JP5020274B2 - Semantic drift occurrence evaluation method and apparatus - Google Patents

Description

  The present invention relates to a semantic drift generation evaluation method and apparatus.

  Conventionally, as an acquisition method of specific knowledge such as a person name, place name, organization name, and semantic knowledge regarding these relationships, an instance belonging to the same category is extracted based on a small number of seed instances that are initial values and belong to a predetermined category, A bootstrap algorithm, which is a method of repeatedly increasing instances by extracted instances, has been proposed (for example, Non-Patent Document 1). In the bootstrap algorithm, an instance can be extracted from a general text sentence. More specifically, when a search log is used as an instance acquisition source, a word belonging to a predetermined category among a plurality of words constituting a search query included in the search log is set as a seed instance. Then, in the search query including the seed instance, a character string other than the seed instance is extracted as a pattern, and the instance is extracted from the extracted pattern. Then, a high fitness is assigned to those that can acquire instances belonging to a predetermined category with a high fitness, and a low fitness is assigned to those that can acquire irrelevant instances. Then, by using the patterns in descending order of the matching degree, the instances are acquired with a high matching ratio. Further, the process of acquiring a new instance from the acquired instance is repeatedly executed in the same manner as the method of acquiring an instance from the seed instance described above. Thus, the bootstrap algorithm has an advantage that a large number of instances can be obtained from a small amount of seed instances.

Mamoru Komachi, Kumi Suzuki, Improvement of semi-supervised acquisition of semantic knowledge from search logs, Journal of the Japanese Society for Artificial Intelligence, 2008, No.3, pp.217-225, Internet <URL: http: //www.jstage.jst .go.jp / article / tjsai / 23/3/217 / _pdf / -char / en />

  However, in the method described in Non-Patent Document 1, there is a possibility that the extracted pattern is a generic pattern that is a pattern that appears in a plurality of categories. Since the generic pattern is a pattern that co-occurs with a category other than the predetermined category, the method described in Non-Patent Document 1 can acquire an instance unrelated to the seed instance belonging to the predetermined category. Once an instance that is not related to the seed instance is acquired, a pattern for acquiring an instance that is highly related to an instance that is not related to the predetermined category is extracted, and the acquired instance is less related to the predetermined category There is a problem that a semantic drift may occur. Furthermore, semantic drift can also occur when an ambiguous instance that can belong to a plurality of categories is acquired regardless of the presence or absence of a generic pattern.

  The present invention has been proposed in view of such conventional problems, and an object thereof is to provide an evaluation method and apparatus capable of recognizing that a semantic drift has occurred.

  The present invention provides the following solutions.

  (1) Generation of semantic drift in which the meaning of a category transitions when an instance highly relevant to the predetermined category is acquired using a search log in a method of acquiring instances included in the predetermined category by a bootstrap algorithm A semantic drift occurrence evaluation method for evaluating a state, wherein a first extraction step of extracting a new instance based on the search log, and a new extraction extracted in the first extraction step by the bootstrap algorithm A second extraction step of extracting a related keyword of an instance from a pre-stored related keyword dictionary when repeating the execution of the first extraction step using a simple instance, and a numerical value assigned to the related keyword Step for generating related keyword vectors as elements If, based on the associated keyword vector at each iteration, it means drift occurs evaluation method comprising a step of evaluating the degree of meaning drift in each iteration.

  According to the semantic drift occurrence evaluation method described in (1), the first extraction process for extracting a new instance is performed based on the search log. Then, the execution of the first extraction process using the extracted instance is repeated by the bootstrap algorithm. When the execution of new instance extraction is repeated by the bootstrap algorithm, the related keyword of the instance is extracted from the related keyword dictionary stored in advance. And the related keyword vector which makes the numerical value allocated with respect to the extracted related keyword an element is produced | generated. Then, the degree of semantic drift in each iteration is evaluated based on the related keyword vector in each iteration.

  According to such a method, since the degree of semantic drift in each iteration is evaluated for the related keyword vector in each iteration and the related keyword vector immediately before the iteration, it is recognized that the semantic drift has occurred based on the evaluation contents. it can.

  (2) In the step of evaluating the degree of semantic drift, a cosine similarity is calculated for the related keyword vector in each iteration and the related keyword vector immediately before the iteration, and the degree of semantic drift is evaluated based on the cosine similarity. The semantic drift generation evaluation method according to (1), characterized in that:

  (2) According to the semantic drift occurrence evaluation method described in (2), the cosine similarity is calculated for the related keyword vector in each iteration and the related keyword vector immediately before the iteration, and the degree of semantic drift is evaluated based on the cosine similarity. . In this way, it is possible to measure how much the related keyword has transitioned from the previous iteration in the related keyword vector in each iteration.

  (3) The step of evaluating the degree of semantic drift includes a related keyword vector in each iteration and a related keyword vector of a seed instance that is an initial value when the new instance is extracted in the first extraction step. The cosine similarity is calculated, and the semantic drift occurrence evaluation method according to (1), wherein the degree of semantic drift is evaluated based on the cosine similarity.

  According to the semantic drift occurrence evaluation method described in (3), a cosine is used for a related keyword vector in each iteration and a related keyword vector of a seed instance that is an initial value when the new instance is extracted by the first extraction step. Similarity is calculated, and the degree of semantic drift is evaluated based on the cosine similarity. By doing in this way, it is possible to measure how much the related keyword of the instance extracted by repetition has shifted from the related keyword of the seed instance.

  (4) Generation of semantic drift in which the meaning of a category changes when an instance highly relevant to the predetermined category is acquired using a search log in a method of acquiring instances included in the predetermined category by a bootstrap algorithm A semantic drift occurrence evaluation device for evaluating a state, wherein an instance extraction unit for extracting a new instance based on the search log and a new instance extracted by the instance extraction unit by the bootstrap algorithm are used. Repetitive execution control means for repeating the execution of the instance extracting means, related keyword extracting means for extracting a related keyword of an instance from a pre-stored related keyword dictionary, and a relation having a numerical value assigned to the related keyword as an element keyword A vector generation means for generating a vector, based on the associated keyword vector in each of the iterations, the meaning drift generating evaluation device and a means drift evaluating means for evaluating the degree of meaning drift in each of the iterations.

  According to such a configuration, the same effect as in (1) can be expected by constructing the device.

  ADVANTAGE OF THE INVENTION According to this invention, the evaluation method and apparatus which can recognize that the semantic drift has generate | occur | produced can be provided.

It is a figure which shows the structural example of the meaning drift generation | occurrence | production evaluation apparatus 1 which concerns on this embodiment. It is a figure showing search log DB21 concerning this embodiment. It is a figure which shows the related keyword dictionary DB22 which concerns on this embodiment. It is a figure which shows the hardware constitutions of the meaning drift generation | occurrence | production evaluation apparatus 1 which concerns on this embodiment. It is a flowchart which shows the flow of the process which the semantic drift generation | occurrence | production evaluation apparatus 1 which concerns on this embodiment performs. It is a flowchart which shows the flow of the process of instance extraction in the instance extraction part 11 of the control part 10 which concerns on this embodiment. It is a figure which shows the measurement result of a seed similarity and a difference similarity.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[overall structure]
FIG. 1 is a diagram illustrating a configuration example of a semantic drift occurrence evaluation device 1 according to the present embodiment.

  The semantic drift occurrence evaluation device 1 uses a search log to acquire an instance highly relevant to a predetermined category in a method for acquiring instances (words) included in the predetermined category by a bootstrap algorithm. It is an apparatus for evaluating the occurrence state of semantic drift that acquires an instance having a low association with a category.

  The semantic drift occurrence evaluation device 1 includes a control unit 10, a storage unit 20, a display unit 31, and an operation unit 32. The control unit 10 includes an instance extraction unit 11, an iterative execution control unit 12, a related keyword extraction unit 13, a vector generation unit 14, and a semantic drift evaluation unit 15. The storage unit 20 also stores a search log database (hereinafter referred to as a database) 21 and a related keyword dictionary DB 22.

  The instance extraction unit 11 extracts a new instance with reference to the search log DB 21 (see FIG. 2 described later). More specifically, the instance extraction unit 11 extracts from the search log DB 21 a search query that includes instances that constitute a specified instance set. The designated instance set is a set of seed instances at the first time point when the process is started by the semantic drift occurrence evaluation device 1. After the instance set is extracted by the instance extraction unit 11, the instance extraction unit 11 An extracted instance set. The instance extraction unit 11 extracts words other than the instances included in the instance set as patterns from the extracted search query, and generates a pattern set composed of the extracted patterns. Then, based on the pattern set, a search query including a pattern constituting the pattern set is extracted from the search log DB 21. Then, words other than the pattern are extracted as instances from the extracted search query, and an instance set composed of the extracted instances is generated. The pattern reliability and the instance reliability are calculated at the time of generating the pattern set and the instance set, respectively. Then, a set of highly reliable instances is extracted as a new instance set belonging to a predetermined category.

により算出され、インスタンス集合Ｉ中の各インスタンスｉとパターンｐとの間の重み付き共起として定義されている。 The pattern reliability and the instance reliability will be described in more detail. Assume that the reliability of the pattern p in the pattern set P is rπ (p) (π is a subscript, the same applies hereinafter), and the reliability of the instance i in the instance set I is the reliability of the pattern rι (i) ( ι is a subscript, the same applies below), and rπ (p) is based on the intuition that a reliable pattern co-occurs with a reliable instance,

And is defined as a weighted co-occurrence between each instance i in the instance set I and the pattern p.

により算出される。 pmi (i, p) is a mutual information amount (PMI) between the instance i and the pattern p, and max pmi is a maximum mutual information amount between the pattern set and the instance set. pmi (i, p) is

Is calculated by

  | I, p | is the number of times the instance i and the pattern p are simultaneously searched, that is, the number of search queries including the instance i and the pattern p included in the search log DB 21. An asterisk is a wild card.

により算出される。 Similarly to the calculation method of the reliability of the pattern p, the reliability of the instance i is defined as an instance in which a highly reliable instance co-occurs with a highly reliable pattern,

Is calculated by

  In the present embodiment, rπ (p) and rι (i) are recursively defined as can be confirmed from the mathematical formulas shown in Equation 1 and Equation 3. The pattern reliability and the instance reliability are alternately calculated by the instance extraction unit 11 between a process for generating a pattern set and a process for generating an instance set.

  The repetitive execution control unit 12 controls repetitive execution of the instance extracting unit 11 using the instance set newly extracted by the instance extracting unit 11 by the bootstrap algorithm. More specifically, the iterative execution control unit 12 counts the number of processes of the instance extraction unit 11 and determines whether or not the number of times specified by the administrator of the semantic drift occurrence evaluation device 1 has been reached. When the specified number of times has not been reached, the instance extraction unit 11 repeatedly performs instance extraction. When the specified number of times has been reached, the instance extraction unit 11 ends the instance extraction. By repeating the instance extraction in this way, a large number of instances can be extracted.

  The related keyword extraction unit 13 refers to the related keyword dictionary DB 22 (see FIG. 3 described later) and extracts related keywords of the instance. More specifically, the related keyword extraction unit 13 refers to the related keyword dictionary DB 22 and extracts one or a plurality of related keywords for each instance extracted by the seed instance or the instance extraction unit 11. The seed instance is an initial value for acquiring an instance in the semantic drift occurrence evaluation apparatus 1 and is manually determined, and belongs to a predetermined category.

  There are two timings at which the related keyword extraction unit 13 extracts related keywords. That is, when the instance is a seed instance, the semantic drift occurrence evaluation device 1 extracts the related keyword before the instance extraction unit 11 extracts the first instance. When the instance is an instance extracted by the instance extraction unit 11, the related keyword is extracted immediately after a new instance is extracted by the instance extraction unit 11.

  The vector generation unit 14 assigns a numerical value to the related keyword extracted by the related keyword extraction unit 13 and generates a related keyword vector having the assigned numerical value as an element. More specifically, the vector generation unit 14 calculates the number of related keywords extracted by the related keyword extraction unit 13 in each instance in the instance set. Then, a numerical value obtained by dividing 1 by the calculated number of related keywords is assigned to the related keyword, and a related keyword vector having the assigned numerical value as an element is generated. Then, the keyword vectors of all instances included in the instance set are totaled.

  For example, when there are two instances X and Y included in the instance set A, four keywords a, b, c, and d are extracted as related keywords of X, and a and e keywords are extracted as related keywords of Y. Assume that two are extracted. Then, 0.25 which is a number obtained by dividing 1 by 4 is assigned to each of the related keywords of X as a related keyword vector, and a (0.25), b (0.25), c (0. 25) and d (0.25). In addition, 0.5, which is a number obtained by dividing 1 by 2, is assigned to each of the related keywords of Y as a related keyword vector, which becomes a (0.5) and e (0.5). The keyword vector of the instance set is a result of aggregating related keyword vectors of the instances X and Y, that is, a (0.75), b (0.25), c (0.25), d (0.25). ), E (0.5).

により算出される。このように、各反復の前後のキーワードベクトルに基づいて算出されたコサイン類似度は、差分類似度と呼ばれる。 The semantic drift evaluation unit 15 evaluates the degree of semantic drift in each iteration based on the related keyword vector assigned to the related keyword in each iteration. More specifically, the semantic drift evaluation unit 15 includes a related keyword vector generated by the vector generation unit 14 after iterative execution of the instance extraction unit 11, and a related keyword vector generated by the vector generation unit 14 immediately before the iterative execution. Cosine similarity is calculated. Here, the cosine similarity is a cosine distance between vectors. If the related keyword vector generated in each repetitive execution is A and the related keyword vector immediately before the repetitive execution is B, the cosine similarity sim (A, B) is

Is calculated by Thus, the cosine similarity calculated based on the keyword vectors before and after each iteration is referred to as difference similarity.

  For example, the keyword vector of the instance set A extracted in a certain iteration is a (0.75), b (0.25), c (0.25), d (0.25), e (0.5). If the keyword vector of the instance set B immediately before the iteration is a (0.33), b (0.33), and c (0.33), the cosine similarity sim ( A, B) is calculated as 0.72 based on the equation (4).

  Further, the semantic drift evaluation unit 15 cosine-likes the related keyword vector generated by the vector generation unit 14 in each iteration of the instance extraction unit 11 and the related keyword vector generated by the vector generation unit 14 in the set of seed instances. Calculate the degree. In this case, the expression for the cosine similarity is the same as the expression shown in Equation 4. Thus, the cosine similarity calculated based on the keyword vector of the seed instance and the keyword vector in a certain iteration is referred to as seed similarity.

  The difference similarity and the seed similarity both take values of 0 or more and 1 or less. The difference similarity measures how much the related keyword vector has transitioned from the previous iteration in the related keyword vector in each iteration, and the transition of the related keyword at a position where the numerical value is extremely lowered, that is, a predetermined category and It is considered that there is a semantic drift in acquiring instances that are not relevant. That is, when the related keywords are similar, the change in the keyword vector is small before and after the iteration, so that the cosine similarity is a value close to 1. If the related keywords are not similar, the keyword vector changes greatly before and after the iteration, and the cosine similarity becomes a value close to zero. That is, by monitoring the cosine similarity in the repeated execution of the instance extraction unit 11, it can be recognized that the keyword vector has changed greatly, that is, that a semantic drift has occurred.

  The seed similarity measures how much the related keyword of the instance extracted by iteration has transitioned from the related keyword of the seed instance. By introducing the seed similarity, the related keyword of the seed instance and the seed instance is measured. Based on the above, the degree of semantic drift of the instance set extracted by the repeated execution of the instance extraction unit 11 can be evaluated.

  In the present embodiment, the degree of semantic drift in each iteration is evaluated by calculating the cosine similarity, but the present invention is not limited to this. For example, the degree of semantic drift in each iteration may be evaluated by calculating the Euclidean distance and the Calbuckle-Librer distance. In this case, it is necessary to normalize the vector generated by the vector generation unit 14. Vector normalization refers to dividing each generated vector by the number of elements.

  FIG. 2 is a diagram showing the search log DB 21 according to the present embodiment. The search log DB 21 stores search queries included in a search log extracted from a predetermined search engine. The search log DB 21 has an “instance 1” field for storing the first word (instance) constituting the search log and an “instance 2” field for storing the second word (instance) constituting the search log. include. That is, in the present embodiment, the search log DB 21 stores a search query composed of two words divided into respective words (instances). The search log DB 21 is referred to when the instance extraction unit 11 extracts an instance.

  In the present embodiment, only the search query composed of two words is stored in the search log DB 21 and a new instance is extracted based on the search query. However, the present invention is not limited to this. For example, the search log DB 21 may store a search query including three or more words, a word decomposed by morphological analysis, or the like, and extract a new instance from the word. By doing so, the source data increases, so that various patterns can be extracted.

  FIG. 3 is a diagram showing the related keyword dictionary DB 22 according to the present embodiment. The related keyword dictionary DB 22 is a DB that stores an instance and a related keyword assigned to the instance in association with each other, and stores an instance and a related keyword extracted in advance from a predetermined dictionary site. The related keyword dictionary DB 22 includes an “instance” field indicating an instance and a “related keyword” field indicating a related keyword assigned to the instance. The related keyword dictionary DB 22 is referred to when the related keyword extraction unit 13 extracts related keywords of seed instances or newly generated instances.

  In the present embodiment, the instance and the related keyword are extracted from a predetermined dictionary site in advance. However, the present invention is not limited to this. For example, a keyword obtained from a snippet included in the search result may be used. Here, a snippet is an introductory sentence of a Web page included in a search result page in a search engine, and is an element that is highly likely to contain a keyword related to a search query. That is, a snippet is analyzed, a related keyword of a search query (instance) is extracted, and the search query and the extracted related keyword may be stored in the related keyword DB 22.

  The display unit 31 performs display related to the function of the semantic drift occurrence evaluation device 1 and visually displays information output by each function to a person who operates the semantic drift occurrence evaluation device 1. The operation unit 32 receives a direct input from a person who operates the semantic drift occurrence evaluation device 1.

[Hardware configuration of semantic drift generation evaluation device]
FIG. 4 is a diagram illustrating a hardware configuration of the semantic drift occurrence evaluation apparatus 1 according to the present embodiment. The semantic drift occurrence evaluation apparatus 1 in which the present invention is implemented may be a standard one, and an example of the configuration is shown below.

  The semantic drift occurrence evaluation device 1 includes a CPU (Central Processing Unit) 1010 (a plurality of CPUs such as a CPU 1012 may be added in a multiprocessor configuration), a bus line 1005, a communication I / F (I / F). F: Interface) 1040, main memory 1050, BIOS (Basic Input Output System) 1060, display device 1022, I / O controller 1070, input device 1100 such as keyboard and mouse, hard disk 1074, optical disk drive 1076, and semiconductor memory 1078 Prepare. The hard disk 1074, the optical disk drive 1076, and the semiconductor memory 1078 are collectively referred to as the storage unit 20.

  The control unit 10 is a part that comprehensively controls various functions related to the semantic drift occurrence evaluation apparatus 1, and cooperates with the hardware described above by appropriately reading and executing various programs stored in the hard disk 1074. Various functions according to the present invention are realized.

  The communication I / F 1040 is a network adapter when the semantic drift occurrence evaluation device 1 transmits and receives information to and from other servers via a communication network. The communication I / F 1040 may include a modem, a cable modem, and an Ethernet (registered trademark) adapter.

  The main memory 1050 temporarily stores data generated when the CPU 1010 executes various programs. The BIOS 1060 records a boot program executed by the CPU 1010 when the semantic drift occurrence evaluation apparatus 1 is started, a program depending on the hardware of the semantic drift occurrence evaluation apparatus 1, and the like.

  The display device 1022 includes a display device such as a cathode ray tube display device (CRT) or a liquid crystal display device (LCD), and functions as the display unit 31.

  The I / O controller 1070 can be connected to a storage unit 20 that is a storage device such as a hard disk 1074, an optical disk drive 1076, and a semiconductor memory 1078.

  The input device 1100 accepts input by the administrator of the semantic drift occurrence evaluation device 1 and functions as the operation unit 32.

  The hard disk 1074 stores various programs for causing the hardware to function as the semantic drift occurrence evaluation apparatus 1, a program for executing the functions of the present invention, and the above-described DB. The semantic drift occurrence evaluation apparatus 1 can also use a hard disk (not shown) separately provided as an external storage device.

  As the optical disk drive 1076, for example, a DVD-ROM drive, a CD-ROM drive, a DVD-RAM drive, and a Blu-ray Disc (registered trademark) drive can be used. When the optical disk drive 1076 is used, the optical disk 1077 corresponding to the optical disk drive 1076 is used. A program or data may be read from the optical disk 1077 by the optical disk drive 1076 and provided to the main memory 1050 or the hard disk 1074 via the I / O controller 1070.

  The computer in the present invention refers to an information processing device including a storage device, a control unit, and the like, and the semantic drift occurrence evaluation device 1 is information including the control unit 10 and the storage unit 20 as described above. The information processing apparatus is constituted by a processing apparatus and is included in the concept of the computer of the present invention.

  The semantic drift occurrence evaluation device 1 is not limited in the number of hardware, and may be configured by one or a plurality of hardware as necessary. In the case of a plurality of hardware, each hardware may be connected via a communication network. For example, the functions of the present embodiment may be realized by using separate servers (devices) for each function and linking the servers by transmitting and receiving signals between the servers.

[Processing flow]
FIG. 5 is a flowchart showing a flow of processing performed by the semantic drift occurrence evaluation apparatus 1 according to the present embodiment.

  In step S1, the control unit 10 (related keyword extraction unit 13) refers to the related keyword dictionary DB 22 and extracts related keywords of seed instances.

  In step S2, the control unit 10 (vector generation unit 14) assigns a numerical value to the related keyword extracted in step S1 and generates a related keyword vector. More specifically, the vector generation unit 14 calculates the number of related keywords extracted for each instance in the instance set extracted by the related keyword extraction unit 13. Then, the number obtained by dividing 1 by the number of the related keywords is assigned to the related keywords as a related keyword vector. Then, the keyword vectors of all the instances included in the set of seed instances are aggregated.

  In step S3, the control unit 10 (instance extraction unit 11) extracts a new instance with reference to the search log DB 21. The instance extraction will be described in detail with reference to FIG.

  In step S4, the control unit 10 (related keyword extraction unit 13) refers to the related keyword dictionary DB 22 and extracts the related keywords of the new instance extracted in step S3.

  In step S5, the control unit 10 (vector generation unit 14) assigns a numerical value to the related keyword extracted in step S4 and generates a related keyword vector.

  In step S6, the control unit 10 (the semantic drift evaluation unit 15) calculates the cosine similarity based on the related keyword vector assigned to the related keyword in each iteration, and evaluates the degree of the semantic drift in each iteration.

  In step S <b> 7, the control unit 10 (repetitive execution control unit 12) counts the number of processings of the instance extraction unit 11. The count of the number of processes is reset to 0 when the semantic drift occurrence evaluation apparatus 1 starts the process.

  In step S8, the control unit 10 (repetitive execution control unit 12) determines whether or not to continue the process. More specifically, it is determined whether or not the number of processes of the instance extraction unit 11 counted in step S7 has reached the number designated by the administrator of the semantic drift occurrence evaluation apparatus 1. If the determination result is YES, the process ends, and if the determination result is NO, the process proceeds to step S3.

  FIG. 6 is a flowchart showing a flow of instance extraction processing in the instance extraction unit 11 of the control unit 10 according to the present embodiment.

  In step S <b> 31, the control unit 10 (instance extraction unit 11) refers to the search log DB 21 and extracts a search query including instances that constitute the designated instance set. In step S32, the control unit 10 (instance extraction unit 11) extracts words other than the instances included in the specified instance set as patterns from the search query extracted in step S31, and is configured by the extracted patterns. Generate a pattern set.

  In step S33, the control unit 10 (instance extraction unit 11) calculates the reliability according to Equation 1 for all patterns included in the pattern set generated in step S32. In step S34, the control unit 10 (instance extraction unit 11) refers to the search log DB 21 and extracts a search query including a pattern included in the pattern set generated in step S32.

  In step S35, the control unit 10 (instance extraction unit 11) extracts words other than the patterns included in the pattern set generated in step S32 as instances from the search query extracted in step S34, and extracts the extracted instances. An instance set composed of is generated. In step S36, the control unit 10 (instance extraction unit 11) calculates the reliability according to the equation shown in Equation 3 for all instances included in the instance set generated in step S35.

  In step S37, the control unit 10 (instance extraction unit 11) extracts instances having high reliability as an instance set based on the reliability calculated in step S36, and ends the instance extraction processing.

[Experimental result]
Subsequently, in the semantic drift occurrence evaluation apparatus 1 according to the present embodiment, an experimental result when the degree of semantic drift in each iteration is evaluated when the generation of instances is repeatedly executed by the iteration execution control unit 12 is shown. In the experiment, as data stored in the search log DB 21 Yahoo! Among search logs for August 2008 of (registered trademark) search, a search log composed of two words separated by a space character was used. As a data to be stored in the related keyword dictionary DB 22, a dump of Wikipedia (registered trademark) July 24, 2008 was used.

  In addition, instances that are included in the instance set generated by the instance extraction unit 11 and have the highest reliability within the top 500 are determined as instances newly extracted by the instance extraction unit 11. In calculating the reliability of the instance, all patterns included in the pattern set were used. In addition, the number of iterations of the instance extraction process is 50. In addition, the names of five baseball players whose predetermined category is “baseball players” are used as seed instances.

  FIG. 7 is a diagram illustrating experimental results of seed similarity and difference similarity. The vertical axis indicates the difference similarity and the seed similarity, and the horizontal axis indicates the number of iterations.

  As shown in FIG. 7, the difference similarity and the seed similarity calculated by the semantic drift evaluation unit 15 after the fifth iteration execution of the instance extraction unit 11 are performed by the semantic drift evaluation unit 15 before the fifth iteration execution. It can be confirmed that the calculated difference similarity and seed similarity are greatly reduced. As a result of scrutinizing the replaced 236 instances in the fifth iteration, the instance newly extracted by the instance extraction unit 11 does not include a baseball player, that is, occurrence of semantic drift. It could be confirmed. Therefore, the semantic drift evaluation unit 15 shown in the present embodiment can recognize that the semantic drift has occurred by calculating the difference similarity and evaluating the degree of change of the difference similarity.

  In addition, since the seed similarity measures how much the related keyword of the instance extracted by the iteration has shifted from the related keyword of the seed instance, the related keyword of the instance extracted in the fifth iteration is determined. It can be confirmed that a significant transition has been made from the related keywords of the seed instance. Therefore, the semantic drift evaluation unit 15 can recognize that the semantic drift has occurred by calculating the seed similarity and evaluating the degree of change in the seed similarity.

  The embodiment of the present invention has been described above, but the present invention is not limited to the present embodiment, and modifications, improvements, and the like within the scope that can achieve the object of the present invention are included in the present invention.

DESCRIPTION OF SYMBOLS 1 Semantic drift generation | occurrence | production evaluation apparatus 10 Control part 11 Instance extraction part 12 Iterative execution control part 13 Related keyword extraction part 14 Vector generation part 15 Semantic drift evaluation part 21 Search log DB
22 Related keyword dictionary DB
31 Display unit 32 Operation unit

Claims (4)

In the method of acquiring an instance included in a predetermined category by a bootstrap algorithm, when an instance highly relevant to the predetermined category is acquired using a search log, an occurrence state of a semantic drift in which the meaning of the category changes is evaluated. Meaning drift generation evaluation method
A first extraction step of extracting a new instance based on the search log;
When repeating the execution of the first extraction step using the new instance extracted in the first extraction step by the bootstrap algorithm,
A second extraction step of extracting a related keyword of the instance from a related keyword dictionary stored in advance;
Generating a related keyword vector whose elements are numerical values assigned to the related keywords;
A semantic drift generation evaluation method comprising: evaluating a degree of semantic drift in each iteration based on a related keyword vector in each iteration.   The step of evaluating the degree of semantic drift includes calculating cosine similarity for the related keyword vector in each iteration and the related keyword vector immediately before the iteration, and evaluating the degree of semantic drift based on the cosine similarity. The semantic drift generation evaluation method according to claim 1, wherein:   The step of evaluating the degree of semantic drift includes cosine similarity with respect to a related keyword vector in each iteration and a related keyword vector of a seed instance that is an initial value when the new instance is extracted in the first extraction step. The semantic drift occurrence evaluation method according to claim 1, wherein the degree of semantic drift is evaluated based on the cosine similarity. In the method of acquiring an instance included in a predetermined category by a bootstrap algorithm, when an instance highly relevant to the predetermined category is acquired using a search log, an occurrence state of a semantic drift in which the meaning of the category changes is evaluated. Meaning drift generation evaluation device
An instance extracting means for extracting a new instance based on the search log;
Repetitive execution control means for repeating the execution of the instance extraction means using the new instance extracted by the instance extraction means by the bootstrap algorithm;
Related keyword extraction means for extracting a related keyword of an instance from a pre-stored related keyword dictionary;
Vector generation means for generating a related keyword vector whose elements are numerical values assigned to the related keywords;
A semantic drift generation evaluation device comprising semantic drift evaluation means for evaluating the degree of semantic drift in each of the iterations based on the related keyword vector in each of the iterations.

Images