JP4963341B2 - Document relationship visualization method, visualization device, visualization program, and recording medium recording the program - Google Patents

Description

  The present invention relates to an inter-document relationship visualization method and a visualization device for visualizing relationships between categorized documents.

In recent years, a large number of documents have been electronically accumulated, and these document groups are usually classified for each related document so that a user can easily find a target document. For example, in a search site called “open directory project” or “goo®”, a large number of web pages are classified by topic by a directory-type search engine. Also, documents other than web pages, such as electronic books, patent documents, papers, and electronic files for computers, are classified into categories as in the case of web pages.
However, the classified document group includes, for example, documents classified due to human error, or documents that do not belong to any of the categories already set (this is referred to as “singular document”). Is included). In addition, there are cases where the classification system is biased toward specific fields. As described above, it is often the case that documents are not properly classified.
Document misclassification can be found by manually checking each document. However, since it is difficult to manually check a large number of documents, the suitability of document classification is often not evaluated.

  Therefore, as a method for evaluating the appropriateness of document classification, a visualization method that expresses the relationship between documents in an easy-to-understand manner is used. This visualization method highlights the structural features of the underlying document and may provide important clues for discovering new knowledge about the document group (e.g., document classification evaluation).

Conventionally, there are the following two methods as such a visualization method. One is a method of obtaining the similarity between documents from the word frequency vector of the document and visualizing the relationship between the documents based on the similarity. That is, it is a technique called MDS (Multi-Dimension Scaling) (for example, refer nonpatent literature 1). The other is a method in which links between documents are expressed as a document network with links between nodes, and the document network is visualized from the distance between the nodes (for example, see Non-Patent Document 2).
M. chalmers and P. chitson, “BEAD: Explorations in information visualiza- tion”, SIGIR '92, Proceedings of the Fifteenth Annual International ACM SIGIR Conference on Resarch and Development in Information Retrieval, ACM Press, 1992, p. 330-337 JBTenenbaum, V. de Silva and JCLangford, A global geometric framework for nonlinear dimensionality reduction, Science, 290, 2000, p. 2319-2323

  However, the methods disclosed in Non-Patent Document 1 and Non-Patent Document 2 both have a problem that many nodes are concentrated and the relationship between documents is difficult to understand.

  Therefore, the present invention has been made to solve the above-described problems, and the purpose thereof is to make it possible to express the relationship between categorized documents in an easy-to-understand manner, an inter-document relationship visualization method, a visualization device, and a visualization It is to provide a program and a recording medium on which the program is recorded.

In order to solve the above problems, the present invention relates to a document generation model representing a word frequency distribution of words appearing in a document for each category, and a document generation model classified into categories registered in advance. A storage device that stores information on a two-dimensional or three-dimensional visualization space for visually expressing the relationship between each document and each category, and the probability that each document in the document group belongs to a registered category A document by a visualization device comprising: a posteriori probability estimation unit that creates a posteriori probability vector whose components are each and a probability that does not belong to any of the registered categories; and a visualization unit that performs visualization processing of the document group. In the inter-relationship visualization method, the posterior probability estimation unit reads the document generation model for each category from the storage device, and For each document in the book group, the probability of belonging to a registered category is calculated based on the document generation model for each category, and the document generation of the category with the highest posterior probability among the document generation models for each category is calculated. Based on the model, calculating a probability that does not belong to any of the registered categories, creating a posterior probability vector having each calculated probability as a component, and the visualization unit creating the created posterior probability vector In accordance with each probability represented by the above, on the visualization space that is information stored in the storage device, the higher the probability that each document belongs to each category, the greater the distance from the center coordinate of the category of the document. Arranging the coordinates of each of the documents and the center coordinates of the categories so that the A step of the respectively of the coordinate information the external output of the categories visualization space disposed between respective center coordinates, characterized in that the run.

The present invention also relates to a document generation model that represents a word frequency distribution of words appearing in a document for each category, and each document in the document group and each category for a group of documents that are classified in advance. A storage device for storing information of a two-dimensional or three-dimensional visualization space for visually expressing the relationship between the two, a probability of belonging to a registered category for each document in the document group, and a registered A visualization apparatus comprising: a posterior probability estimation unit that creates a posterior probability vector having a probability that does not belong to any of the categories; and a visualization unit that performs visualization processing of the document group, wherein the posterior probability The estimation unit reads the document generation model for each category from the storage device, and for each document in the document group, for each category. The probability of belonging to a registered category is calculated based on the document generation model of, and among the document generation models for each category, the registered category is based on the document generation model of the category with the highest posterior probability. A probability that does not belong to any of the above, including an estimation function to create a posterior probability vector having each calculated probability as a component, the visualization unit according to each probability represented by the created posterior probability vector, On the visualization space, which is information stored in the storage device, the higher the probability that each document belongs to each category, the shorter the distance from the center coordinate of the category of the document to the document. Visualization function to arrange each coordinate and center coordinate of each category, and each document coordinate and each category And an output function of the information of the visualization space for the external output arranged and heart coordinates, characterized in that it comprises a.

  The present invention is also a visualization program for causing a computer to execute the above-described inter-document relationship visualization method. Furthermore, the present invention is a computer-readable recording medium on which a visualization program for causing a computer to execute the inter-document relationship visualization method is recorded.

  According to the present invention, the relationship between categorized documents can be easily expressed.

Hereinafter, embodiments of the present invention will be described with reference to FIGS.
FIG. 1 is a block diagram showing a system including a visualization device according to an embodiment of the present invention.
In FIG. 1, the visualization apparatus 10 is connected to a user terminal 30 via a communication network 20 such as the Internet.
The visualization device 10 includes a communication device 11, a storage device 12, and a processing device 13. For example, a computer such as a server device is used for this. The communication device 11 is an input / output interface, the storage device 12 is a memory, a hard disk, or the like, and the processing device 13 is a CPU or the like.
In addition, although the single visualization apparatus 10 is shown in FIG. 1, you may comprise the visualization apparatus 10 so that a distributed process may be performed using a some computer.

The storage device 12 stores document generation models for document groups classified into categories registered in advance for each category. The category is used for administratively classifying electronic documents such as Web pages, electronic books, and patent gazettes. For example, a topic corresponds to this.
As will be described in detail later, the document generation model represents a word frequency distribution of words appearing in the document. The document generation model is stored in the storage device 12 in a file format, for example.

  The processing device 13 includes a document generation model construction unit 131, a posterior probability vector estimation unit 132, and a visualization unit 133. The functions of these units 131 to 133 will be described later.

  The user terminal 30 is a computer such as a personal computer, and has the following general configuration. That is, the user terminal 30 includes an input device such as a keyboard, a display device such as a computer display, a storage device such as a hard disk and a memory, and a processing device such as a CPU.

Next, characteristic processing of the visualization apparatus 10 will be described.
[Document generation model construction process]
In the visualization device 10, the document generation model construction unit 131 performs a document generation model construction process using functions from the following formulas (1) to (3). Here, a case where a document generation model for each category is constructed from a document group classified in advance will be described.

First, a group of documents classified in advance is defined as follows. Let the classified document group be D = [d ₁ , d ₂ ,..., D _N ]. dn represents the nth document, and N represents the total number of given documents. Also, let C = [1, 2,..., K,. Where K is the total number of categories. Each document is classified into one category, and the category of the nth document is c _n ∈C. Further, a set of all words included in a given document is set to W [w ₁ , w ₂ ,..., W _v ]. w _j represents the j-th word, and V represents the total number of words.

The content of the document is expressed by a word frequency vector. That is, here, for example, a document expression called BOW (Bag-of-Words) in which the order of words in the document, dependency, and the like are ignored is used. In text classification, document representation by BOW is often used. Here, it is assumed that the word frequency vector of the nth document is x _n = (x _n1 , x _n1 ,..., X _nv ). x _ni represents the number of the j-th word included in the n-th document.

For example, a naive Bayes model (hereinafter abbreviated as “NB model”, NB: Naive Bayes) is used as a probability model for document generation for each category, that is, a document generation model. The NB model represents the contents of a document with a word frequency vector. Note that the NB model is used as the document generation model, but for example, an n-gram model or the like may be used.
In the NB model, it is assumed that the generation probability of the document d _n belonging to the category c _n is a multinomial distribution of the following equation (1).

The parameter θ _k = [θ _k1 , θ _k2 ,..., Θ _kv ] of the k-th category document generation model is estimated by maximizing the likelihood L _k of the following equation (2). In equation (2), a prior probability hyperparameter λ _k is introduced.

With λ _k in equation (2), it is possible to prevent the generation probability of a word that has never appeared in a given document from being zero, and to improve generalization performance. Note that the generalization performance means performance for consolidating to upper nodes.

The maximum likelihood estimated value of θ _kj is expressed by the following equation (3).

In order to prevent over-learning, the hyperparameter λ _k in equation (3) is estimated here by a cross-validation method.

[A posteriori probability vector estimation process]
In the visualization device 10, the posterior probability vector estimation unit 132 performs posterior probability vector estimation processing using functions from the following formulas (4) to (12).
A posteriori probability vector is a vector that represents the probability of a document belonging to each category (probability of belonging to a registered category) and the probability of not belonging to any category (probability of not belonging to a classified category). It is. Here, for example, by the maximum entropy method (see “K. Nigam, J. lafferty and A. McCallum, Using maximum entropy for text classification, In IJCAI-99 Workshop on Machine Learning for Information Filtering, 61-67, 1999”). An estimation method will be described.
In the maximum entropy method, when a function representing a relationship between an nth document and a kth category (called a feature function) is a feature f (d _n , k), the following equation (4) The parameter is estimated so that the entropy of the probability P (k | d _n ) is maximized while satisfying the above constraint. The probability P (k | d _n ) is a probability belonging to the kth category when the nth document is given. Note that the function of Equation (4), which is the maximum entropy condition, is stored in the storage device 12.

Then, there exists a unique distribution that maximizes the entropy of the probability P (k | d _n ). This distribution is expressed by the following equation (5).

Β in the equation (5) is a parameter to be estimated. In Expression (5), k takes an integer value from 1 to K + 1. 1 to K represent existing categories, and K + 1 represents a category other than the existing categories. P (K + 1 | d _n ) represents the probability that the nth document does not belong to any of the existing categories.

Next, the feature (d _n , k) between the n-th document and the k-th category (classified one) is expressed by the following equation (6). Here, the log likelihood of the nth document when the kth category in the NB model is given is used.

According to equation (6), if the word frequency of the nth document is similar to the document generation model of the kth category, the feature (d _n , k) increases, and the posterior probability (k | d _n ) belonging to the kth category. Also gets higher. Therefore, it is considered appropriate to use log-likelihood for the feature to estimate the posterior probability.
When limited to the existing K categories, the posterior probability in the NB model is expressed by the following equation (7) by Bayes' theorem.

According to Equation (7), and P _NB (k│d _n) when the uniform prior probability P (k), and P (k│d _n) by the maximum entropy method when a beta = 1 is Will be equal. Therefore, it can be said that this is a natural extension of the NB model. By the maximum entropy method, it becomes possible to appropriately estimate the posterior probability P _NB (k | d _n ) including the probability of not belonging to any existing category under a preferable condition of maximizing entropy.

Further, as a feature for estimating the posterior probability that the nth document does not belong to any existing category, the log likelihood with 3 sigma in the next model, that is, the 3 sigma value is used. This is a model (document generation model) of a category (likelihood c _n = argmax _k P _NB ( _k | d _n ) with the highest posterior probability of the nth document in the NB model, where the average μ and the standard deviation σ looking at a normal distribution, the probability that the random variable in accordance with the normal distribution enters the section (mu + 3 [sigma] and mu-3 [sigma]) is a 0.9 97. Thus, 3-sigma value, feature for existing categories Documents that are relatively higher than are considered to have a lower probability of belonging to an existing category.
Therefore, in order to obtain a 3-sigma value, a random variable X that satisfies the following equation (8) is considered.

In Expression (8), the average value of X when X is tried M times asymptotically approaches a normal distribution as M increases by the central limit theorem. That is, the following expressions (9) to (11) are obtained. In equation (9), X _i represents the value of X of the i-th trial, and N (μ, σ) represents a normal distribution with mean μ and standard deviation σ.

Next, the feature f (d _n , K + 1) is expressed by the following equation (12). The feature f (d _n , K + 1) is for obtaining the probability of the nth document not belonging to the existing category. The feature f (d _n , K + 1) has a 3 sigma value represented by Expression (12). In equation (12), M _n represents the number of words in the nth document.

As described above, the visualization device 10 estimates the unknown parameter β after determining the feature by the equations (1) to (12). Then, the posterior probability P _NB (k | d _n ) is obtained. A vector obtained from the obtained posterior probability P _NB (k | d _n ) of the nth document is referred to as a posterior probability vector. The posterior probability vector q _n is represented by q _n = (P (1 | d _n ),..., P (K + 1 | d _n )).

  In addition, the posterior probability vector estimation process is described in the non-patent literature `` K. Nigam, J. Lafferty and A. McCallum, Using maximum entropy for text classification, In IJCAI-99 Workshop on Machine Learning for Information Filtering, 61-67, 1999. It is assumed to be performed by a known method disclosed in the above.

[Visualization processing by posterior probability preservation embedding method]
Further, in the visualization device 10, the visualization unit 133 performs document group visualization processing using functions from the following equations (13) to (16) based on the posterior probability vector q _n .
First, the visualization space will be described. The visualization space realizes visualization of a document group, and here, the Euclidean space is used.
The coordinates in the visualization space that is the center of the document group belonging to the k-th category is φ _k = (φ _k1 ,..., Φ _kD ). D represents the dimension of the visualization space. Further, the coordinates of the n documents _{r n = (r n1, ···} , r nD) and.

Consider a case where the relationship between posterior probabilities is embedded in a low-dimensional, that is, two- or three-dimensional visualization space so that the user can intuitively grasp the relationship between documents. The relationship between the posterior probabilities embedded in the visualization space is a state in which the relationship of the posterior probabilities P _NB (k | d _n ) obtained by the above formulas (1) to (12) is maintained.
Therefore, the Euclidean distance u _nk of the formula (13), so as to satisfy the following conditions, so as to place the φ and r _n. In the visualization space, if the probability of belonging to the kth category of the nth document increases, the Euclidean distance _unk between the kth category and the nth document is made smaller (near). On the other hand, if the probability that the nth document belongs to the kth category decreases, the Euclidean distance _unk between the kth category and the nth document is increased (distant).

Here, consider the probability s _nk = ρ (u _nk ) that the nth document belongs to the kth category for the visualization space described above. ρ (u _nk ) is a monotonic phenomenon function regarding u> 0 (including 0 in this case). Here, 0 <ρ (u _nk ) <1 (including 0 and 1 here), ρ (0) = 1, ρ (infinity) = 1, Σρ (u _nk ) = 1 (k = 1, ..., K + 1). Then, if the Euclidean distance u _nk decreases, the probability s _nk approaches 1; conversely, if u _nk increases, the probability s _nk gradually approaches 0. In this way, the Euclidean distance _unk between each document and each category is adjusted in the visualization space according to the probability _snk belonging to the registered category.
A typical example of such ρ (u _nk ) is shown in the following formula (14). In the present embodiment, the function of Expression (14) is used.

(14)

Here, if a vector representing which category the nth document belongs to is s _n = (s _n1 ,..., Sn _{(k + 1)} ), in the visualization space, s _n is a posterior probability vector q. That is, it should be an approximation of _n . Therefore, q _n and s _{n are} considered as discrete probability distributions, and the distance between these two probability distributions is minimized for all documents. As a result, s _n approximates q _n . The Cullback distance KL (q _n , s _n ) representing the distance between the two probability distributions is shown in the following equation (15).

  In this case, an objective function (which should be maximized here) Lv for obtaining the coordinates R of all documents is represented by the following equation (16).

In equation (16), in _{fact, R = [r 1, ...} , r N] L v and _{Φ = [φ 1, ...,} φ K] relating By repeated until L _v and converge about the R Will be asked. Given Φ, L _v will be strictly convex upward with respect to R and will have the favorable property that the convergence value guarantees a global optimal solution. Coordinates obtained by compressing to low dimensions (that is, two-dimensional or three-dimensional) with MDS based on Θ (estimated) = [θ ₁ (estimated),..., Θ _k (estimated)] as the initial value of Φ Will be used. This is because a value when categories having similar appearance probabilities of words in the document are arranged close to the visualization space is used as an initial value. This makes it possible to visualize the relationship between categories. An explanatory diagram showing this visualization technique is shown in FIG.

FIGS. 2A and 2B show the case of a document group (cluster) classified into five categories of “sumo”, “baseball”, “rugby”, “soccer”, and “professional wrestling”. Yes. In FIGS. 2A and 2B, the document is actually visualized by hue in category units. For example, documents classified as “sumo”, “baseball”, “rugby”, “soccer”, and “professional wrestling” are color-coded in blue, red, green, light blue, and yellow, respectively.
Then, when the visualization processing is performed on the document group shown in FIG. 2A, the similarity between categories is reflected and the document group is arranged as shown in FIG. In FIG. 2B, document groups belonging to two categories of “sumo” and “professional wrestling” are arranged adjacent to each other. This is because both “sumo” and “professional wrestling” belong to the same sports genre as martial arts, and thus it is determined that the similarity between the two categories is high.

Here, the “classification error”, “multiple category”, and “singular document” shown in FIG.
A “classification mistake” is a document that is misclassified. Such a document is arranged around a document group of a category different from the registered category. In FIG. 2B, the “pro-wrestling” document is shown as a misclassification, and it is arranged around the “sumo” document group. Therefore, it is highly likely that the document belongs to the “sumo” category, not “pro-wrestling”.

  A “multiple category” is a document that is considered to be a document belonging to multiple categories. Such a document is arranged between a plurality of categories by visualization. In FIG. 2B, documents arranged between “soccer” and “baseball” are shown as multiple categories. Therefore, it is highly likely that this document belongs not only to “baseball” but also to “soccer”.

  A “singular document” is a document that has a high probability of not belonging to a registered category. Such a document is arranged in a place different from a place (area) where classified document groups gather in the visualization space. In FIG. 2B, the document “baseball” is arranged at a location different from the document group of the five types of categories. Such unique documents can be classified into the following two cases based on their contents. One is a case where useful contents not included in a document belonging to a registered category are described. The other is a case where contents irrelevant to the classification system or contents not worth document management (nonsense contents) are described. In the former case, since useful contents are described, it is significant for the user to find a specific document.

Next, processing of the entire operation from the above-described document generation model construction processing to visualization processing will be described.
FIG. 3 is a diagram illustrating an operation procedure of the visualization device 10. Note that the operation of the visualization device 10 is realized by the processing device 13 sequentially executing the visualization program stored in the storage device 12. The visualization program may be read from a computer-readable recording medium. Examples of the recording medium include a CD-ROM, a semiconductor memory, and a magnetic disk.

  First, the user operates the user terminal 30 to access the visualization device 10 via the communication network 20. Thereafter, in the processing device 13 of the visualization device 10, the document generation model construction unit 131 receives a request for a document group (categorized data) to be processed from the user terminal 30. Then, the document generation model construction unit 131 performs the construction processing of the document generation model for each category for the requested document group using the functions from the above formulas (1) to (3) (S1: "Building function"). Thereby, the word frequency of the word which appears in a document will be calculated | required for every document.

Subsequently, the posterior probability vector estimation unit 132 reads out the document generation model for each category constructed in S <b > 1 from the storage device 12. Then, the posterior probability vector estimation unit 132 performs an estimation process of the posterior probability vector q _n based on the read document generation model using the functions from Equation (4) to Equation (12) (S2: This) Is called "estimation function"). This will be described in detail. The posterior probability vector estimation unit 132 has the highest posterior probability P _NB (k | d _n ) of the document based on the posterior probability vector q _n according to the maximum entropy condition (equation (4)) read from the storage device 12. A posteriori probability vector q _n is estimated using a category document generation model. As a result, for each document, the probability of belonging to the registered category and the probability of not belonging to the registered category can be displayed as a vector.

Next, the visualization unit 133 uses the functions from Equation (13) to Equation (16) described above, and based on the posterior probability vector q _n estimated in S2, each document in the document group requested in S1. (S3: This is called “visualization function”). This will be described in detail. The visualization unit 133 adjusts and arranges the distance between each document and each category on the visualization space according to the probability belonging to the registered category represented by the posterior probability vector q _n estimated in S2. As a result, the similarity between the document and the category can be reflected in the visualization space.

  And the visualization part 133 transmits the visualization result of S3 to the terminal 30 for users via a communication network as an output process of a visualization result (S4: This is called "output function."). Thereby, for example, in the user terminal 30, the visualization space reflecting the relationship between the registered categories is displayed on the display device (see FIG. 2). Therefore, the user can discover new knowledge about the document group such as misclassification, multiple categories, unique documents, and the like. From the above, it is possible to grasp the relationship between categorized documents more accurately. Details of specific knowledge obtained from misclassification, multiple categories, and unique documents will be described later.

[Evaluation of visualization of the present invention]
Next, in order to evaluate the effectiveness of the present invention, webs classified into the top category of the Open Directory Project (ODP) Japanese web page (here, Open directory project, http://dmoz.org/). Visualization was performed using pages.
Here, sampling of web pages was performed as follows. First, web pages with 50 or fewer words were excluded from Japanese web pages registered in ODP. In addition, web pages classified into multiple categories were excluded. And 100 web pages for each category were randomly sampled. As a result of sampling, the obtained web pages were 1300 pages belonging to the following category.

  The categories used are arts, online-shop, computer, sports, news, business, recreation, health, various There are 13 types: reference, home, society, science, and regional. The game categories included in the ODP top category were excluded because the web page did not reach 100 pages.

[Visualization result]
FIG. 4 is a diagram illustrating a visualization result of the sampled web page group. In FIG. 4, web pages are color-coded by category. For example, the “arts” web page group is red, the “online-shop” web page group is blue, the “computer” web page group is pink, and the “sports” web page group is green.
In FIG. 4, web pages belonging to the same category form a cluster (a group of documents belonging to the same category). The total number of clusters is 14. This is because the number of clusters for 13 types is added to the number of clusters for 13 unique categories and the number of clusters for unique documents is 1.
Here, looking at the positional relationship of the clusters, the clusters of the related categories are arranged close to each other. For example, each cluster of “online-shop” and “business” and each cluster of “sports” and “health” are arranged close to each other.

And in these clusters, as shown in FIG. 4, there is a web page a different from the classified category. This is considered a misclassification. Specifically, the web page a shown in FIG. 4 is classified as “regional”, but is considered to have a high possibility of belonging to the “health” category. Then, when the content of the web page a was confirmed, it turned out that the web page a is a hospital page. If the local residents of the hospital classify the web page a from the viewpoint of using the web page a, it is true that the web page a is a “regional” page. However, many web pages belonging to the “health” category include hospital pages. Therefore, when considering the classification of the web page a in consideration of the substance of such classification, it seems that it is more appropriate to include the web page a in the “health” category.
Thus, the web page with the possibility of misclassification could be found through the visualization result of FIG.

In FIG. 4, the web page b is shown almost in the middle of each cluster of “sports” and “recreation”, although it is classified as “recreation”. The web page b was a soccer toto page. Certainly, the soccer lottery is a kind of recreation, but it is also closely related to sports. This was thought to be the reason why web page b was shown approximately halfway between the “sports” and “recreation” clusters. If you are interested in soccer, you may be interested in the web page b of the Circer Lottery.
In this way, it is possible to find a web page that has the possibility of multiple classification from the visualization result of FIG. Also, it is possible to find out what is difficult to classify into one specific category, judging from the content of the web page itself. Furthermore, when reviewing the category classification, it is possible to confirm what content should be focused on.

  Furthermore, it is possible to find a web page that has become a unique document from the visualization result of FIG. In FIG. 4, for example, a web page related to the introduction of a mail magazine corresponds to this. This web page was classified into the computer category. However, looking at the text used in the web page, it included a lot of other content, not just about the computer. Therefore, it turns out that the web page is not about the computer.

[Comparison with conventional methods]
Next, 1300 pages of web pages obtained by the above sampling were visualized by the MDS method (see Non-Patent Document 1) which is the conventional method described above. Here, two methods were used: a method of visualizing the word frequency vector by dimensional compression and a method of visualizing the posterior probability vector described above by dimensional compression.

  FIG. 5 is a diagram showing the result of visualizing the word frequency vector by dimensional compression using the MDS method. In FIG. 5, the clusters of each category are concentrated and the websites overlap with each other, so it is difficult to understand the similarity of web page groups. For this reason, the characteristics of the document are not known. This is because visualization was performed without reflecting the relationship of categories.

  FIG. 6 is a diagram showing the result of visualizing the posterior probability vector by dimensional compression by the MDS method. In FIG. 6, a cluster was formed for each category, and it was possible to find a classification error. However, the clusters were concentrated and it was difficult to find similar relationships among categories, multiple categories, and unique documents.

In this way, as a result of comparison with the conventional method, as shown in FIGS. 5 and 6, since the clusters are concentrated in any case, the web page group is compared with the case of FIG. 4 according to the present invention. It was difficult to understand the similar relationship of.
In the conventional method, since the visualization process is performed by comparing the documents without considering the category number K, when the number of documents is N, the calculation amount O is O (N ² ). In the present invention, since each document is compared with each category for visualization processing, the calculation amount O is O (NK). For this reason, in the conventional method, as the number of documents increases, the amount of calculation increases by a power of 2, whereas the present invention does not increase so much. Therefore, the visualization apparatus 10 can be easily applied to a large number of documents.

  The present invention is not limited to the embodiment described above. The data structure of the storage device 12 and the order of program processing can be variously changed by a known technique.

  Moreover, although the visualization apparatus 10 demonstrated the case where it comprised with a server apparatus, you may make it comprise with computers, such as a single personal computer, for example.

It is a block diagram which shows the system containing the visualization apparatus which concerns on embodiment of this invention. It is explanatory drawing which shows the method of visualization by the visualization apparatus of FIG. It is a figure which shows the operation | movement procedure of the visualization method in the visualization apparatus of FIG. It is a figure which shows the example of the result visualized by the visualization apparatus of FIG. It is a figure which shows the result of having dimensionally compressed and visualized the word frequency vector by MDS method. It is a figure which shows the result of having dimension-compressed and visualized the posterior probability vector by MDS method.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Visualization device 12 Storage device 13 Processing device 20 Communication network 30 User terminal

Claims (8)

For document groups classified into pre-registered categories, a document generation model representing the word frequency distribution of words appearing in the documents for each category, and the relationship between each document in the document group and each category visually A storage device for storing information of a two-dimensional or three-dimensional visualization space for expressive representation,
For each document in the document group, each probability of belonging to a registered category;
A probability that does not belong to any of the registered categories, and a posteriori probability estimator that creates a posteriori probability vector with components as components,
A visualization unit for performing visualization processing of the document group;
A method for visualizing the relationship between documents by a visualization device comprising:
The posterior probability estimation unit
Reads the document generation model for each of the categories from the storage device, for each document d _n in the documents,
The document d _n feature function f such correlation is larger the higher the model parameters word frequency and category k of (d _{n, k),} and represents the relationship between the document d _n and registered categories other than Category Using the feature function f (d _n , K + 1) of the following equation,
μ: average of the parameters of the _{c n}
sigma: standard deviation of parameters _{c n}
M _n : number of words in document d _n
c _n : The document generation model of the category having the highest posterior probability of the document d _n satisfies the constraint of the following expression,
N: Total number of documents
c _n: | as the entropy of the _{(d n} k) is the maximum, the probability belonging to each registered category k of the document _{d n} P document _{d n} category probability is given to P (k _{| d} n) , and it does not belong to any of the registered category probabilities P | a step of calculating the (K + 1 d _n), to create a posterior probability vector q _n of a document d _n to the probability component the calculated,
The visualization unit is
Function ρ the distance _{u nk} between center coordinates phi _k coordinates _{r n} and category k of the document _{d n} document _{d n} into a probability _{s nk} in category k, greater than zero probability _{s nk} and the distance _{u nk} monotonically decreases becomes the sum of k from 1 to K + 1 is using a function ρ becomes 1, the distance between the vector s _n obtained from the probability s _nk and the posterior probability vector q _n of each document d _n with respect to For an objective function L _v that minimizes the sum total in all documents of KL _n , L _v related to the coordinate set R of the document group having the coordinate r _n as an element and a category group having the central coordinate φ _k as an element and determining the set of coordinates R and center coordinates set Φ by repeating relates center coordinate set Φ to L _v and converge,
Externally outputting information of a visualization space in which the coordinate set R of the document group and the central coordinate set Φ of the category group are arranged;
A method for visualizing the relationship between documents, characterized in that Feature function f (d _{n, k)} representing the relationship between each category k of the document d _n and registered is represented by the following formula,
x _nj: word in the document _{d n} frequency
θ _kj : Parameter of document model of category k The probability P (k | d _n ) and the probability P (K + 1 | d _n ) are calculated by the parameter β shown by the following equation:
The inter-document relationship visualization method according to claim 1, wherein the inter-document relationship visualization method is obtained by estimating. The function ρ is expressed by the following equation:
Before Symbol objective function L _v is shown by the following formula
The method for visualizing the relationship between documents according to claim 1 or 2. For a group of pre-registered categorized documents, a document generation model representing the word frequency distribution of words that appear in the document for each category, and the relationship between each document in the document group and each category visually A storage device for storing information of a two-dimensional or three-dimensional visualization space for expressing
For each document in the document group, each probability of belonging to a registered category;
A probability that does not belong to any of the registered categories, and a posteriori probability estimator that creates a posteriori probability vector with components as components,
A visualization unit that performs a visualization process of the document group,
The posterior probability estimator is
Reads the document generation model for each of the categories from the storage device, for each document d _n in the documents,
The document d _n feature function f such correlation is larger the higher the model parameters word frequency and category k of (d _{n, k),} and represents the relationship between the document d _n and registered categories other than Category Using the feature function f (d _n , K + 1) of the following equation,
μ: average of the parameters of the _{c n}
sigma: standard deviation of parameters _{c n}
M _n : number of words in document d _n
c _n : The document generation model of the category having the highest posterior probability of the document d _n satisfies the constraint of the following expression,
N: Total number of documents
c _n: | as the entropy of the _{(d n} k) is the maximum, the probability belonging to each registered category k of the document _{d n} P document _{d n} category probability is given to P (k _{| d} n) and the probability P that does not belong to any of the registered category (K + 1 | d _n) is calculated, with the estimation function for creating a posterior probability vector q _n of a document d _n to the probability component the calculated,
The visualization unit includes:
Function ρ the distance _{u nk} between center coordinates phi _k coordinates _{r n} and category k of the document _{d n} document _{d n} into a probability _{s nk} in category k, greater than zero probability _{s nk} and the distance _{u nk} monotonically decreases becomes the sum of k from 1 to K + 1 is using a function ρ becomes 1, the distance between the vector s _n obtained from the probability s _nk and the posterior probability vector q _n of each document d _n with respect to For an objective function L _v that minimizes the sum total in all documents of KL _n , L _v related to the coordinate set R of the document group having the coordinate r _n as an element and a category group having the central coordinate φ _k as an element and visualization function of obtaining the coordinate set R and center coordinates set Φ by repeating until the L _v with respect to the center coordinates set Φ converges,
A visualization apparatus comprising: an output function for externally outputting information on a visualization space in which the coordinate set R of the document group and the central coordinate set Φ of the category group are arranged. Feature function f (d _{n, k)} representing the relationship between each category k of the document d _n and registered is represented by the following formula,
x _nj: word in the document _{d n} frequency
θ _kj : Parameter of document model of category k The probability P (k | d _n ) and the probability P (K + 1 | d _n ) are calculated by the parameter β shown by the following equation:
The visualization device according to claim 4, wherein the visualization device is obtained by estimating. The function ρ is expressed by the following equation:
Before Symbol objective function L _v is shown by the following formula
The visualization apparatus according to claim 4 or 5, wherein   A visualization program for causing a computer to execute the inter-document relationship visualization method according to any one of claims 1 to 3. A computer-readable recording medium recording a visualization program for causing a computer to execute the inter-document relationship visualization method according to any one of claims 1 to 3.