JP6040141B2 - Keyword assigning device, keyword assigning method, and program - Google Patents

Description

  The present invention relates to a technique for assigning a related keyword to a document.

  Giving relevant keywords to a document is extremely important in data analysis and document search, for example. For example, in academic papers, in addition to the main text, related keywords are usually assigned manually. These keywords are useful information, such as knowing the trend of research trends simply by analyzing the frequency.

  If a document is created manually, it is easy to manually assign some related keywords. However, it takes a lot of time and labor to manually assign keywords to a large number of documents already created or a document obtained by converting a voice archive into a text using a voice recognition technology, which is extremely expensive. Therefore, a technique for automatically assigning keywords to documents is required. Such a technique is also called document classification or text classification (see Non-Patent Document 1).

  There are two main methods for automatically assigning keywords. One method is a type of method called latent semantic indexing. Typical methods of latent semantic indexing include probabilistic latent semantic analysis (Probabilistic Latent Semantic Analysis, pLSA) and latent Dirichlet Allocation (Latent Dirichlet Allocation, LDA) (see Non-Patent Document 2). Although latent semantic indexing can give various keywords, there is a problem that keywords meaningful to people are not always given.

  Another method is a method using a classifier (see Non-Patent Document 1). A number of keywords are determined in advance, and for each keyword, a classifier is prepared for determining whether a certain document is related to the keyword. The input document is applied to each classifier, and a keyword determined to be relevant is assigned. If the method uses a classifier, keywords can be determined in advance, and keywords that are meaningful to people can be set.

Nagata, Masaaki, Hirahiro Jun, “Text Classification:“ Trade Fair of Learning Theory ””, “Special Issue: Information Theory of Learning and Its Applications”, Information Processing, Vol. 42, No. 1, pp. 32-37 , 2001 Tomoharu Iwata, Takeshi Yamada, Nobuo Ueda, “Visualization of Documents Based on Topic Models”, IPSJ Transactions, Vol. 50, No. 6, pp. 1649-1659, 2009

  In the keyword assignment technique using a classifier, it is necessary to prepare as many classifiers as the number of keywords and apply the classifier to the input document brute force, so the processing time proportional to the number of keywords is required. Necessary. As a result, there is a problem that the number of keywords is practically limited to several tens to several hundreds or thousands due to time constraints for calculation.

  On the other hand, there is a case where it is desired to change a keyword to be preferentially given according to the situation. For example, there may be a case where a keyword related to a person name or a place name is given priority. When considering general-purpose keyword assignment technology, it is necessary to respond to such a request.

  An object of the present invention is to dramatically increase the number of keywords that can be assigned to a document. Furthermore, a specific keyword is preferentially given according to the situation.

  In order to solve the above problems, the keyword assignment device of the present invention includes a parameter calculation unit, a graph construction unit, a score calculation unit, and a graph scanning unit. The parameter calculation unit calculates a model parameter for each keyword using a plurality of documents to which the keyword is assigned. The graph construction unit calculates the degree of relativeity between the model parameters, and constructs a graph with links between nodes existing in the vicinity based on the degree of relativeity between the model parameters with the model parameter as each node. The score calculation unit calculates the score of the input document using the model parameter. The graph scanning unit searches the graph based on the score and identifies a keyword corresponding to the input document and the nearest node.

  According to the present invention, the number of keywords that can be assigned to a document can be dramatically increased. Furthermore, a specific keyword can be preferentially given according to the situation.

FIG. 1 is a diagram for explaining a graph search technique. FIG. 2 is a diagram illustrating a functional configuration of the keyword assigning apparatus according to the first embodiment. FIG. 3 is a diagram illustrating a processing flow of the keyword assigning method according to the first embodiment. FIG. 4 is a diagram illustrating a functional configuration of the keyword assigning apparatus according to the second embodiment. FIG. 5 is a diagram illustrating a processing flow of the keyword assigning method according to the second embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the component which has the same function in drawing, and duplication description is abbreviate | omitted.

[Points of Invention]
The present invention enables a keyword related to an input document to be selected at a high speed from a huge number (for example, about one million) of keywords. According to this invention, although a huge number of keywords can be handled, since a keyword group is prepared in advance by a human, it is possible to prevent assignment of meaningless keywords.

  In the present invention, the graph search technique is applied to a keyword assignment technique based on a classifier. Thereby, a large increase in the number of keywords and high-speed keyword assignment can be achieved at the same time.

The basic points of the present invention are the following two points.
1. A graph is constructed based on the degree of relativeity between model parameters, and keywords related to the input document are searched at high speed using the graph.
2. Bias can be set for keywords that should be prioritized during graph search.

  The former is a technical point for dramatically increasing the number of keywords, and the latter is a technical point for changing keywords to be preferentially assigned.

[Keyword assignment technology by classifier]
A classifier (hereinafter also referred to as a model) is a device that determines how much each keyword is appropriate for a given document. Specifically, the classifier returns a score S_Φ_λ (d) for document d and keyword label λ. A score S_Φ_λ (d) is calculated using a classifier for all keyword labels λ, and a keyword having a higher score S_Φ_λ (d) is assigned to the document d.

Φ_λ is a model parameter, and a specific arithmetic expression of the score S_Φ_λ (d) is unique to the model. Typical models used for assigning keywords to documents include Support Vector Machine (SVM), Boosting, Naive Bayes, and the like. The support vector machine and boosting are described in Non-Patent Document 1 above. Naive Bayes is described in Reference 1 below.
[Reference 1] Masumi Shirakawa, Kotaro Nakayama, Takahiro Hara, Shojiro Nishio, “Wikipedia Category Graph Analysis for Document Classification by Naive Bayes”, Proceedings of the 26th Annual Conference of the Japanese Society for Artificial Intelligence, 26th Japan Society for Artificial Intelligence National Convention, June 2012

  To learn a model, a database consisting of a plurality of documents with keywords is required. Learning is to estimate the value of the model parameter Φ_λ using a document database. Specifically, first, a predetermined statistic (feature vector) such as a word frequency or a tf-idf value is obtained from each document. The model parameter Φ_λ is determined using these statistics and the keyword label λ. The specific procedure for parameter estimation depends on the type of model and can be implemented by known techniques.

  In calculating the score S_Φ_λ (d) at the time of keyword estimation, a statistic (feature vector) obtained from the document, that is, a numerical value of the document is used.

  A specific example of a calculation formula for the score S_Φ_λ (d) is shown below.

  In the case of a linear model (for example, support vector machine, boosting, etc.), the score S_Φ_λ (d) can be calculated by the following equation (1).

Here, · is an inner product operator, and f (d) is a feature vector obtained from the document d.

  In the case of the naive Bayes model, the score S_Φ_λ (d) can be calculated by the following equation (2).

Here, P represents a probability, and f_k (d) is the kth element of the feature vector f (d). In Equation (2), P (λ) and {P_k (x | λ)} correspond to the model parameter Φ_λ.

The probability P is generally obtained by maximum likelihood estimation. For example, P (λ) can be obtained by dividing the frequency of the keyword λ by the total number of frequencies of all keywords. P_k (f_k (d) | λ) is expressed by the following equation (3) using the ratio of f_k (D_λ) to the sum Σ _k f_k (D_λ) for all the elements of the statistics obtained from all documents D_λ relating to the keyword λ. ).

  In addition to the above methods, there are the following methods for assigning keywords. First, a model parameter Φ_λ is obtained by a predetermined arithmetic expression from the document set D_λ related to the keyword λ. The model parameter Φ_λ may be, for example, the feature vector itself of the document set D_λ, or may be an average of feature vectors obtained from each document and a covariance matrix. Next, the similarity between the input document d and the document set D_λ is evaluated with a predetermined similarity (or distance scale). Finally, a keyword having high similarity with the input document d is assigned. As the similarity (or distance scale), for example, cosine similarity, Euclidean distance or Mahalanobis distance is widely used.

  This method is different from the above-described general classifier in that it does not include a process of estimating the model parameter Φ_λ during model learning. However, since it is common to classify based on a predetermined score S_Φ_λ (d) defined by a predetermined model parameter Φ_λ, it is treated as a kind of classifier here.

[Graph Search Technology]
The graph search is a technique for finding a certain input sample and the nearest sample at high speed in the presence of a large number of samples in a certain metric space.

  First, a large number of samples to be searched are regarded as nodes, and links are established in advance to nodes that are close to each other. When searching, the distance between the input sample (hereinafter referred to as a query) and the node is measured in order while following the link. First, an initial node is selected, and the distance between the initial node and the query is measured. Next, the distance between the node linked to the initial node and the query is measured. If there are multiple links in the initial node, the distance between all connected nodes and the query is measured. When searching by the greedy method, the node having the closest distance to the query is selected from the nodes whose distances are measured, and the distance between the link destination node extending from the node and the query is further measured. This process is repeated as long as a node closer to the query is found. Then, a sample corresponding to the nearest node finally found is used as a search result. If necessary, the initial node may be changed for searching. In addition to searching by the greedy method, the search may be performed while considering a plurality of upper nodes.

To find the nearest sample in the query exactly, the distance to all samples must be measured. However, by configuring the graph appropriately, in many cases, the nearest sample can be reached with a small number of measurements. Typical graph structures include k-neighborhood graph (see Reference 2 below), k-degree reduced neighborhood graph (k-degree reduced neighborhood graph) with redundant links reduced from k-neighbor graph (see below) Reference 3)).
[Reference 2] Masajiro Iwasaki, “Neighbor search of metric space using approximate k nearest neighbor graph”, Transactions of Information Processing Society of Japan. Database, vol. 3, no. 1, pp. 18-28, 2010
[Reference 3] Kazuo Aoyama, Kazumi Saito, Hiroshi Sawada, Naonori Ueda, “Fast approximate similarity search based on degree-reduced neighborhood graphs”, Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining, pp. 1055- 1063, 2011

  FIG. 1 shows an example of a graph search by a greedy method using a k neighborhood graph with k = 2. A circle drawn with a bold line and hatched is an initial node S, and a white circle drawn with a bold line is a query d (that is, an input document). First, the distance between the initial node S and the query d is measured. Next, the distance to the query d is similarly measured for the nodes N1, N2, and N3 linked to the initial node S (dotted arrow). Among the nodes S, N1, N2, and N3 whose distances are measured, the transition is made to the nearest node N3 (thick arrow), and the distance to the query d is measured again for the linked nodes N4 and N5. In the example of FIG. 1, when the bold arrow is traced three times, the node N6 in which no node closer to the query d can be found at the link destination is reached. Here, the search is terminated, and the node N6 is output as the nearest node.

[First embodiment]
1st Embodiment of this invention is the keyword provision apparatus which applied the graph search technique to the keyword provision technique based on a classifier. Thereby, a large increase in the number of keywords and high-speed keyword assignment can be achieved simultaneously.

  With reference to FIG. 2, an example of a functional configuration of the keyword assigning apparatus 1 according to the first embodiment will be described. The keyword assigning apparatus 1 includes a database storage unit 10, a parameter calculation unit 11, a graph construction unit 13, a graph storage unit 14, a query input unit 20, an initialization unit 21, a score calculation unit 22, a graph scanning unit 23, and an upper keyword storage unit. 24 and a result output unit 25. The keyword assigning device 1 may be configured to include the parameter modifying unit 12. A modification including the parameter modification unit 12 will be described later. The keyword assigning device 1 is, for example, a special program configured by reading a special program into a known or dedicated computer having a central processing unit (CPU), a main storage device (Random Access Memory, RAM), and the like. Device. For example, the keyword assigning apparatus 1 executes each process under the control of the central processing unit. The data input to the keyword assigning device 1 and the data obtained in each process are stored in, for example, the main storage device, and the data stored in the main storage device is read out as needed and used for other processing. Is done. Each storage unit included in the keyword assigning device 1 includes, for example, a main storage device such as a RAM (Random Access Memory), an auxiliary storage device configured by a semiconductor memory element such as a hard disk, an optical disk, or a flash memory, or It can be configured with middleware such as a relational database or key-value store. Each storage unit included in the keyword assigning device 1 only needs to be logically divided, and may be stored in one physical storage device.

  With reference to FIG. 3, an example of the processing flow of the keyword assignment method executed by the keyword assignment apparatus 1 according to the first embodiment will be described according to the order of procedures actually performed.

<Graph construction processing>
The database storage unit 10 stores a database including a plurality of documents to which keywords are assigned.

  In step S11, the parameter calculation unit 11 uses the plurality of documents stored in the database storage unit 10 to calculate model parameters for all keywords attached to the document. The model parameter calculation method is the same as the model learning method in the keyword assignment technique using the conventional classifier described above. The calculated model parameter is input to the graph construction unit 13.

  In step S <b> 13, the graph construction unit 13 calculates the degree of relativeity between the model parameters, and constructs a graph in which the model parameters are used as nodes and links are established between nodes that are close to each other.

  In the conventional keyword assignment technique based on the classifier, it is only necessary to evaluate the similarity of the input document d in the model parameter Φ_λ related to the keyword λ using the score S_Φ_λ (d). That is, it is only necessary to focus on the function of the model parameter Φ_λ and the input document d. However, when constructing a graph in the present invention, there is no input document d. A graph must be constructed with respect to model parameters, in which case a space is defined that defines the degree of affinity between any two model parameters. This is a new point that does not appear in conventional keyword assignment techniques.

  Specifically, the degree of relativeity between model parameters is configured as follows. The distance between the model parameters is non-relative, and is used as the relative degree by reversing the sign. In the following, the degree of affinity between the two model parameters Φ_λ and Φ_λ ′ is expressed as F (Φ_λ || Φ_λ ′).

  When a probability model such as a naive Bayes model is used, a similarity measure generally used in the probability model may be used. As such a similarity measure, for example, Kullback-Leibler divergence, f-divergence, Pearson function, Lp norm distance, and the like can be used.

  When Kullback-Leibler divergence is used, the degree of relative F (Φ_λ || Φ_λ ′) is calculated by the following equation (4).

Here, P is the probability of having Φ_λ as a parameter, and P ′ represents the probability of having Φ_λ ′ as a parameter. In the case of a document document, x is often a word. At this time, integration corresponds to addition of discrete symbols.

  When f-divergence is used, the degree of relativeity F (Φ_λ || Φ_λ ′) is calculated by the following equation (5).

Here, g is an arbitrary function.

  When the Pearson function is used, the degree of relative F (Φ_λ || Φ_λ ′) is calculated by the following equation (6).

  When the Lp norm distance is used, the degree of relative F (Φ_λ || Φ_λ ′) is calculated by the following equation (7).

  When the model parameter is a vector or when it is regarded as a vector, a similarity measure generally used for a vector may be used. As such a similarity measure, for example, Minkowski distance (including Euclidean distance and Lp norm distance), Mahalanobis distance, Manhattan distance, Chebyshev distance, standard Euclidean distance, cosine similarity, Pearson correlation coefficient, etc. are used. be able to. The Pearson correlation coefficient is obtained by dividing the cosine similarity of a vector by an average value.

  As long as the degree of relativeity between parameters is defined, the graph construction method itself may be applied with a conventional graph construction technique. The constructed graph is stored in the graph storage unit 14.

<Graph search processing>
In step S <b> 20, the query input unit 20 receives a document to which a keyword is assigned. Hereinafter, the input document is also called a query.

  In step S21, the initialization unit 21 initializes information used in the graph search process. Specifically, determination of an initial node for graph search, initialization of the upper keyword storage unit 24, feature vectorization of the input document, and the like are performed.

  In step S22, the score calculation unit 22 calculates the score of the input document using the model parameters corresponding to each of the predetermined node and the link destination node of the predetermined node in the graph stored in the graph storage unit 14. To do. The predetermined node is an initial node determined by the initialization unit 21 in the first process, and a node that has been transitioned by the immediately preceding process in the second and subsequent processes. The calculated score is input to the graph scanning unit 23.

  In step S23a, the graph scanning unit 23 searches the graph stored in the graph storage unit 14 based on the input score.

  The search on the graph requires a contrivance with respect to the conventional graph search technique. Each node of the graph corresponds to a model parameter, but the input (query) to the keyword assigning device is a document. In the conventional graph search, it is assumed that the node and the query correspond to the same event. More specifically, the graph search is a technique based on the premise that the same relativeity space is used for the graph construction and the graph search.

  In the present invention, the above assumption is removed as an engineering approximation. That is, the graph construction is performed based on the degree of affinity F (Φ_λ || Φ_λ ′) between parameters, but the graph search is performed based on the score S_Φ_λ (d) calculated using the model parameter Φ_λ and the document d. To do. This is due to the fact that most of the elements of the model parameter Φ_λ used in the keyword assignment technique correspond to a certain word or n-gram. Both the function F (Φ_λ || Φ_λ ') between model parameters and the function S_Φ_λ (d) between the model parameter Φ_λ and the document d are, as expected, vectors (or sets) whose elements correspond to words or n-grams. It is a function that expresses the relationship. That is, both functions do not form completely different spaces. In the present invention, this point is utilized, and different degree-of-kindness spaces are used at the time of graph construction and graph search.

  When constructing a graph, it is necessary to define a similarity space, but the constructed graph has only information about which neighboring nodes of a certain node, and does not hold information such as how close it is. In other words, the similarity space used at the time of graph construction at the time of graph search is unnecessary, and even if it is not a nearest neighbor even if it is viewed from the distance space at the time of graph search, a link is established to a node in the vicinity, It can be expected that reasonable search results will be obtained.

  However, as a matter of course, the degree of closeness F (Φ_λ || Φ_λ ′) between model parameters may coincide with the score S_Φ_λ (d). For example, as described above, the model parameter is the feature vector itself. For example, there is a technique for assigning keywords based on the cosine similarity of the tf-idf vector obtained from the input document. In this case, both the model parameter Φ and the feature vector f (d) obtained from the document d are tf-idf vectors, so if the cosine similarity is used for both the graph construction and the graph search, The degree of relative F (Φ_λ || Φ_λ ′) and the score S_Φ_λ (d) match. This example is a special form of the present invention, but does not depart from the present invention.

  The graph scanning unit 23 sequentially traces the link destinations of the nodes, and if a node approaching the query is not found by the link destination, the graph scanning unit 23 gives another initial node to perform the search again, or restarts the search from the second candidate. Or, complicated processing is executed. A method for searching for such a graph is the same as the conventional graph search technique.

  The graph scanning unit 23 specifies a keyword corresponding to the input document searched as described above and the nearest node. The identified keyword and score are stored in the upper keyword storage unit 24. In the upper keyword storage unit 24, keywords and scores are stored in descending order of score.

  In step S23b, the graph scanning unit 23 determines whether to end the graph search. The end condition of the graph search is, for example, whether the number of measurements of the distance between the query and each node has reached a predetermined upper limit, whether the number of times to reach the vertex node has reached a predetermined upper limit, etc. It is. A vertex node is a node where there is no node closer to the query at the link destination of the node.

  In step S25, the result output unit 25 outputs a predetermined number of keyword / score pairs stored in the higher keyword storage unit 24 in descending order of score.

  As described above, the keyword assigning apparatus 1 according to the first embodiment constructs a graph having the model parameter as each node based on the degree of relativeity between the model parameters, and based on the score calculated using the input document and each model parameter. By searching the graph, it is possible to efficiently search for a keyword to be given to the input document. As a result, the number of keywords can be dramatically increased as compared with the keyword assignment technique using the conventional classifier.

[Modification]
The keyword assigning device 1 may be configured to include a parameter modifying unit 12 as indicated by a dotted line in FIG. The parameter modification unit 12 modifies the model parameter output from the parameter calculation unit 11 to generate an analog. The modified model parameter is input to the graph construction unit 13. The analog is a parameter made up of a part of model parameters, a quantized parameter, or the like.

  Consider a parameter Ψ_λ that is similar to the model parameter Φ_λ. For example, this is the case when the parameter Ψ_λ is composed of a part of the model parameter Φ_λ. In this case, the degree of relativeity F (Ψ_λ || Ψ_λ ′) between model parameters is used at the time of graph construction, and the score S_Φ_λ (d) is used at the time of graph search.

  For example, when the score S_Φ_λ (d) is the Mahalanobis distance, the model parameter Φ_λ is an average of feature vectors and a covariance matrix. However, the method of defining the degree of relativeity (distance) between covariance matrices is not always clear. Even if it is defined, an increase in calculation cost can be easily imagined from the large number of elements in the matrix. Therefore, the graph is constructed with the Euclidean distance between the mean vectors, and the search is performed with the Mahalanobis distance.

  By configuring the keyword assigning apparatus 1 so as to include the parameter modifying unit 12, it is possible to obtain an advantage that the processing cost for constructing a graph having a high calculation cost can be reduced.

[Second Embodiment]
When constructing a general-purpose keyword assigning apparatus that handles a huge number of keywords, there is a case where it is desired to change a keyword to be given preferentially according to the situation. For example, there is a case where priority is given to a person name keyword, or a case where priority is given to a concept keyword having a relatively high degree of abstraction such as politics or sports. Although it is possible to prioritize specific keywords by configuring the graph in this way in advance, it is not practical to reconfigure the graph each time because the processing cost increases.

  In the second embodiment of the present invention, the keyword assignment device according to the first embodiment is configured so that a specific keyword can be given preferentially according to the situation.

  With reference to FIG. 4, an example of a functional configuration of the keyword assigning apparatus 2 according to the second embodiment will be described. Similar to the keyword assignment device 1 according to the first embodiment, the keyword assignment device 2 is a database storage unit 10, a parameter calculation unit 11, a graph construction unit 13, a graph storage unit 14, a query input unit 20, an initialization unit 21, A score calculation unit 22, a graph scanning unit 23, an upper keyword storage unit 24, and a result output unit 25 are included. The keyword assigning device 2 further includes a bias storage unit 15 and a bias addition unit 26. The keyword assigning device 2 may be configured to include the parameter modifying unit 12 as in the keyword assigning device 1 according to the first embodiment.

  The bias storage unit 15 stores a keyword to be preferentially assigned and a bias indicating the priority as a set. There are various methods for determining the bias value. For example, the following three methods can be mentioned.

  The first method uses a search word and the number of searches. A keyword with high search frequency is extracted by a search engine provided on the Internet (for example, GOOGLE (registered trademark), YAHOO! (Registered trademark), etc.), and an arbitrary value is determined according to the number of searches. Set as. For example, it is possible to preferentially assign hot topics that are popular in the world by giving a large bias to keywords that are frequently searched.

  The second method uses the thesaurus hierarchy information. The thesaurus is a representation of vocabulary relationships in a tree structure. Generally, the higher the hierarchy, the more abstract concepts are given. A value arbitrarily determined according to the thesaurus hierarchy is set as a bias for the keyword registered in the thesaurus. The thesaurus to be used may be appropriately determined according to the field of the keyword to be given with priority. For example, by giving a larger bias to a keyword with a lower thesaurus hierarchy, it becomes possible to preferentially assign the subdivided keywords. On the other hand, it is possible to give an abstract keyword preferentially by giving a larger bias to a keyword having a higher thesaurus hierarchy.

  The third method uses a rule based on part of speech or the like. Give bias according to keyword attributes such as names of people and places. The keyword that gives the bias and the value of the bias that is given to the keyword may be determined arbitrarily.

  With reference to FIG. 5, an example of the processing flow of the keyword assigning method executed by the keyword assigning apparatus 2 according to the second embodiment will be described according to the order of procedures actually performed.

  The processing from step S11 to step S22 is the same as the graph construction processing in the keyword assigning apparatus 1 according to the first embodiment.

  In step S <b> 26, the bias addition unit 26 adds the bias b_λ for the keyword λ stored in the bias storage unit 15 to the score S_Φ_λ (d) calculated by the score calculation unit 22. Specifically, the bias b_λ is added to the score S_Φ_λ (d). Alternatively, the score S_Φ_λ (d) is multiplied by the bias b_λ. As a result, a keyword given a large bias is preferentially assigned to the top of the search result.

  The processing from step S23a to step S25 is the same as the graph construction processing in the keyword assigning apparatus 1 according to the first embodiment.

  In this way, the keyword assignment device 2 according to the second embodiment preferentially assigns a specific keyword according to the situation while handling an enormous number of keywords like the keyword assignment device 1 according to the first embodiment. Can be granted. As a result, it is possible to achieve detailed keyword assignment that matches the user's needs.

  The present invention is not limited to the above-described embodiment, and it goes without saying that modifications can be made as appropriate without departing from the spirit of the present invention. The various processes described in the above embodiment may be executed not only in time series according to the order of description, but also in parallel or individually as required by the processing capability of the apparatus that executes the processes or as necessary.

[Program, recording medium]
When various processing functions in each device described in the above embodiment are realized by a computer, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, various processing functions in each of the above devices are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

DESCRIPTION OF SYMBOLS 1, 2 Keyword assignment apparatus 10 Database storage part 11 Parameter calculation part 12 Parameter modification part 13 Graph construction part 14 Graph storage part 15 Bias storage part 20 Query input part 21 Initialization part 22 Score calculation part 23 Graph scanning part 24 Upper keyword storage Unit 25 result output unit 26 bias addition unit

Claims (6)

A parameter calculation unit that calculates a model parameter for each keyword using a plurality of documents to which the keyword is assigned;
A graph construction unit that calculates a degree of relativeity between the model parameters, and constructs a graph in which links are established between nodes that are adjacent to each other based on the degree of relativeity between the model parameters, with the model parameter as each node;
A score calculation unit for calculating the score of the input document using the model parameters;
A graph scanning unit that searches the graph based on the score and identifies the keyword corresponding to the input document and the nearest node;
Keyword assignment device including The keyword assigning device according to claim 1,
A parameter modifying unit that selects a part of the model parameter or quantizes the model parameter and modifies the model parameter;
The said keyword construction part calculates the relative degree between the said modified model parameters, and construct | assembles the said graph The keyword provision apparatus. The keyword assignment device according to claim 2,
The parameter calculation unit uses the mean and covariance matrix of the feature vectors of the input document as the model parameters,
The graph construction unit uses the Euclidean distance between the mean of the feature vectors as the relative degree,
The said score calculation part uses the Mahalanobis distance of the said input document and the said model parameter as said score, The keyword provision apparatus. It is a keyword provision apparatus in any one of Claim 1 to 3,
A keyword adding device further comprising: a bias adding unit that applies a predetermined bias to the score to the keyword. A parameter calculating step for calculating a model parameter for each keyword using a plurality of documents to which the keyword is assigned;
The graph construction unit calculates the degree of relativeity between the model parameters, and constructs a graph with links between nodes existing in the vicinity based on the degree of affinity between the model parameters, with the model parameters as nodes. Construction steps and
A score calculation step in which the score calculation unit calculates the score of the input document using the model parameters;
A graph scanning step in which the graph scanning unit searches the graph based on the score and identifies the keyword corresponding to the node closest to the input document;
Keyword assignment method including   The program for functioning a computer as a keyword provision apparatus in any one of Claim 1 to 4.

Images