JP2010267017A - Device, method and program for classifying document - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To perform association between documents having similar contents while reflecting a change of interest of an author, about a lot of documents. <P>SOLUTION: A document classification device 1000 includes: a storage part 2 storing a prescribed calculation expression and a prescribed calculation end condition used in the plurality of documents and a probability distribution model; an initial setting part 3 randomly setting initial values of document classes to which the respective documents belong and topic classes to which words belong; a document class evaluation part 4 estimating the document class to which the document belongs in each document; a topic class evaluation part 5 estimating the topic class to which the word belongs in each word; a convergence decision part 6 making the document class evaluation part 4 and the topic class evaluation part 5 repeat the estimation until satisfying the prescribed calculation end condition; and an output part 7 outputting a calculation result including contents of the document class. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention associates a large amount of documents (electronic data such as Web pages, papers, patent specifications, blog articles, etc.) between documents with similar contents and between authors who created documents with similar contents. Regarding technology.

  In general, documents to be searched include metadata such as authors, their affiliations, cited references, and creation times. There is known a method of using information such as metadata to associate between documents having similar contents and to associate between authors who created documents having similar contents.

  The Author-Topic model disclosed in Non-Patent Document 1 is a technique for associating an author with a topic by giving each author a possibility of handling (selecting) each topic as a probability. Here, a probability distribution for selecting a topic for each author is called a topic selection probability distribution. That is, in the topic selection probability distribution calculated for a certain author, association is realized by assuming that the topic having the highest probability is the topic most relevant to the author.

  In addition, the Author-Persona-Topic model disclosed in Non-Patent Document 2 takes into account multiple personalities of personas for each author and introduces a topic selection probability distribution for each personality to allow multiple authors to This is a technique that makes it possible to have a topic selection probability distribution.

M. Steyvers, P. Smyth, M. Rosen-Zvi, and T. Griffiths: “Probabilistic Author-Topic Models for Information Discovery”, ACM KDD (International conference on Knowledge Discovery and Data mining), pp. 306-315 (2004 ). D. Mimno and A. McCallum: “Expertise Modeling for Matching Papers with Reviewers”, ACM KDD (International conference on Knowledge Discovery and Data mining), pp. 500-509 (2007).

  However, in the method of Non-Patent Document 1, the topic selection probability distribution is fixed to one for one author. Therefore, there is a problem that the change of the author's interest cannot be reflected.

  In the method of Non-Patent Document 2, a plurality of topic selection probability distributions can be provided for each author. However, since a persona is set closed for each author, personas can be associated between different authors. Can not. In other words, if there is an unconnected author (an author who is not in a co-authorship or friendship) who creates a sentence with similar content, the persona is set on a different scale (different keyword) for each author. As a result, it is impossible to determine whether the author's personas are similar. Therefore, there is a problem that it is not possible to make associations between authors who create documents having similar contents and are not connected at all.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to associate documents having similar contents with respect to a large number of documents while reflecting a change in interest of the author. Another object is to associate the authors who created documents with similar contents.

  In order to solve the above problems, the present invention probabilistically assigns a plurality of documents, which are electronic data, to each of a plurality of document classes, which are a set of documents having similar contents. Each word constituting the document is stochastically assigned to one or more of a plurality of topic classes that are word sets having similar contents, each of the document authors is assigned to one or more document classes, and each document class is assigned 1 A document classification device that performs calculation based on a predetermined probability distribution model associated with one or more topic classes, and is used to estimate a plurality of documents and a document class to which each document belongs in the probability distribution model. The first calculation formula, the second calculation formula used to estimate the topic class to which each word belongs and used in the probability distribution model, and a predetermined calculation end condition are stored. An initial setting unit that randomly sets initial values of a memory unit, a document class to which each document belongs, and a topic class to which each word belongs, and all document classes for each document based on the first calculation formula For each word, calculate the probability of belonging to each, and estimate the document class with the highest probability as the document class to which the document and the author belong, and the second calculation formula , Calculate the probability of belonging to each topic class, and estimate the topic class with the highest probability as the topic class to which the word belongs, and the document class until a predetermined calculation end condition is satisfied. A convergence determination unit that causes the evaluation unit and the topic class evaluation unit to repeat estimation and a convergence determination unit satisfy a predetermined calculation termination condition. When it is determined that, characterized in that it comprises an output unit for outputting a calculation result including the document and the author belongs to every document class.

  According to this invention, each document class can be associated with one or more topic classes, while each document author belongs to one or more document classes. In other words, since one author can be indirectly associated with one or more topic classes, if one author is associated with multiple topic classes, the change in the interests of the author As a result, the documents having similar contents can be associated with each other (each document is classified into a plurality of document classes).

  Further, the present invention probabilistically assigns each of a plurality of documents to a plurality of document classes that are a set of documents having similar contents, and each word constituting each document is assigned to a plurality of documents that are electronic data. Probably belong to one or more of multiple topic classes that are similar word sets, and each of the document authors stochastically belongs to one of the author classes that are the set of authors who created similar documents A document classification apparatus that performs calculation based on a predetermined probability distribution model that probabilistically belongs an author class to one or more document classes and associates each document class with one or more topic classes. A first calculation formula for estimating the document class used in the probability distribution model and the document class to which each document belongs, and the word used in the probability distribution model and belonging to each word. A storage unit for storing a second calculation formula for estimating a class, a third calculation formula for estimating the author class to be used for each author and belonging to the author, and a predetermined calculation end condition An initial setting unit that randomly sets initial values of an author class to which each author belongs, a document class to which each document belongs, and a topic class to which each word belongs, and for each author based on the third calculation formula The author class evaluator for calculating the probability of belonging to each of the author classes and estimating the author class with the highest probability as the author class to which the author belongs, and the document based on the first calculation formula For each document class, the probability of belonging to each document class is calculated, and the document class with the highest probability is set as the document class to which the document and its author class belong. Based on the document class evaluation unit to be estimated and the second calculation formula, the probability of belonging to each of all topic classes is calculated for each word, and the topic class having the highest probability is determined as the topic class to which the word belongs. A topic class evaluation unit that estimates, a convergence determination unit that causes the author class evaluation unit, document class evaluation unit, and topic class evaluation unit to repeat estimation until a predetermined calculation termination condition is satisfied, and a convergence determination unit And an output unit that outputs a calculation result including the content of the author class and the content of the document class when it is determined that the calculation end condition is satisfied.

  According to this invention, in addition to the above invention, each author of a document can belong to any author class, and each author class can be associated with one or more document classes. That is, an author who has created a document having similar contents can be associated with the author who has created a document having similar contents by belonging to the same author class.

  Further, according to the present invention, the convergence determination unit has an error between the previous probability distribution and the latest probability distribution calculated by at least one of the author class evaluation unit, the document class evaluation unit, and the topic class evaluation unit as a predetermined calculation end condition. It is characterized in that it is less than a predetermined threshold value or that the number of repeated calculations reaches a predetermined number.

  According to this invention, as a predetermined calculation termination condition, by setting a specific and appropriate condition such as the threshold value related to the error of the probability distribution described above or the number of repeated calculations, the calculation accuracy, processing time, etc. Conditions can be set according to the purpose.

The present invention is also a document classification program for causing a computer to function as the document classification apparatus described above.
According to this invention, a computer in which this program is installed can realize each function based on this program.

  According to the present invention, it is possible to associate documents having similar contents with respect to a large number of documents while reflecting a change in interest of the author. In addition, it is possible to perform association between authors who created documents having similar contents.

It is explanatory drawing of the document and topic classification | category model in 1st Embodiment. It is a block diagram of the document classification device in 1st Embodiment. It is a block diagram of a document class evaluation part. It is a block diagram of a topic class evaluation part. It is a flowchart which shows the process of the document and topic classification | category method in 1st Embodiment. It is a flowchart which shows a document class process. It is a flowchart which shows a document class sampling process. It is a flowchart which shows a topic class update process. It is a flowchart which shows a topic class sampling process. It is explanatory drawing of the author / document / topic classification model in 2nd Embodiment. It is a block diagram of the document classification device in 2nd Embodiment. It is a flowchart which shows the process of the document and topic classification | category method which considered the author class in 2nd Embodiment. It is a flowchart which shows an author class update process. It is a figure which shows distribution of an author and a document class in an Example. It is a figure which shows distribution of the author class and document class in an Example. It is a figure which shows distribution of the document and topic class in an Example. It is a figure which shows distribution of the document class and topic class in an Example.

  Hereinafter, modes for carrying out the present invention (hereinafter referred to as “embodiments”) will be described by dividing them into a first embodiment and a second embodiment with reference to the drawings. In the following, a token corresponds to a concept obtained by abstracting words. That is, the document is composed of a sequence of tokens, and a specific word is entered into the token to become a specific document. In the description of the present embodiment, “evaluation” means “estimation”.

<First Embodiment>
In the first embodiment, a set of documents having similar contents is set as a document class, a set of similar words (tokens) is set as a topic class, and each document is evaluated by evaluating the document class and topic class to which each author belongs. Make associations between content documents. Since it can be determined that documents belonging to the same document class are highly related (similar contents are handled), association between documents having similar contents is realized by classifying each document and each author into a document class. be able to.

  The document class is, for example, “sports” or “politics”. The topic class associated with the document class “sports” is, for example, “baseball”, “soccer”, and the like. The words belonging to the topic class “baseball” are, for example, “professional baseball”, “major league”, and the like.

Specifically, first, the probability _P document class _{c d} Field of document d is _{j (c d = j | c} \d, z, γ, δ) is calculated. Here, c _\D is a set of document classes excluding the document d, z is the set of topics, the γ and δ is a hidden parameter (described in detail later). For each document d, probabilities are calculated for all document classes, and the document class with the highest probability is associated as the document class to which the document belongs.

Similarly, the probability P (z _di = t | j, z _{\ di} , w, β, δ) that the topic class related to the i-th token di of the document d is t is calculated. Here, z _di represents the topic class to which the token di belongs (hereinafter also referred to as a token (word) z _di ), z _{\ di} is a set of topics excluding z _di from z, and w is a set of words. And β is a hidden parameter (details will be described later). For each word, the probabilities are calculated for all topic classes, and the topic class with the highest probability is associated as the topic class to which the word belongs. Since a topic class is associated with each word in a document, a plurality of topic classes may be associated with one document.

  In the first embodiment, a method (document / topic classification) that simultaneously performs association between a document and a topic (association between a document and a topic handled by the document) and association between documents (association between documents having similar contents). Method: LIT0) will be described.

  First, the document / topic classification model in the first embodiment will be described with reference to FIG. In the first embodiment, the following a. ~ E. The association is performed using a model that assumes

a. One document is stochastically associated with one document class.
b. A word (token) (in the document) is associated with a topic class. Different words may be associated with the same topic class.
c. One document class can select one or more topic classes. A plurality of topic classes can be associated with one document class.
d. A topic class can select one or more tokens. Since a plurality of words (tokens) may be associated with one topic class, different words (tokens) may be selected for the same topic class (for example, words w ₁ and w in FIG. 1). ₃ ).
e. Each document class has a different probability distribution for topic class selection. For example, in the example of FIG. 1, even if the document class j ₁ and the document class j ₂ are both have selected topic class t _4, the probability that the document class j ₁ selects the topic class t _4, the document class j ₂ is not necessarily the same probability of selecting a topic class t _4.

(Constitution)
Next, the configuration of the document classification device according to the first embodiment will be described. As shown in FIG. 2, the document classification device 1000 is a computer device, and includes an input unit 1, a storage unit 2, a calculation unit 100, and an output unit 7.

  The input unit 1 is a means for inputting information, and includes, for example, a keyboard, a mouse, a disk drive device, and the like.

  The storage unit 2 is a means for storing information. For example, the storage unit 2 includes a general hard disk device and the like, and includes an operation program, document data to be associated (including metadata such as author information), and the like. Information such as parameters and equations is stored. As described above, the document data is electronic data such as Web pages, papers, patent specifications, blog articles, and the like.

The arithmetic unit 100 is a main control device composed of, for example, a CPU (Central Processing Unit) and a RAM (Random Access Memory), and includes an initial setting unit 3, a document class evaluation unit 4, a topic class evaluation unit 5, And a convergence determination unit 6.
The document class evaluation unit 4 includes a sample author selection unit 41, a sample document selection unit 42, an initial setting unit 43, and a document class update unit 44 (details will be described later).
The topic class evaluation unit 5 includes a sample author selection unit 51, a sample document selection unit 52, an initial setting unit 53, and a topic class update unit 54 (details will be described later).

  The output unit 7 is a means for outputting information, for example, a graphic board (output interface) and a monitor connected thereto.

(processing)
Next, processing of the document classification apparatus 1000 will be described with reference to FIG. Note that the step numbers in FIG. 5 and the reference numerals in FIG. 2 are matched (for example, step S1 and input unit 1, step S3 and initial setting unit 3, etc., and so on in FIGS. 12 and 11).

[Input unit 1]
First, the input unit 1 accepts input of the number J of document classes to be classified and the number T of topic classes to be classified by the user (step S1).

[Initial setting part 3]
Next, the initial setting unit 3 randomly sets (selects) the document class to which each document belongs and the initial value of the topic class (step S3). These initial values can be arbitrarily selected, but there is one document class for one document. For each word in each document, one topic class to which the word belongs is set.

[Document Class Evaluation Unit 4]
Next, the document class evaluation unit 4 performs document class update processing (step S4). Specifically, the probability P (c _d = j | c) that the document class to which the document d belongs is j (j = 1, 2, 3,..., J) using the Gibbs sampling method. _{\ D} , z, γ, δ) are evaluated. In Gibbs sampling, select one document to be sampled, calculate the conditional probability (the following formula (3c)), and then compare the randomly generated value (random number) with the conditional probability value. The process of updating the document class is repeated until convergence (determined in step S6).

Here, the document class update processing in step S4 will be described with reference to FIG.
First, the sample author selection unit 41 selects one author a that is not yet selected from the set A of authors (step S41).
Next, the sample document selection unit 42 selects one sample target document d that has not yet been selected from the documents created by the author a (step S42).

In the subsequent processing, the probability P (c _d = j | c _{\ d} , z, γ, δ) for all document classes j (j = 1, 2, 3,..., J) is selected for the selected document d. evaluate.

  First, in the initial setting unit 43, initial values of parameters used in the subsequent probability calculation are set (step S43). For example, the total number of documents stored in the storage unit 2 is set as the value of the parameter n. Also, initial values of the hidden parameters β, γ, and δ used for document class evaluation are determined. Here, γ is a vector of J dimension (J is the total number of document classes input by the input unit 1), and δ is a vector of T dimension (T is the total number of topic classes input by the input unit 1). As the values of the elements γ and δ, arbitrary values satisfying the following relational expressions (1) and (2) are set.

_{0 <γ j ≦ n aj\d j} = 1,2,3, ..., J ··· formula (1)
0 <δ _t ≦ n _jt \ _d t = 1, 2, 3,..., T (2)
Here, n _{aj \} d is the total number of documents created by the author a, excluding the document d to be sampled, out of all documents belonging to the document class j. Also, n _{jt \} d is the number of selected topic classes t in all documents belonging to the document class j excluding the document d to be sampled.

  Next, the document class update unit 44 performs document class sampling processing (step S44). Here, the document class sampling process in step S44 will be described with reference to FIG.

  As shown in FIG. 7, the document class updating unit 44 first selects one document class j from the J document classes (step S441), and performs the following processing.

The probability P (c _d = j | c _{\ d} , z, γ, δ) is calculated by the following equation (3c) (first calculation equation) (step S442). For this purpose, first, the following equation (3a) is calculated.

Next, for j = 1, 2,..., J, the following formula (3b) is calculated to calculate the probability of adding the probabilities of the document class having a smaller value of j than itself.

Then, for all j = 1, 2,..., J, normalization is performed according to equation (3c). As a result, all the probabilities P (c _d = j | c _{\ d} , z, γ, δ) are values of 1 or less.

Here, the symbol ∝ means that the value on the left side is proportional to the value on the right side. n _dt is the number of times the topic class t has been selected in the document d, and n _d is the total number of tokens included in the document d. gamma _j denotes the j dimension gamma, [delta] _t means the t dimension [delta].
The hidden parameters γ and δ set by the initial setting unit 3 are set so that the denominator of the equation (3a) becomes 0 when n _aj \ _d and n _{jt \ d} become 0 in the equation (3a). It plays the role of a smoothing parameter to avoid the probability being unable to be calculated.

After step S442, a uniform random number of 0 ≦ R <1 is generated (step S443), and the value of the probability P (c _d = j | c _{\ d} , z, γ, δ) calculated in step S442 is greater than R. If it is larger (Yes in step S444), the class to which the document d belongs is changed to j (step S445). Step probability _P calculated in _{S442 (c d = j | c} \d, z, γ, δ) when the value of the following R (No at step S444), skip document update class without (S445 Then, the process proceeds to step S446.

  If all the document classes have not been updated for the document d, that is, if there is a document class that has not been processed in step S442 and subsequent steps (No in step S446), the document class j ′ that has not been checked is selected. Select (Step S441) and repeat Steps S442 to S445. When the update process is completed for all the document classes j = 1, 2,..., J (Yes in step S446), the process returns to FIG.

  The document class evaluation unit 4 selects a document d ′ that has not been selected as a sample target from among the documents created by the author a as a new sample target document (No in step S45 in FIG. 6 → step S42), and step S43. , S44 is repeated. These processes are repeated for all documents created by the author a (until “Yes” in step S45).

  Next, when all the authors are not sampled, that is, when there is an author who has not finished sampling (No in step S46), the document class evaluation unit 4 returns to step S41 and has not yet selected. Choose the author a 'for which there is no sample. And the process of step S42-step S45 is repeated. When sampling has been completed for all the authors (Yes in S46), the process returns to FIG. 5, and step S4 is completed, and the process proceeds to step S5.

[Topic class evaluation unit 5]
After step S4, the topic class evaluation unit 5 performs topic class update processing (step S5). Specifically, the i-th token di in a document d belongs to the topic class t (t = 1, 2, 3,..., T) by a technique using Gibbs sampling as in the document class evaluation unit 4. The probability P (z _di = t | j, z _{\ di} , w, β, δ) is evaluated.

Here, the topic class update processing in step S5 will be described with reference to FIG.
First, the sample author selection unit 51 selects one sample target author a that has not yet been selected from the set A of authors (step S51).
Next, the sample document selection unit 52 selects one sample target document d that has not yet been selected from the documents created by the author a (step S52).

  Next, the initial setting unit 53 sets initial values of parameters used in subsequent probability calculations (step S53). For example, the total number of documents stored in the storage unit 2 is set as the value of the parameter n. In addition, initial values of hidden parameters β, γ, and δ used for topic class evaluation are determined. Here, γ and δ are the same as γ and δ in the document class evaluation unit 4, and arbitrary values satisfying the expressions (1) and (2) are set. Β is a vector of V dimensions (V is the number of different words (number of types) included in all documents), and the value of each element of β is set to an arbitrary value satisfying the following relational expression.

0 <β _v ≦ n _tv \ _di j = 1, 2, 3,..., J (Equation 4)
Here, n _tv \ _di is the total number of tokens that currently select the word v, excluding the i-th token of the document d to be sampled, among the tokens belonging to the topic class t.

  Next, the topic class update unit 54 performs topic class sampling processing (step S54). Specifically, the following processing is repeated for all tokens in the document d (all tokens in the document d excluding tokens whose probability has already been calculated).

Here, the topic class sampling process in step S54 will be described with reference to FIG.
First, the topic class update unit 54 selects one token (assumed to be the i-th word) di in the document d that has not been sampled yet (step S541).

Next, one topic class t that has not been sampled is selected (step S542).
Then, the following equation (5c) (second calculation formula), the probability _{P (z di = t | j} , z \di, w, β, δ) calculates the (step S543). For this purpose, first, the following equation (5a) is calculated.

Next, with respect to t = 1, 2,..., T, a probability obtained by adding the probabilities of topic classes having a smaller value of t than itself is calculated by calculating equation (5b).

Then, for all t = 1, 2,..., T, normalization is performed according to equation (5c). As a result, all the probabilities P (z _di = t | j, z _{\ di} , w, β, δ) are values of 1 or less.

Here, n _jt \ _di is the total number of tokens that currently select the topic class t other than the i-th token of the document d to be sampled among _all the documents belonging to the document class j. β _v means v of dimension β.

The hidden parameters β and δ set by the initial setting unit 53 have a probability that the denominator of the equation (5a) becomes 0 when n _tv \ _di or n _{jt \ di} becomes 0 in the equation (5a). It plays the role of a smoothing parameter to avoid being unable to calculate.

Next, a uniform random number of 0 ≦ R <1 is generated (step S544), and the value of the probability P (z _di = t | j, z _{\ di} , w, β, δ) calculated in step S543 is greater than R. is greater (in step S545 Yes), the topic class that belongs to the i-th token _{z di} of the document d to change to t (step S546). Probability calculated in step _{S543 P (z di = t |} j, z \di, w, β, δ) in the following cases R is value (No at step S545), updates the topic class is not a performed (S546 Is skipped), and the process proceeds to step S547.

If all topic classes have not been updated in step S547, that is, if there is a topic class for which the probability has not yet been calculated for word di (No in step S547), that topic class is selected (step S542), by repeating the steps S543~ step S546, the probability for all the topic class _{t P (z di = t |} j, z \di, w, β, δ) is calculated.

When the sampling process is completed for all topic classes (Yes in step S547), a token for which the conditional probability has not yet been calculated is selected in the currently selected document d (No in step S548, step S541). ), Steps S542 to S547 are repeated. Thereby, the probability P (z _di = t | j, z _{\ di} , w, β, δ) belonging to each topic class can be calculated for all tokens (words) in the document d (t = 1, 2, ..., T). After Yes in step S548, the process returns to FIG. 8, ends step S54, and proceeds to step S55.

  Next, when there is a document that is not selected as a sample target among the documents created by the author a (No in step S55), the topic class evaluation unit 5 creates a new document d ′ that has not yet been selected. A document to be sampled is selected (step S52), and steps S53 and S54 are repeated. This is performed for all the documents created by the author a (until Yes in step S55).

  If there is an author who has not finished sampling after Yes in step S55 (No in S56), an author a 'which has not yet finished sampling is selected (step S51), and similarly, steps S52 to S55 are selected. repeat. When sampling has been completed for all the authors (Yes in step S56), the process returns to FIG. 5 to end step S5 and proceed to step S6.

[Convergence determination unit 6]
The convergence determination unit 6 determines whether or not the probability distributions of the equations (3c) and (5c) calculated by the document class evaluation unit 4 and the topic class evaluation unit 5 have converged (step S6), and these probability distributions. Are converged (during No in step S6), the processing of the document class evaluation unit 4 and the topic class evaluation unit 5 (steps S4 and S5) is repeated.

  When the probability distribution converges (Yes in step S6), the obtained probability distribution of the document class and the topic class is output to the output unit 7. Here, the determination of whether or not the probability distribution has converged is made by comparing the error between the probability distribution one step before and the probability distribution obtained this time, and when the error is equal to or less than a predetermined threshold value It may be determined. Alternatively, the number of repetitions of the document class update process (step S4) and the topic class update process (step S5) is counted, and if the preset number of repetitions is reached, it is determined that the probability distribution has converged and the process is terminated. good. Thus, the calculation end condition may be set according to purposes such as calculation accuracy and processing time.

[Output unit 7]
After Yes in step S6, the output unit 7 outputs the document class and the topic class based on the probability distribution obtained from the convergence determination unit 6 (step S7).
At this time, the document class outputs j having the largest probability P (c _d = j | c _{\ d} , z, γ, δ) evaluated for the document d as the document class to which the document d belongs.

In addition, at this time, topic class, document the i-th token in the d (word) probability was calculated for _{di P (z di = t |} j, z \di, w, β, δ) the highest becomes t The topic class z _di to which the word di belongs, and the probability P (z _di = t | j, z _{\ di} , w, β, δ) for all the words (i = 1, 2,..., N _d ) in the document d. A set obtained by selecting t with the highest value is output as a topic class selected by the document d.

  Thus, according to the document classification apparatus 1000 of the first embodiment, each document class can be associated with one or more topic classes while each document author is associated with one or more document classes. In other words, since one author can be indirectly associated with one or more topic classes, if one author is associated with multiple topic classes, the change in the interests of the author As a result, the documents having similar contents can be associated with each other (each document is classified into a plurality of document classes).

<Second Embodiment>
Next, the document classification apparatus 1000a (1000) of the second embodiment will be described. In addition, the same code | symbol is attached | subjected about the element (a structure and a process) similar to the case of the document classification device 1000 of 1st Embodiment, and duplication description is abbreviate | omitted suitably.

  In the second embodiment, in addition to the contents of the first embodiment, an author class that creates a document having similar contents is an author class, the author belongs to the author class, and the author class selects a document class. As a result, it is possible to determine that authors belonging to the same author class are highly related (similar in interest).

Specifically, the probability author class (community) _{s a} Field of author a is _{h P (s a = h |} s \a, c, α, γ) is calculated. Here, s _{\ a} is a set of author classes excluding author a, c is a set of document classes, and α is a hidden parameter. For each author, probabilities are calculated for all author classes, and h with the highest probability is associated as the author class to which author a belongs.

  In the second embodiment, association between documents and topics (association between a document and a topic handled by the document), association between documents (association between documents having similar contents), and association between authors (similarity) A method (author / document / topic classification method: LIT1) for simultaneously performing association between authors who created a document with the content to be described will be described.

As shown in FIG. 10, in the model (author / document / topic classification model) for associating an author, a document, and a topic class in the second embodiment, a. ~ E. In addition to f. ~ H. Assuming
f. One author is stochastically associated with one author class.
g. Authors handling similar contents (a set of authors selecting the same document class) belong to the same author class (community).
h. One author class can select one or more document classes.

(Constitution)
Next, the configuration of the document classification apparatus 1000a according to the second embodiment will be described. As shown in FIG. 2, the document classification apparatus 1000a according to the second embodiment has an author class evaluation unit for evaluating the author class in the arithmetic unit 100a (100), as compared with the document classification apparatus 1000 according to the first embodiment. 8 is added. The author class evaluation unit 8 includes a sample author selection unit 81, an initial setting unit 82, and a document class update unit 83 (details will be described later).

(processing)
Next, processing of the document classification apparatus 1000a will be described with reference to FIG. First, the outline of the document / topic classification method taking account of the author class will be described. In addition to the document class update process (step S4) and the topic class update process (step S5) of the first embodiment, the class to which the author belongs. The author class update process (step S8) for evaluating is added, and iterative probability calculation and update process are performed as in the first embodiment. Thus, the probability P (s _a = h | s _{\ a} , c, α, γ) that the author class (community) to which an author a belongs is h (h = 1, 2, 3,..., H), The probability P (c _d = j | c _{\ d} , z, γ, δ) that the document class to which the document d belongs is j (j = 1, 2, 3,..., J), and the i th _Evaluate the probability P (z _di = t | j, z _{\ di} , w, β, δ) that the token belongs to (relates to) the topic class t (t = 1, 2, 3,..., T).

[Input unit 1]
First, the input unit 1 receives an input of the number H of author classes to be classified in addition to the number J of document classes to be classified and the number T of topic classes to be classified by the user (step S1a).

[Initial setting part 3]
Next, the initial setting unit 3 sets (selects) the author class to which each author belongs, the document class to which each document belongs, and the initial value of the topic class (step S3a). These initial values can be arbitrarily selected as in the first embodiment, but one document class and a plurality of topic classes can be selected for one document. Also, there is one author class to which one author belongs.

[Author class evaluation section 8]
Next, the author class evaluation unit 8 performs an author class update process (step S8). Here, the author class update processing in step S8 will be described with reference to FIG. Similar to the document class evaluation unit 4 and the topic class evaluation unit 5 of the first embodiment, the author class evaluation unit 8 uses the Gibbs sampling technique to set an author class (community) to which an author a belongs to h (h = 1, 2, 3,..., H), the probability P (s _a = h | s _{\ a} , c, α, γ) is evaluated.

First, the sample author selection unit 81 selects one author a that is not yet selected from the set A of authors (step S81).
Next, the initial setting unit 82 sets initial values of parameters used in subsequent probability calculations (step S82). For example, initial values of the hidden parameters α and γ used for evaluating the document class are determined. Here, as in the first embodiment, γ is an arbitrary value satisfying the above-described expression (1) with a vector of J dimensions (J is the total number of document classes input by the input unit 1). α is a vector of H dimension (H is the total number of author classes input by the input unit 1), and an arbitrary value satisfying the following relational expression (6) is set.
0 <α _h ≦ n h _{\ a} h = 1, 2, 3,..., H (formula 6)
_Here, n h\a is the number of documents belonging to the author class h except for the document author a created.

Next, the document class updating unit 83 selects one author class h that has not been selected yet (update processing has not been completed) (step S83), and performs the following update processing.
The probability P (s _a = h | s _{\ a} , c, α, γ) is calculated by the following formula (7c) (third calculation formula) (step S84). For this purpose, first, the following equation (7a) is calculated.

Next, with respect to h = 1, 2,..., H, the probability is calculated by adding the probabilities of topic classes having a smaller value of h than itself by calculating equation (7b).

Then, for all h = 1, 2,..., H, normalization is performed according to equation (7c). As a result, all the probabilities P (s _a = h | s _{\ a} , c, α, γ) are values of 1 or less.

Here, n _Hj\a among documents all authors belonging to the author class h created, the number of all documents belonging to the document class j at this time, except a document author a created, n _aj author a Is the number of documents belonging to the document class j, and _na is the number of documents created by the author a.

Next, the document class update unit 83 generates a uniform random number of 0 ≦ R <1 (step S85), and the probability P (s _a = h | s _{\ a} , c, α, γ) calculated in step S84. If the value of is greater than R (Yes in step S86), the class to which the author a belongs is updated to h (step S87). If the value of the probability P (s _a = h | s _{\ a} , c, α, γ) calculated in step S84 is R or less (No in step S86), the author class is not updated (S87 Skip) and go to step S88.

Steps S83 to S87 are repeated until all h (h = 1, 2,..., H) have been updated (until No in step S88). If all authors have not been sampled after Yes in step S88, that is, if there is an author a 'that has not yet been sampled (No in step S89), the process returns to step S81. Thereafter, the author a ′ that has not yet been sampled is selected (step S81), and the probabilities P (s _a = h | s _{\ a} , c, α, γ) are obtained for all authors by repeating steps S82 to S88. ) Is evaluated (Yes in step S89), the process returns to FIG. 12, and step S8 ends and the process proceeds to step S4. Steps S4 and S5 are the same as those in the first embodiment.

[Convergence determination unit 6]
As shown in FIG. 12, after step S5, the convergence determination unit 6 determines whether or not each probability distribution calculated in steps S8, S4, and S5 has converged (step S6a), and until the probability distribution converges. The process of steps S8, S4, and S5 is repeated (until Yes in step S6a). When each probability distribution converges, the process is terminated (Yes in step S6a), and the obtained probability distribution of the author class, document class, and topic class is output to the output unit 7. To determine whether each probability distribution has converged, the error between the probability distribution one step before and the probability distribution obtained this time is compared, and if the error is equal to or less than a predetermined threshold, it is determined that the probability distribution has converged. It's also good. Alternatively, when the number of repetitions set in advance is reached, it is possible to determine that the probability distribution has converged and end the processing.

[Output unit 7]
After Yes in step S6, the output unit 7 outputs the author class, document class, and topic class based on the probability distribution obtained from the convergence determination unit 6 (step S7a).
At this time, the author class outputs h having the largest probability P (s _a = h | s _{\ a} , c, α, γ) evaluated for the author a as the author class to which the author belongs.
At this time, the document class and the topic class are as described in the output unit 7 of the first embodiment.

  As described above, according to the document classification apparatus 1000a of the second embodiment, in addition to the effects of the first embodiment, each author of a document belongs to one of the author classes, and one or more author classes are provided. Associated with the document class. That is, an author who has created a document having similar contents can be associated with the author who has created a document having similar contents by belonging to the same author class.

  That is, according to the document classification apparatus 1000a of the second embodiment, a set (document class) of documents related to similar topics or a document with similar contents is created from a large number of documents and data of authors who created the documents. A set of authors (author class) can be extracted. At that time, the name of the document class (for example, “hobbies”, “fashion”, “gourmet”, “sports”, etc. when the target document is a blog, web page, etc .; Clustering can be performed only from document and author information without specifying physics, “mechanical engineering”, “information engineering”, etc.). As a result, the following can be realized.

(Search for documents with similar contents)
By extracting a document belonging to the same document class as the search target document, it is possible to search for a document having a content similar to the search target document.

(Community search)
By clustering blog articles and Web pages, it is possible to classify them into a set of authors (authors) who handle similar contents. By extracting authors that belong to the same author class as you, you can extract a community of authors that handle topics you are interested in, or search for authors that are similar in interest.

(Collaborative filtering (Recommendation system))
By using this method as the user who purchased the product instead of the document instead of the document and applying this method as the user who purchased the product instead of the author, a set of similar products is extracted as the document class and the similar product is purchased as the author class. A set of users can be extracted. This makes it possible to recommend a product purchased by an author belonging to an author class to which a certain user belongs. Further, by specifying the class of the author who purchased the product of the document class to which a certain product belongs, it is possible to specify a user who is likely to like the product.

  In the prior art, when a document is classified only from words in the document, for example, when one author creates a large number of documents, the influence of the author becomes too great. was there. On the other hand, according to the document classification apparatus 1000 of the first and second embodiments, each author or each author class is associated with a document class, so that it is possible to avoid a situation where the influence of one author becomes too great. it can.

  Further, by creating a program to be executed by a computer constituting the document classification apparatus 1000 and installing the program in the computer, the computer can realize each function based on the program.

  Although description of this embodiment is finished above, the aspect of the present invention is not limited to these. The specific configuration and processing can be appropriately changed without departing from the gist of the present invention.

  Next, focusing on the extraction of the author's interest from the document, the results of qualitative evaluation and quantitative evaluation of the proposed model (models in the first and second embodiments) will be described. Two indicators, test set perplexity and KL-Divergence, were used for quantitative evaluation.

  The target document data is an academic paper, the number of authors (researchers) is 364, the number of documents is 14907, and the type of word is 27748. The number of topics T is fixed at 100. The parameters were α = 1 / T, β = 0.1, γ = 1 (LIT), and δ = 1. Further, the number of Gibbs sampling iterations was set to 10,000, and when the number of iterations reached 10,000 in the convergence determination unit 6, it was determined that the convergence was completed and the processing was terminated.

[Qualitative evaluation]
A part of the result of classifying the distribution of each author and the selected document class for each author class when using the method in the second embodiment (FIG. 14) and the distribution of the document class selected for each author class (FIG. 14) 15). These distributions can be calculated using the following formula (8) (corresponding to FIG. 14) and formula (9) (corresponding to FIG. 15), respectively.

なお、ｃは文書クラス、ａは著者であり、ｎ_ａｃは著者ａが作成した中で文書クラスがｃとなった文書の数である。γ_ｃはｃ次元の隠れパラメータγを意味する。＾は、実験結果として得られた推定値の平均値であることを意味する。

Note that c is a document class, a is an author, and _nac is the number of documents created by the author a and having the document class c. γ _c means a c-dimensional hidden parameter γ. ^ Means an average value of estimated values obtained as experimental results.

なお、ｃは文書クラス、ｓは著者クラスであり、ｎ_ｓｃは著者クラスｓの全著者が作成した中で文書クラスがｃとなった文書の数である。

Note that c is a document class, s is an author class, and n _sc is the number of documents created by all authors of the author class s and having the document class c.

  In FIG. 14 and FIG. 15, the document class of the document created by each author or an author belonging to the same author class is counted, and normalization is performed for each line. FIG. 14 shows an author / document class distribution, and FIG. 15 shows an author class / document class distribution. In FIG. 14, the vertical axis corresponds to the ID of each author, and is separated by a line l1 for each author class. The horizontal axis corresponds to the document class ID of the document created by each author. In FIG. 15, the vertical axis corresponds to the ID of the author class. A portion surrounded by a frame e1 in FIG. 15 corresponds to the author class in FIG. The horizontal axis corresponds to the document class ID of the document created by each author.

  14 and 15, when the number of corresponding documents is large, the filled area becomes large, and when the number of corresponding documents is small, the filled area becomes small. That is, a characteristic document class can be known for each author class from these diagrams.

  Similarly, a part of the result of classifying the distribution of each document and the selected topic class for each document class (FIG. 16) and the distribution of the topic class selected for each document class are shown (FIG. 17). These distributions can be calculated using the following equations (10) (corresponding to FIG. 16) and (11) (corresponding to FIG. 17), respectively.

ここで、ｚはトピッククラス、ｄは文書である。ｎ_ｄｚは、文書ｄの中でトピッククラスｔが選択された数である。δ_ｚはｚ次元の隠れパラメータδを意味する。

Here, z is a topic class and d is a document. n _dz is the number of topic classes t selected in the document d. [delta] _z denotes the hidden parameters [delta] of the z dimension.

ここで、ｚはトピッククラス、ｃは文書クラスである。ｎ_ｃｚは、文書クラスｃの全文書の中でトピッククラスｔが選択された数である。

Here, z is a topic class, and c is a document class. n _cz is the number of topic classes t selected in all documents of document class c.

  16 is a document / topic class, and FIG. 17 is a diagram (graph) showing the distribution of document classes / topic classes. In FIG. 16, the vertical axis corresponds to the ID of each document, and is separated by a line 12 for each document class. The horizontal axis corresponds to the topic class ID. In FIG. 17, the vertical axis corresponds to the ID of each document class, and the horizontal axis corresponds to the ID of the topic class. In FIG. 17, the part enclosed by a frame e2 corresponds to the document class of FIG.

  16 and 17, when the number of corresponding topics is large, the filled area becomes large, and when the number of applicable topics is small, the filled area becomes small. That is, a characteristic topic class can be known for each document class from these diagrams.

[Quantitative evaluation]
Next, in order to perform quantitative evaluation as a generation model of the proposed model (the model in the first and second embodiments), test set perplexity is calculated using parameters estimated by each model including the comparative example. Compare these.

  The feature of the proposed model is that the generation probability of a new document can be calculated for each author (first embodiment), and the generation probability of a new document can be calculated for the entire author (second embodiment).

For example, in the first embodiment, the probability P (c _d = j | c _{\ d} , z, γ, δ) is calculated for each document created by the author for each author. Then, it is possible to know the generation probability of how much of the document class the author belongs to. Further, by calculating the probability P (z _di = t | j, z _{\ di} , w, β, δ), the document newly created by the author is composed of a sequence of words belonging to what topic class. You can see. Therefore, each probability calculated in the first embodiment is a model that can calculate the generation probability of a new document for each author.

In the second embodiment, in addition to the calculation in the first embodiment, the probability P (s _a = h | s _{\ a} , c, α, γ) is calculated. As a result, it is possible to evaluate what kind of document class is likely to be generated by an author belonging to a certain author class. This allows the authors belonging to a certain author class to know the probability of the document class to which the new document to be generated next and the topic class included in the document belong. Obtainable.

  Therefore, quantitative evaluation is performed as follows for the second embodiment. First, 10% of all documents are selected at random and set as test documents. The remaining 90% of the documents from which the test document is removed are taken as learning documents. Then, the parameters are estimated using the learning document. Finally, the perplexity of words in the test document is calculated using these parameters. Perplexity is an index used in the language model field, and the lower this value, the higher the performance of the language model. PPX which is Test set perplexity is given by the following equation (12).

ここでθ_ｔはモデルにより文書dに与えられるトピッククラスtの選択確率、φ_ｔｖは単語ｖの出現確率、Ｗは単語数、Ｄ_textは文書数である。なお、非特許文献１の手法（ＡＴ）、非特許文献２の手法（ＡＰＴ）、および、第２実施形態の手法では、著者毎にトピック選択確率を与えることができるので、そのトピッククラスを用いた場合のTest set perplexityを算出するために、式（１２）を修正して次の式（１３）とした。

Here, θ _t is the selection probability of the topic class t given to the document d by the model, φ _tv is the appearance probability of the word v, W is the number of words, and D _text is the number of documents. In the method of Non-Patent Document 1 (AT), the method of Non-Patent Document 2 (APT), and the method of the second embodiment, a topic selection probability can be given for each author. In order to calculate the test set perplexity in the case of the above, the equation (12) is corrected to the following equation (13).

ここで、θ_ｔ｜ｊはモデルにより文書dに与えられる文書クラスjのトピッククラスtの選択確率、θ_ｊ｜ａはモデルにより文書dに与えられる著者aの文書クラスjの選択確率である。評価はテスト文書集合を変えて５回行い、そのperplexityの平均を計算した。なお、θ_ｔ｜ｊとθ_ｊ｜ａは、実際には式（８）と式（９）に定義されている推定値の平均値を利用した。表１に結果を示す。第２実施形態は、perplexityが総じて低く、モデルとして高性能であることが分かる。

Here, θ _{t | j} is the selection probability of the topic class t of the document class j given to the document d by the model, and θ _{j | a} is the selection probability of the document class j of the author a given to the document d by the model. The evaluation was performed 5 times while changing the test document set, and the average of the perplexity was calculated. For θ _{t | j} and θ _{j | a} , the average value of the estimated values defined in Equation (8) and Equation (9) was actually used. Table 1 shows the results. In the second embodiment, it can be seen that the perplexity is generally low and the model has high performance.

  Furthermore, in the second embodiment, since the document class can be given to each document using the distribution of the author class and the document selection distribution of each author class, the Test set perplexity when the document class is used. Formula (12) was modified to the following formula (14) so that PPX could be evaluated.

ここで、θ_ｊ｜ｈはモデルにより文書dに与える著者クラスhの文書クラスjの選択確率である。また、λ_ｈは次の式（１５）のように表すことができる。

Here, θ _{j | h} is the selection probability of document class j of author class h given to document d by the model. Further, λ _h can be expressed as the following equation (15).

なお、ｎ_ｈは、著者クラスｈに属する著者が作成した文書数である。また、α_ｈはｈ次元の隠れパラメータαを意味する。

Note that n _h is the number of documents created by an author belonging to the author class h. Α _h means an h-dimensional hidden parameter α.

The evaluation was performed 5 times by changing the test document set, and the average of the perplexity was used. Table 2 shows the results.

  Here, since the author class cannot be calculated in the conventional technique, the comparison is not performed, but the perplexity of the second embodiment is considered to be a low numerical value in view of the numerical size.

  Table 3 shows the results of calculating the distance between the word distributions for all topic combinations using KL-Divergence. If this value is large, it means that the difference in word distribution is large, so that the difference between topics can be clearly distinguished. From this result, it can be seen that the result of the second embodiment is higher in the accuracy of extracting independent topics than in the prior art.

DESCRIPTION OF SYMBOLS 1 Input part 2 Memory | storage part 3 Initial setting part 4 Document class evaluation part 5 Topic class evaluation part 6 Convergence determination part 7 Output part 8 Author class evaluation part 100,100a Operation part 1000,1000a Document classification apparatus

Claims (7)

For multiple documents that are electronic data,
Each of the plurality of documents belongs to one of a plurality of document classes that are document sets having similar contents, and each of the words constituting the plurality of documents is one of a plurality of topic classes that are word sets having similar contents. Based on a predetermined probability distribution model that probabilistically belongs, each author of the document is stochastically belonging to one or more of the document classes, and each of the document classes is associated with one or more of the topic classes. A document classification device for performing calculations,
The plurality of documents, a first calculation formula for estimating the document class to be used for each document used in the probability distribution model, and the topic class to be used for the word and used for the probability distribution model. A storage unit for storing a second calculation formula for estimating, and a predetermined calculation end condition;
An initial setting unit that randomly sets initial values of a document class to which each of the documents belongs and a topic class to which each of the words belongs;
Based on the first calculation formula, the probability of belonging to each of all the document classes is calculated for each document, and the document class having the highest probability is determined as the document class to which the document and the author belong. A document class evaluator that estimates as
Based on the second calculation formula, for each word, the probability of belonging to each of the topic classes is calculated, and the topic class having the highest probability is estimated as the topic class to which the word belongs. A class evaluation department;
A convergence determination unit that causes the document class evaluation unit and the topic class evaluation unit to repeat the estimation until the predetermined calculation end condition is satisfied;
An output unit for outputting a calculation result including a document and an author belonging to each document class when the convergence determination unit determines that the predetermined calculation end condition is satisfied;
A document classification apparatus comprising: For multiple documents that are electronic data,
Each of the plurality of documents is probabilistically belonging to any of a plurality of document classes that are document sets having similar contents, and each of the words constituting the plurality of documents is a plurality of topic classes that are word sets having similar contents And each of the authors of the document is stochastically belonging to one of the author classes that is a set of authors who have created similar documents, and the author class is assigned to one or more of the author classes. A document classification device that performs a calculation based on a predetermined probability distribution model that probabilistically belongs to a document class and associates each of the document classes with one or more topic classes,
The plurality of documents, a first calculation formula for estimating the document class to be used for each document used in the probability distribution model, and the topic class to be used for the word and used for the probability distribution model. And a second calculation formula for estimating the author class, a third calculation formula for estimating the author class to which each author belongs and used in the probability distribution model, and a predetermined calculation end condition are stored. A storage unit;
An initial setting unit that randomly sets initial values of an author class to which each of the authors belongs, a document class to which each of the documents belongs, and a topic class to which each of the words belongs;
Based on the third formula, for each author, the probability of belonging to each of the author classes is calculated, and the author class having the highest probability is estimated as the author class to which the author belongs. A class evaluation department;
Based on the first calculation formula, the probability of belonging to each of all the document classes is calculated for each document, and the document class having the highest probability is determined as the document class to which the document and the author class belong. A document class evaluator that estimates as
Based on the second calculation formula, for each word, the probability of belonging to each of the topic classes is calculated, and the topic class having the highest probability is estimated as the topic class to which the word belongs. A class evaluation department;
Until the predetermined calculation termination condition is satisfied, the author class evaluation unit, the document class evaluation unit, and the topic class evaluation unit, a convergence determination unit that repeats the estimation,
An output unit that outputs a calculation result including the content of the author class and the content of the document class when the convergence determination unit determines that the predetermined calculation end condition is satisfied;
A document classification apparatus comprising: The convergence determination unit has, as the predetermined calculation end condition, an error between a previous probability distribution and a latest probability distribution calculated by at least one of the author class evaluation unit, the document class evaluation unit, and the topic class evaluation unit is predetermined. The document classification apparatus according to claim 2, wherein the threshold value is equal to or less than a threshold value of the number of times or the number of repeated calculations reaches a predetermined number. For multiple documents that are electronic data,
Each of the plurality of documents belongs to one of a plurality of document classes that are document sets having similar contents, and each of the words constituting the plurality of documents is one of a plurality of topic classes that are word sets having similar contents. Based on a predetermined probability distribution model that probabilistically belongs, each author of the document is stochastically belonging to one or more of the document classes, and each of the document classes is associated with one or more of the topic classes. Do the calculations,
The plurality of documents, a first calculation formula for estimating the document class to be used for each document used in the probability distribution model, and the topic class to be used for the word and used for the probability distribution model. A storage unit for storing a second calculation formula for estimating, and a predetermined calculation end condition;
A document classification method by a document classification device having an initial setting unit, a document class evaluation unit, a topic class evaluation unit, a convergence determination unit, and an output unit,
The initial setting unit randomly sets initial values of a document class to which each of the documents belongs and a topic class to which each of the words belongs,
The document class evaluation unit calculates a probability of belonging to each of all the document classes for each document based on the first calculation formula, and selects the document class having the highest probability as the document and the document class. Estimate as the document class to which the author belongs,
The topic class evaluation unit calculates a probability of belonging to each of the topic classes for each word based on the second calculation formula, and the word belongs to a topic class having the highest probability. Estimating as the topic class,
The convergence determination unit causes the document class evaluation unit and the topic class evaluation unit to repeat the estimation until the predetermined calculation end condition is satisfied,
The document classification method, wherein the output unit outputs a calculation result including a document and an author belonging to each document class when the convergence determination unit determines that the predetermined calculation end condition is satisfied. . For multiple documents that are electronic data,
Each of the plurality of documents is probabilistically belonging to any of a plurality of document classes that are document sets having similar contents, and each of the words constituting the plurality of documents is a plurality of topic classes that are word sets having similar contents And each of the authors of the document is stochastically belonging to one of the author classes that is a set of authors who have created similar documents, and the author class is assigned to one or more of the author classes. Performing a calculation based on a predetermined probability distribution model that stochastically belongs to a document class and associates each of the document classes with one or more of the topic classes;
The plurality of documents, a first calculation formula for estimating the document class to be used for each document used in the probability distribution model, and the topic class to be used for the word and used for the probability distribution model. And a second calculation formula for estimating the author class, a third calculation formula for estimating the author class to which each author belongs and used in the probability distribution model, and a predetermined calculation end condition are stored. A storage unit;
A document classification method by a document classification device having an initial setting unit, an author class evaluation unit, a document class evaluation unit, a topic class evaluation unit, a convergence determination unit, and an output unit,
The initial setting unit randomly sets initial values of an author class to which each of the authors belongs, a document class to which each of the documents belongs, and a topic class to which each of the words belongs,
The author class evaluation unit calculates the probability of belonging to each of the author classes for each of the authors based on the third calculation formula, and the author belongs to the author class having the highest probability. Estimated as the author class,
The document class evaluation unit calculates a probability of belonging to each of all the document classes for each document based on the first calculation formula, and selects the document class having the highest probability as the document and the document class. Estimate as the document class to which the author class belongs,
The topic class evaluation unit calculates a probability of belonging to each of the topic classes for each word based on the second calculation formula, and the word belongs to a topic class having the highest probability. Estimating as the topic class,
The convergence determination unit causes the author class evaluation unit, the document class evaluation unit, and the topic class evaluation unit to repeat the estimation until the predetermined calculation end condition is satisfied,
The output unit outputs a calculation result including the content of the author class and the content of the document class when the convergence determination unit determines that the predetermined calculation end condition is satisfied. Classification method. The convergence determination unit has, as the predetermined calculation end condition, an error between a previous probability distribution and a latest probability distribution calculated by at least one of the author class evaluation unit, the document class evaluation unit, and the topic class evaluation unit is predetermined. 6. The document classification method according to claim 5, wherein the threshold value is equal to or less than a threshold value of the number of times or the number of repeated calculations reaches a predetermined number. A document classification program for causing a computer to function as the document classification device according to any one of claims 1 to 3.

Images