2) If the number of classes described in a certain (specific) article mi is N (mi), the knowledge score N of the user in that article is 1 / N (mi).
3) If the total set of articles of a user is M, the knowledge score N (Ci) for each class (specific class) Ci of the user is given by the following three formulas.
4) If a set of child classes, related classes, and fusion classes of a specific class Ci is C, a user's knowledge score N (Ci + C) for the specific class Ci is given by the following four equations.

If a complementary taxonomy is used, as will be described later, user knowledge for a specific class Ci can be analyzed based on knowledge about related classes and fusion classes in addition to manually created taxonomies.
Next, in step S9020, the user knowledge is extracted along the taxonomy using the user articles and user IDs classified in the taxonomy class, as in step S1020.

Next, a flowchart of FIG. 14 will be described as a third embodiment of the arithmetic processing for achieving the taxonomy complement method, know-how analysis method, and know-how analysis method of FIG.
This calculation process is a modification of the subroutine performed in step S7010 of FIG. 11. First, in step S10010, as in step S2010, an article on the mailing list and a part of the built taxonomy are input. The articles are classified into classes on the taxonomy, and the classified user articles and user IDs are output.

Next, in step S10020, in the same manner as in step S2020, a child class that inherits the property is automatically generated for the class on the taxonomy, or the relationship between the class on another taxonomy is analyzed, and the properties of both are analyzed. By creating a fusion class that has both and automatically generating it as a child class of the class on the taxonomy, the taxonomy is complemented.
Next, in step S10030, by performing the arithmetic processing of FIG. 15 described later, the articles on the mailing list are classified into the classes on the taxonomy including the complemented taxonomy classes, and the classified user articles and Output the classification result.

Next, the calculation process of FIG. 15 performed in step S10030 will be described. In this calculation process, first, in step S11010, as in step S8030, the articles on the mailing list are classified into the classes on the taxonomy including the complemented taxonomy classes, and the classification results are output.
Next, in step S11020, the articles classified into the complementary class and the classification results are presented to the user. As will be described later, the user can acquire know-how related to each class by browsing a mail set classified into classes on the complementary taxonomy. Moreover, the information search by browsing the class on the taxonomy can grasp the system development topic according to expert knowledge even by a user who is not very detailed about the object.

  Next, in step S11030, articles and classification results that are particularly classified into the fusion class Cij are presented to the user. As will be described later, since the development class topics for products are gathered in the fusion class Cij, know-how necessary for system development can be found. When a user browses a mail set in a class, the mail set is ranked and presented in the order of the mail set in which the name attribute of the class appears in particular.

  Next, the effect | action by the arithmetic processing of each said embodiment is demonstrated. First, a procedure for manually creating a taxonomy shown in FIG. 16 will be described. First, the designer determines the development project that is the subject of knowledge management. Regarding the granularity of the target project, it is considered appropriate that the information exchanged on the mailing list etc. needs to be discussed in common among engineers. Next, a taxonomy for the product developed in the target project and a taxonomy for the task are constructed as shown in FIG. Here, an instance (entity) on a taxonomy is a topic of a development target product or a development task, and a class is defined by a taxonomy for the topic. In order to facilitate the construction of the taxonomy, the relationship between the classes on the taxonomy is limited to the fact that the lower class inherits the properties of the upper class. For example, the class “router failure management device” in FIG. 16 inherits the attribute “network management” in the upper class “network (NW in the figure) management device”. A class is given multiple name attributes. For example, the class “network management device” has two attributes “network management device” and “network operation device”.

  For classification of e-mails into taxonomies such as steps S1020, S1030, S2010, and S2030, a technique for classifying blog articles into taxonomies described in Japanese Patent Laid-Open No. 2008-27142 previously proposed by the present inventors Is applicable. However, unlike blogs, e-mail is characterized by discussing various topics in chronological order. In the present embodiment, it is considered that topics related to system development often change in units of about one week, and the unit of the e-mail set to be classified is set as a weekly e-mail set of the same thread, and this unit is set as a mail set. Name it.

First, as preprocessing for an email article, the citation is removed (such as a line beginning with>) or the signature (a line including an address or email address) is removed. Then sort the mail sets in the mailing list into taxonomies. The mail set classification procedure will be described using the example shown in FIG.
1) Classify all mail sets as taxonomies. As a classification method, if a description in an article of a certain (specific) mail set has a name attribute of a certain (specific) class Ci on the taxonomy, the article is classified into the class Ci. Note that articles may be classified into a plurality of classes.

  2) Next, classification error filtering is performed. For example, the name attribute of the class “getter” in FIG. 17 is often used not only as a product that acquires network management information from a router, but also as a method that returns a field value in the java language. , Classification errors will increase. Therefore, the following filtering algorithm is introduced by utilizing the property of taxonomy that the properties between classes close to each other in the class hierarchy are close.

  That is, when the name attribute of a (specific) class Ci in a certain (specific) e-mail m is described, the description of the adjacent class Cj of the specific class Ci in the mail set S to which the e-mail m belongs. Check if there is any. When there is a description, the specific electronic mail m and the mail set S are classified as topics related to the specific class Ci, and when there is no description, it is an error. For example, if the description for “getter” in FIG. 17 exists in a certain (specific) e-mail m and the description “network management” is included in the mail set S, for example, the mail set S has the class “network”. It is classified as an email belonging to the subclass “getter” of “Management”. Here, the range of the adjacent class can be changed using the number of classes existing on the taxonomy between the filter target class and the adjacent class, that is, the number of hops. For example, when the hop count is 1, the adjacent classes are a parent class and a child class, and when “getter” is a filter target class, it is a “router failure management device”. When the number of hops is 2, the grandfather class and the sibling class are also included. In FIG. 17, “setter” and “network management apparatus” are also included.

3) From such classification results, system development knowledge for each user can be extracted by sorting the classification results based on the user IDs of the mail articles in the mail set.
The relationship analysis between users, such as step S1040, can be performed as follows. FIG. 18 is a conceptual diagram of an interest system graph reconstructed by giving semantic information to user relationships. Here, interest ontologies between users are compared, and a class or instance having a high degree of interest matching is used as interest matching information, while a class or instance that does not match interest is given to the edge as deviation information. Semantic information is also given to the relationship between user categories. The interest ontology of the user set belonging to the user category is compared between the categories, and a class or instance having a large number shared by the user category is used as interest matching information, while a class or instance that does not match interest is given to the edge as misalignment information. . In the above, by adding a class as semantic information as well, if the instances of each user category become excessive, the search user can confirm the class and easily find a search direction that satisfies his interests. it can.

  In FIG. 18, the data of the user graph includes a plurality of user categories including a user category 10, a user category 20,. Each of the user category 10, user category 20,..., User category N includes a plurality of lower user categories. In FIG. 18, for the convenience of explanation, only the user category 10 is illustrated as subgraphs of the user category 11 and the user category 12 that are user categories below the user category 10. In FIG. 18, the relationship of connections between subgraphs and users is indicated by solid lines, and the semantic relationship between subgraphs and users is indicated by broken lines.

The user category 11 of the subgraph has user 1 and user 2, and the relationship between user 1 and user 2 is accompanied by a dashed line connecting them:
Match: Instance 1 of class A
Misalignment: Instance 2 of class A (User 1)
Misalignment: Instance 3 of class A (User 2)
It is as written.

  “Match: instance 1 of class A” means that instance 1 belonging to class A matches between user 1 and user 2. “Displacement: instance 2 of class A (user 1)” means that user 1 has instance 2 belonging to class A, but user 2 does not. “Displacement: instance 3 of class A (user 2)” means that user 2 has instance 3 belonging to class A, but user 1 does not.

Also, the user category 12 of the subgraph has users 3 and 4 and the relationship between these users 3 and 4 is accompanied by a broken line connecting them:
Match: Instance 2 of class A
Misalignment: Instance 4 of class B (User 3)
Misalignment: Instance 5 of class B (user 4)
It is as written.

  “Match: instance 2 of class A” means that instance 2 belonging to class A matches between user 3 and user 4. “Displacement: instance 4 of class B (user 3)” means that the user 3 has the instance 4 belonging to the class B, but the user 4 does not. “Displacement: instance 5 of class B (user 4)” means that the user 4 has the instance 5 belonging to the class B, but the user 3 does not.

Furthermore, the relationship between the user category 11 and the user category 12, which are subgraphs, is accompanied by a dashed line connecting them:
Match: Instance 2 of class A
Misalignment: Instance 1 of class A (category 11)
Misalignment: Instance 3 of class A (category 11)
Misalignment: Instance 4 of class B (category 12)
Misalignment: Instance 5 of class B (category 12)
It is as written.

  “Match: instance 2 of class A” means that instance 2 belonging to class A matches between category 11 and category 12. “Displacement: instance 1 of class A (category 11)” means that category 11 has instance 1 belonging to class A, but category 12 does not. “Displacement: instance 3 of class A (category 11)” means that the category 11 has the instance 3 belonging to the class A, but the category 12 does not. “Displacement: instance 4 of class B (category 12)” means that category 12 has instance 4 belonging to class B, but category 11 does not. “Displacement: instance 5 of class B (category 12)” means that the category 12 has the instance 5 belonging to the class B, but does not have the category 11.

  In the above description, for example, that instance 1 belongs to class A means that it is based on the taxonomy of the instance, that is, an expert defines it with expert knowledge. In the interest system graph as shown in FIG. 18, a simple user graph is replaced with a hierarchical community (category) distribution to give a relationship between the communities. Therefore, conventionally, since this relationship does not exist, users who are separated from each other have to follow all the users interposed between them in order to refer to each other, but by giving the relationship as described above, If it matches the user's own interest, a desired community (category) can be searched at once.

In the interest system graph as shown in FIG. 18, it is assumed that the user 1 belonging to the user category 11 is a search user, and the instance 4 and the instance 5 under the class B are interested. The user 1 sees the relationship between the users, and since the deviation information of the user 2 does not include the information of the instance 4 and the instance 5, when the deviation information of the adjacent user category 12 is seen, it is understood that the user 1 has the class B information. The user 3 and the user 4 can be searched to find the instance 4 and the instance 5.
In the above example, the scale of the user graph is not so large, so the effect does not seem to be remarkable, but when the scale of the graph is large, the effect of having a search means for tracing semantic information is great. .

  In such relationship analysis between users, it is possible to form an interest system graph by accumulating and managing graph-based data in accordance with user preferences from user preference information. In addition, as described above, in order to rely on a taxonomy defined by specialists based on expertise, a management system is mapped according to the intended intention. Therefore, it becomes easy for the user to find a desired content by an intuitive search operation. In particular, in the conventional recommendation (recommendation), there were many so-called push-type recommendations with a sudden feeling. By clearly indicating the relationship between users, the user can actively track the relationship, The so-called toughness of such push type recommendations can be eliminated. For example, the user obtains a sense of satisfaction by referring to the meaning of “reason for matching”, or is interested in new information while referring to the meaning of “reason for deviation”.

  Next, the taxonomy auto-complement of this embodiment and its effect | action are demonstrated. To complete the taxonomy, use termEX as described above to extract development phrases from the mailing list. Then, a phrase set K that can be an important topic regarding the development project is extracted from the extracted phrases by the engineer of the development project. Next, a (specific) phrase Kl is selected from the phrase set K, and the relationship between the phrase Kl and the mail set S (Ci) classified into a (specific) class Ci on the taxonomy is calculated. The relevance is obtained by the appearance frequency score U (Kl) of the phrase Kl according to the above-mentioned formula 1. Then, as described above, while changing different classes Cj, the application frequency score U (Kl) is calculated heuristically X times, and the appearance frequency score average value U (Kl) a is greater than or equal to the heuristic threshold α. In some cases, the phrase Kl is extracted as a related topic characteristic only for the specific class Ci, and is given to the related class Cr as a child class of the specific class Ci.

The advantage of using taxonomy is that it can detect not only simple co-occurrence but also co-occurrence based on the distance between classes on the taxonomy. For example, in FIG. 19, if “getter” is a specific class Ci, when the number of hops is set to 4 or more, general phrases such as “log file” are removed, and a relatively characteristic “getter” such as “SNMP”. The phrase remains. If the number of hops is reduced to about 2, the phrase “SNMP” disappears and only the phrase “error analysis” remains.
On the other hand, for example, topics that belong to both the class on the product taxonomy and the class on the task taxonomy are topics related to product development tasks and are considered to include know-how. Extract the fusion class Cij.

  Then, if the mail set is classified again into these complementary taxonomies, knowledge for each user can be extracted along the complementary taxonomies. At that time, as described above, by introducing a knowledge score N (Ci + C) indicating the depth of the knowledge of the user for each class constituting the taxonomy, the user having knowledge for each class is analyzed. If a complementary taxonomy is used, knowledge for a certain (specific) class Ci can be analyzed based on knowledge about related classes and fusion classes in addition to manually created taxonomies. As a result, for a certain (specific) class Ci on the taxonomy, if the developer frequently describes the name of the related class Cr of the class Ci, even if the developer does not describe the class name very often, It becomes possible to extract as a user who has a lot of knowledge about the (specific) class Ci (now analysis). Considering that the person in charge of development tends to write on topics that are finer than the class names on the taxonomy, know-how analysis using complementary taxonomy, especially related classes, is effective.

  Further, as described above, the user can acquire know-how regarding each class by browsing the mail set classified into the class on the complementary taxonomy. Information search by browsing the classes on the taxonomy allows users who are not familiar with the subject to understand system development topics according to their expertise. In addition, since the topics of development tasks for products are aggregated in the fusion class Cij, know-how necessary for system development can be found.

  In addition, in the above-described relationship analysis between similar users, an improvement in accuracy can be expected by using a complementary taxonomy. In other words, as described above, the person in charge of development tends to write on topics that are finer than the topics included in the taxonomy class. In addition, the similarity can be determined also in the related class Cr and the fusion class Cij.

  Also, an important advantage of using taxonomy is that the relationship between users can be expressed in more detail as shown in FIG. Conventionally, in relations between users, it was only possible to give the points that the knowledge is similar or different between similar users by using keywords. "postgres", "oracle", the difference is "zone", "address design", all must be listed with keywords. However, it is painful for the user to check all the keywords, and the keywords are often not understood. On the other hand, by using the taxonomy, as shown in FIG. 20, the point that the knowledge is similar between the user A and the user B is the (path setting & address design) class under the (path setting), The difference is that (oracle) under (DB) class and (zone & design) under (DNS server) can grasp the relationship between users based on expertise.

  Next, verification of this embodiment was performed using a mailing list used in a development project of a system for managing a network service and a taxonomy created by an engineer. The mailing list used for the verification was 23833 e-mails, 13342 mail sets, and was used for 2 years and 2 months when 111 people posted. Three engineers created a small taxonomy consisting of only 49 classes for products and 57 classes for tasks in about 6 hours. Furthermore, as a phrase set, preprocessing was performed manually for about 3 hours based on a simple rule such as deleting only one character from a phrase cut out from the mailing list with termEX. The number of phrases that can finally become the topic of development knowledge extracted is 3952 phrases, and a complementary taxonomy consisting of about 1575 classes was constructed by a taxonomy completion algorithm.

  The verification confirmed three points: 1) validity of related classes, 2) effectiveness of know-how extraction, and 3) effectiveness of know-how extraction. As a verification method, among the results of the algorithm in each verification, the engineer of the development project confirmed the highest result of the score, and confirmed the percentage of correct answers (matching rate). By changing the number of higher confirmations of the result and confirming the result, it is possible to determine how much the confirmation rate can be obtained by the confirmation number. In addition, in order to suppress the variation in the results of the engineers to be verified, the verification was performed while performing a cross check by three engineers deeply involved in the development project. Prior to the verification, the number of trials X was preliminarily compared with 3, 5, and 10, and the number of related classes was too small for 3 and too large for 10. Also, the threshold value α is changed to 0.8, 0.85, 0.9, and 0.95. If the threshold value α is too low, the accuracy of the result is lowered, but if it is too high, topics related to the class are extracted too much. Therefore, the verification was performed as 0.85. Further, the upper Y was verified as 5. As a comparison object, it compared with the method which does not complement a taxonomy, and the method which complements a taxonomy using TF / iDF.

  First, the validity of the complemented related class Cr was verified for a certain (specific) class Ci. The correct answer is that the topic related to class Ci is (a) a topic that inherits the properties of class Ci and is to be confirmed as know-how, and (b) a topic that does not inherit the properties of class Ci but is to be confirmed as know-how. Two types of standards were prepared. As verification targets, select 10 classes, which are about 10% of all classes on the taxonomy, change the results to the top 5 topics and 10 topics for each, output and verify, and the verification results are shown in the table below. Summarized in 1. According to this, in the embodiment of the present invention, the correct answers belonging to (a) can be extracted more than twice as much as TF / iDF. The correct answers belonging to (b) can be obtained with the same accuracy in both the embodiment of the present invention and the TF / iDF. In the subsequent verification, the top 10 topics were used and the taxonomy was supplemented.

  Next, the effect on the know-how analysis was verified. As a standard for the correct answer, the case where a user with a high knowledge score is a user who has a deep knowledge with respect to a class on a manually created taxonomy is regarded as a correct answer. As a comparison object, a comparison was made between a case where no taxonomy was complemented and a case where the taxonomy was complemented using TF / iDF. In addition, as the verification target, 10 classes corresponding to 20% of all classes of the product taxonomy are selected, and for each of them, the users with particularly large knowledge scores are changed to the top 5, 10, 15, 20 and output, The verification results are summarized.

FIG. 21 shows a result composed of an average value of all results. According to this, compared with the case where the original taxonomy is used (when complementation is not performed), in the embodiment of the present invention, about 3% to 10%, and 3% to 8% better than TF / iDF. You can get a rate.
Furthermore, the effectiveness when the knowledge similarity between users was calculated was verified. As a verification target, about 10 people who are about 10% of the members of the mailing list are targeted, and other engineers similar to each user are changed to the top 5, 10, 15, 20 based on the similarity, and output, The verification results are summarized. Here, when selecting 10 users as verification targets, a user who is closely related to the entire project and a user who was only involved in a certain (specific) product development were selected and verified.

  In FIG. 22, the result which consists of the average value of all the results is shown. According to this, it can be seen that the matching rate of 10% to 20% or more is obtained in the embodiment of the present invention, compared with the case where the taxonomy is originally used (when no complement is performed). In addition, a higher precision can be obtained compared to TF / iDF. From these results, it was shown that supplementing the taxonomy with the present invention leads to appropriate know-how detection.

Furthermore, it was verified whether the related class, the fusion class, and the mail set classified into each of the five classes corresponding to 10% of the classes on the product taxonomy actually contain know-how about the class. In the verification, among the mail threads classified as threads, the top three and six mail sets were changed and output, and the correct answer ratio was confirmed.
Table 2 below shows the results consisting of the average of all results. This shows that the related classes include more know-how than the original taxonomy classes. This is reasonable because deeper classes have more detailed know-how. However, like the fusion class 1 in the table, the percentage of correct answers was low in the fusion class. As a result of checking the results, it was found that the accuracy of the fusion class of the “problem” and “test” classes on the task was poor. This is because the phrases “problem” and “test” did not completely eliminate the error due to the polysemy even when the filtering algorithm was used. Therefore, when these classes and fusion classes were excluded and verified, it was confirmed that the accuracy was improved as shown in fusion class 2 in Table 2. The problem with polysemy is that the co-occurrence of the class name attribute between the task and the product should be confirmed in the same e-mail, not in the mail set.

It is a block diagram which shows an example of the saw shirt network to which this invention is applied. It is a block diagram of the computer system used for the terminal or server of FIG. It is a schematic configuration of a taxonomy complement method, a know-how analysis method, and a know-how analysis method constructed in a computer system by arithmetic processing according to a program. It is a flowchart which shows 1st Embodiment of the arithmetic processing of FIG. It is a flowchart which shows the subroutine of the arithmetic processing of FIG. It is a flowchart which shows the subroutine of the arithmetic processing of FIG. It is a flowchart which shows the subroutine of the arithmetic processing of FIG. It is a flowchart which shows the subroutine of the arithmetic processing of FIG. It is a flowchart which shows the subroutine of the arithmetic processing of FIG. It is a flowchart which shows the subroutine of the arithmetic processing of FIG. It is a flowchart which shows 2nd Embodiment of the arithmetic processing of FIG. It is a flowchart which shows the subroutine of the arithmetic processing of FIG. It is a flowchart which shows the subroutine of the arithmetic processing of FIG. FIG. 12 is a flowchart illustrating a subroutine of the arithmetic processing of FIG. 11 as a third embodiment of the arithmetic processing of FIG. 3. It is a flowchart which shows the subroutine of the arithmetic processing of FIG. It is explanatory drawing of the small taxonomy constructed | assembled initially. It is explanatory drawing of the mail set classification | category to each class of a taxonomy. It is a conceptual diagram of an interest system graph reconstructed by giving semantic information to user relationships. It is explanatory drawing of the adjacent class setting by the number of hops. It is explanatory drawing of the relationship analysis between users using a taxonomy. It is explanatory drawing of the effect by this invention. It is explanatory drawing of the effect by this invention.

Explanation of symbols

  1 is a communication network, 2 to 4 and 7 to 9 are terminals, 5 is a server, 11 is an article classification unit, 12 is a knowledge extraction unit, 13 is a knowledge similarity calculation unit, 14 is an inter-user relationship analysis unit, and 60 is a CPU. , 62 is a RAM, 64 is a ROM, 66 is an interface, 68 is a bus, 70 is a storage device, 72 is an output device, and 74 is an input device.

Claims (4)

A taxonomy completion method executed by a computer based on software, in which a child class that inherits the properties of the parent class is set as a child class belonging to the parent class, and the child class has a name attribute common to the parent class. If a taxonomy is constructed by assigning, a set of e-mails related to the same thread for a given period as a mail set, and the description in a specific mail set article has a name attribute of a specific class on the taxonomy, Categorize all mail sets as taxonomies in a way that classifies a particular mail set into that particular class, and determines whether a particular phrase in a particular class can inherit the name attribute of that particular class. A child whose name attribute is the specific phrase if the specific phrase can inherit the name attribute of the specific class. It performs processing for complementing the lath as a child class of the particular class, the determination and the complementary treatment, and belongs to the parent class of the particular class Ci in number article article number equivalent to the particular class Ci The other class Cj is selected, and the appearance frequency Ki of the specific phrase K in the mail set classified into the specific class Ci, and the appearance frequency Kj of the specific phrase K in the mail set classified into the other class Cj, When the appearance frequency score U (K) is calculated and the average value of the appearance frequency scores U (K) calculated by repeating a predetermined number of times while changing other classes Cj is equal to or greater than a threshold value, the specific phrase K A taxonomy complementing method comprising complementing a related class having a name attribute as a child class of the specific class Ci . A taxonomy completion method executed by a computer based on software, in which a child class that inherits the properties of the parent class is set as a child class belonging to the parent class, and the child class has a name attribute common to the parent class. If a taxonomy is constructed by assigning, a set of e-mails related to the same thread for a given period as a mail set, and the description in a specific mail set article has a name attribute of a specific class on the taxonomy, Classify all mail sets as taxonomies in such a way that a specific mail set is classified into the specific class, determine whether there is a relationship between two specific classes, and if there is a relationship, A process that generates a child class common to two classes is performed, and the two specific classes are different child classes belonging to the same parent class. The relevance of the mail set of each of the two child class is calculated based on the simpson coefficient, its relevance is the case within a predetermined upper among all comparison class belonging to the one parent class, the A taxonomy completion method comprising complementing a fusion class composed of a common mail set of two different child classes as a common child class belonging to the two different child classes. A taxonomy complement program for causing a computer to execute the taxonomy complement method according to claim 1 . A computer-readable storage medium storing the taxonomy complement program according to claim 3 .

Images