JP2002007433A - Information sorter, information sorting method, computer readable recording medium recorded with information sorting program and information sorting program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enhance sorting precision regardlessly of the contents and quantity of the information on an object to be sorted. SOLUTION: This information sorter is provided with a feature element extracting part 40 to extract feature elements for every sorting category from each of plural sample texts to be included in sorting sample data 30 in which a group 10 of sample texts is preliminarily associated with plural sorting categories, a sorting method deciding part 50 to decide a sorting method with the highest sorting precision from plural sorting methods based on the sorting sample data 30, a sort learning information generating part 60 to generate sorting learning information 70 showing features for every sorting category based o the feature elements extracted by the feature element extracting part 40 according to the sorting method decided by the sorting method deciding part 50, and an automatic sorting part 90 to sort a group 80 of new texts being the object to be sorted for every sorting category according to the sorting method decided by the sorting method deciding part 50 and the sorting learning information 70.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an information classification apparatus, an information classification method, a computer-readable recording medium on which an information classification program is recorded, and an information classification program used for classifying a large amount of text information and the like. In particular, an information classification device, an information classification method, and a computer-readable recording medium that records an information classification program that can increase the classification accuracy and efficiency by selecting a classification method with the highest classification accuracy from a plurality of classification methods , And an information classification program.

2. Description of the Related Art Recently, a huge amount of text information can be easily obtained by using the Internet. For this reason, there is a need for a technology that grasps the contents of a large amount of text information and efficiently extracts necessary text information from the content. This is because if these text information is classified into a predetermined classification category, it is convenient for a search when utilizing the text information later, or when finding related text information.

Conventionally, such a large amount of text information is classified by a person in charge of classification, a creator of the text information or a user of the text information to determine the contents of the new text information, and the classification is made up of a plurality of classification categories. Each was manually classified into the best classification category in the system. As another classification method, there is a method in which the content of new text information is analyzed using a computer system, and text information corresponding to a classification category is automatically classified based on the analysis result. The former classification method has a very high cost, and the latter classification method has problems in the number of classification categories and the classification accuracy for obtaining practical results.
For this reason, means and methods for effectively solving such problems have long been desired.

[0004]

2. Description of the Related Art At present, when a large amount of digitized text information has been distributed, classification based on the semantic content of text information has become an important issue from the viewpoint of efficient search / use of text information. ing. 2. Description of the Related Art Conventionally, as a means for solving such a problem, an information classification device that automatically executes a text information classification operation has been used in various fields.

Conventionally, a method for deriving a text information classification method based on a given text information classification example and then classifying new text information based on this classification method is disclosed in Japanese Patent Application Laid-Open No. H11-32821. No.
JP-A-11-296552, JP-A-11-167
Various classification methods are disclosed in, for example, Japanese Patent Application Laid-Open No. 581 and Japanese Patent Application Laid-Open No. 11-161671. Here, the following (1)
Items to (3) list conventional classification methods. (1) Statistical classification method based on stochastic model (2) Classification method that performs automatic classification by learning (3) Create rules for classifying text information into each classification category, and use this rule Classification method for automatic classification

[0006] In the classification method of the item (1), general classification tendency can be found, but fine classification tendency cannot be found. According to the classification method of item (2), high classification accuracy can be obtained when the number of classification categories is less than several tens, but when the number of classification categories increases to several tens or more, the classification accuracy decreases. In addition, the classification method of the item (3) requires a great deal of cost for creating and maintaining rules. in this way,
The classification methods of items (1) to (3) each have advantages and disadvantages.

FIG. 18 is a block diagram showing a configuration of a conventional information classification device. In this figure, classification sample data 2 is correct data relating to a classification composed of a plurality of texts in which which text is to be classified into which classification category. The feature element extraction unit 1 extracts, from each text, feature elements (words) representing the features of each classification category from the classification sample data 2.

Here, in extracting feature elements, it is necessary to efficiently extract feature elements that can enhance the discrimination ability of each classification category. Therefore, the feature element extraction unit 1 uses the feature element extraction method for enhancing the discrimination ability based on the appearance frequency of the feature element. A plurality of feature element extraction methods have been conventionally proposed. In addition, a method of designating some parts of speech for the attribute of the feature element is also employed.

The classification learning information generation unit 3 includes a feature element extraction unit 1
The feature of each classification category is calculated from the feature element extracted by the above, and the classification learning information 4 as the calculation result is generated. As the classification learning method in the classification learning information generation unit 3, a plurality of methods have been conventionally proposed.
The classification learning information 4 is information indicating the correspondence between the state of the feature element and the classification category. The automatic classification unit 5 is a classification target according to one classification method fixedly set in advance.
The new text group 6 including a plurality of texts is classified into classification categories based on the classification learning information 4, and the classification result data 7 is output.

[0010]

By the way, as described above, in the conventional information classification device (see FIG. 18),
Although it has been described that there are a plurality of feature element extraction methods of the feature element extraction unit 1, the classification accuracy in the classification result data 7 varies depending on the content and amount of the new text group 6 to be classified. Therefore, it is difficult to uniquely define a universal extraction method that maintains high classification accuracy for the new text group 6 having any content and amount.

Similarly, the classification learning information generating unit 3 has described that there are a plurality of classification learning methods. However, the classification accuracy in the classification result data 7 depends on the content and amount of the new text group 6. Fluctuates, it is difficult to uniquely define a universal classification learning method that maintains high classification accuracy. From this, in the conventional information classification device,
Inevitably, one of a plurality of classification methods (feature element extraction method, classification learning method) is fixedly used.

Therefore, in the conventional information classifying apparatus, the classification of the new text group 6 is performed by one fixed classification method. Therefore, the classification accuracy varies depending on the content and the amount of the new text group 6, and as a result, the classification is performed. There has been a problem that accuracy is reduced.

The present invention has been made in view of the above,
An object of the present invention is to provide an information classification device, an information classification method, a computer-readable recording medium on which an information classification program is recorded, and an information classification program capable of improving the classification accuracy regardless of the content and amount of information to be classified. And

[0014]

To achieve the above object, according to the present invention, a plurality of sample texts included in classification sample information in which a plurality of sample texts and a plurality of classification categories are associated in advance. A feature element extraction unit that extracts a feature element for each classification category from each of the texts; a classification method determination unit that determines a classification method with the highest classification accuracy from among a plurality of classification methods based on the classification sample information; A classification learning information generation unit configured to generate classification learning information representing a feature for each classification category based on the feature element extracted by the feature element extraction unit in accordance with the classification method determined by the classification method determination unit; According to the classification method determined by the method determination means and the classification learning information, a new text group to be classified is divided for each classification category. Characterized in that it comprises a classifying means for.

According to the present invention, a plurality of classification methods are made usable, and the classification method determining means determines a classification method having the highest classification accuracy from the plurality of classification methods based on the classification sample information. After that, the new text group is classified for each classification category according to this classification method, so that the classification accuracy can be improved as compared with the related art regardless of the content and amount of the information to be classified.

According to a second aspect of the present invention, there is provided the first aspect of the present invention.
In the information classification device described in the above, the feature element extracting means extracts each of the feature elements by a plurality of feature element extraction methods, and based on these extraction results, among the plurality of feature element extraction methods, includes A feature element extraction method having high discrimination ability is selected, and a feature element corresponding to the selection result is set as an extraction result.

According to the present invention, a plurality of feature element extraction methods can be used by the feature element extraction means, and feature elements respectively corresponding to the plurality of feature element extraction methods are extracted. Since the feature element corresponding to the feature element extraction method having a high ability to discriminate between categories is used as the extraction result, the classification accuracy of the classification result corresponding to this feature element can be further improved.

The invention according to claim 3 is based on claim 1.
The information classification device described in (1), further comprising an editing unit that edits the feature element extracted by the feature element extraction unit.

According to the present invention, the editing means is provided so that the extracted feature element can be edited (deleted, added, etc.), so that flexible feature element setting can be performed for the classification category.

The invention according to claim 4 is based on claim 1.
In the information classification device according to any one of the items (1) to (3), the classification method determination unit determines a classification method with the highest classification accuracy from a plurality of classification methods by a cross validation method.

According to the present invention, a plurality of classification methods are made available, and the classification method having the highest classification accuracy is selected from the plurality of classification methods based on the classification sample information by the classification method determination means. After determining by the validation method, the new text group is classified according to the classification category according to this classification method, so that the classification accuracy can be improved as compared with the past, regardless of the content and amount of classification target information. it can.

Further, the invention according to claim 5 is based on claim 1.
5. The information classification device according to any one of items 4 to 4, further comprising a specification unit that specifies a classification target portion in the sample information and the new text group.

According to the present invention, the classifying sample information and the classification target portion in the new text group are designated by the designation means.
Since the classification can be performed on essentially necessary parts, the classification accuracy can be further improved.

The invention according to claim 6 is the invention according to claim 1.
In the information classification device according to any one of the first to fifth aspects, by clustering a plurality of sample texts, clustering means for generating the classified sample information in which the plurality of sample texts are associated with a plurality of classification categories. It is characterized by having.

According to the present invention, since the classification sample information is generated by the clustering means, the efficiency can be remarkably improved as compared with a case where the classification category is manually generated from a plurality of sample texts. At the same time, the work load on the user can be reduced.

The invention according to claim 7 is the first invention.
5. The information classification apparatus according to any one of claims 1 to 5, wherein the clustering unit clusters the classified sample information, a comparison unit that compares a clustering result of the clustering unit with a desired clustering result, and a comparison of the comparison unit. Changing means for changing the classification sample information as necessary based on the result.

According to the present invention, the clustering result of the clustering means is compared with a desired clustering result, and when the comparison result is, for example, inconsistent, the classification sample information can be changed by the changing means. Since the new text group can be classified based on the proper classification sample information, the classification accuracy can be extremely increased.

The invention according to claim 8 is the invention according to claim 1.
7. The information classification apparatus according to any one of items 1 to 7, further comprising a clustering result display unit that clusters a new text group in the classification result of the classification unit and displays the clustering result.

According to the present invention, the clustering result is displayed by the clustering result display means, so that the user can easily grasp the distribution of the classification results.

The invention according to claim 9 is based on claim 1.
The information classification device according to any one of to 8, further comprising an optimization unit that optimizes a classification result of the classification unit,
The classification learning information generating means regenerates the classification learning information based on the optimized classification result, and the classification means
In accordance with the classification method determined by the classification method determination means and the regenerated classification learning information, a new text group to be classified is classified for each classification category.

According to the present invention, the classification learning information is regenerated based on the classification result optimized by the optimizing means, and the new text group is classified again according to the classification learning information. Classification accuracy can be improved.

According to a tenth aspect of the present invention, in the information classification apparatus according to the ninth aspect, the difference between the classification result before the optimization and the classification result after the optimization is visually recognizable. It is characterized by comprising a difference recognition information display means for displaying as recognition information.

According to the present invention, the difference between the classification results before and after the optimization is displayed as the difference recognition information so that the user can recognize the difference at a glance, so that the user can quickly respond based on the difference. As a result, classification accuracy can be improved.

Further, according to the present invention, a feature element is extracted for each classification category from each of a plurality of sample texts included in classification sample information in which a plurality of sample texts and a plurality of classification categories are associated in advance. A feature element extraction step, based on the classification sample information, a classification method determination step of determining a classification method having the highest classification accuracy from among a plurality of classification methods, and a classification method determined in the classification method determination step. A classification learning information generating step of generating classification learning information representing a feature for each classification category based on the characteristic element extracted in the characteristic element extraction step; a classification method determined in the classification method determining step; A classification step of classifying a new text group to be classified into classification categories according to the learning information.

According to the present invention, a plurality of classification methods are made available, and in the classification method determining step, a classification method having the highest classification accuracy is determined from the plurality of classification methods based on the classification sample information. After that, the new text group is classified for each classification category according to this classification method, so that the classification accuracy can be improved as compared with the related art regardless of the content and amount of the information to be classified.

According to a twelfth aspect of the present invention, a feature element is extracted for each classification category from each of a plurality of sample texts included in classification sample information in which a plurality of sample texts and a plurality of classification categories are associated in advance. A feature element extracting step, based on the classification sample information, a classification method determining step of determining a classification method having the highest classification accuracy from a plurality of classification methods, and a classification method determined in the classification method determination step. A classification learning information generating step of generating classification learning information representing a feature for each classification category based on the characteristic element extracted in the characteristic element extraction step; a classification method determined in the classification method determining step; A classification step of classifying a new text group to be classified into classification categories in accordance with the learning information. A computer-readable recording medium recording the order information classification program.

According to the present invention, a plurality of classification methods are made available, and the classification method having the highest classification accuracy is determined from the plurality of classification methods based on the classification sample information in the classification method determination step. After that, the new text group is classified for each classification category according to this classification method, so that the classification accuracy can be improved as compared with the related art regardless of the content and amount of the information to be classified.

According to a thirteenth aspect of the present invention, a feature element is extracted for each classification category from each of a plurality of sample texts included in classification sample information in which a plurality of sample texts and a plurality of classification categories are associated in advance. A feature element extraction procedure, a classification method determination procedure for determining a classification method having the highest classification accuracy from among a plurality of classification methods based on the classification sample information, and a classification method determined in the classification method determination procedure. A classification learning information generating step of generating classification learning information representing a feature for each classification category based on the characteristic element extracted in the characteristic element extraction step; a classification method determined in the classification method determination step; The computer executes a classification procedure for classifying a new text group to be classified into classification categories according to the learning information. It is because of information classification program.

According to the present invention, a plurality of classification methods are made available, and the classification method having the highest classification accuracy is determined from the plurality of classification methods based on the classification sample information in the classification method determination procedure. After that, the new text group is classified for each classification category according to this classification method, so that the classification accuracy can be improved as compared with the related art regardless of the content and amount of the information to be classified.

[0040]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, with reference to the drawings, an information classification device, an information classification method, and a computer-readable recording medium recording an information classification program according to the present invention will be described.
An embodiment of the information classification program will be described in detail.

FIG. 1 is a block diagram showing the configuration of an embodiment according to the present invention. In this figure, a sample text group 10 is a set of a plurality of unclassified texts. The clustering unit 20 includes the sample text group 10
Are clustered to generate classification sample data 30. The classification sample data 30 is correct answer data relating to a classification including a plurality of texts in which which text is to be classified into which classification category.

The feature element extraction unit 40 extracts feature elements (words) representing the features of each classification category from each text from the classification sample data 30 in the same manner as the feature element extraction unit 1 (see FIG. 18). . However, while the feature element extraction unit 1 extracts feature elements according to one feature element extraction method, the feature element extraction unit 40 extracts feature elements according to each of a plurality of feature element extraction methods. , The feature element extraction unit 1.

The classification learning information generation unit 60 is similar to the classification learning information generation unit 3 (see FIG. 18) in that the feature element extraction unit 4
0, the feature of each classification category is calculated from the feature element extracted, and the classification learning information 7 as the calculation result is obtained.
Generate 0. However, while the classification learning information generation unit 3 calculates a feature according to one classification learning method, the classification learning information generation unit 60 calculates a feature according to each of a plurality of classification learning methods. Different from the generation unit 3.

The classification method determining unit 50 determines a classification method having the highest classification accuracy from among a plurality of classification methods by, for example, well-known cross validation. Details of the operation of the classification method determining unit 50 will be described later. New text group 80, as shown in FIG. 2, a plurality of new text TX ₁ to be classified (text name text1) ~ new text TX ₁₀ (text name text10), consisting of .... Returning to FIG. 1, the automatic classification unit 90 includes the classification method determination unit 5
The new text group 80 is classified into classification categories based on the classification method and the classification learning information 70 determined by 0, and this is output as the classification result data 100 (see FIG. 3).

The clustering unit 110 clusters the classification result data 100 and generates a clustering result C
(See FIG. 4). The display unit 120 is a display that displays the clustering result C from the clustering unit 110 and various data from each unit. 5 to 7 show display examples of the display unit 120. Input unit 1
Reference numeral 30 denotes a mouse, a keyboard, and the like for performing an editing operation described later, a window operation on the display unit 120, and the like.

Next, the operation of the above-described embodiment will be described with reference to the flowcharts shown in FIGS. When the sample text group 10 is input to the clustering unit 20 shown in FIG. 1, the clustering unit 20 clusters a plurality of texts in the sample text group 10 in step SA1 shown in FIG.
At step SA2, the clustering unit 20 categorizes each cluster into categories. In step SA3, the clustering unit 20 classifies the classification sample data 30 (correct data) regarding a classification including a plurality of texts in which classifications are classified into which texts.
Is output to the feature element extraction unit 40.

Thus, in step SA4, the feature element extraction unit 40 executes a feature element extraction process of extracting, from each text, a feature element (word) representing a feature of each category in the category sample data 30. That is, in step SB1 shown in FIG.
0 extracts a feature element (word) candidate representing the feature of the classification category by performing a morphological analysis on the classification sample data 30. In step SB2, the feature element extraction unit 40
Executes a process of unifying synonyms in the extracted feature element candidates.

In step SB3, the feature element extraction unit 40
Counts, for each of the plurality of extracted feature element candidates, the same word feature element. Step S
In B4, the feature element extraction unit 40 executes a ranking process for narrowing down a plurality of feature element candidates for each classification category.
In this ranking process, for a plurality of candidate feature elements, a method of ranking feature elements by classification category in descending order of appearance frequency, a method of ranking feature elements by classification category in descending order of appearance probability, A statistical method (a method of lowering the ranking of feature elements that also appear in other classification categories) is used to calculate the frequency, and a method of ranking feature elements for each classification category is employed.

In step SB5, the feature element extraction unit 40
Extracts a predetermined number of the above-described feature elements having high rankings from the top for each classification category, and extracts these as feature elements. In step SB6, the feature element extraction unit 40 outputs the extracted feature element as feature element extraction result data. FIG. 11 shows a feature element appearance frequency order list R ₁ ranked in the order of appearance frequency among the three ranking methods described above.
FIG. 14 is a diagram illustrating (corresponding to feature element extraction result data).

FIG. 5 shows a classification category (“Econom”).
ic "," Foreign ", ...," Societ "
y "and" Sport "), and a frequency field indicating the frequency of appearance of feature elements (" market "," yen appreciation ", etc.) in the classification category. In the record corresponding to each classification category, the number of texts classified into the classification category is described. Here, the text refers to a text constituting the sample text group 10 (see FIG. 1). For example, a classification category of “Economic” classifies 27 texts, and a classification category of “Foreign” includes 4 texts.
Three texts are classified.

The leftmost field in the figure is a ranking indicating the order of appearance frequency. For example, "Economi
In the category “c”, the feature element whose ranking of the appearance frequency in the 27 texts is the first is “market”
(Frequency: 61.0) The feature element of the second place is “yen high”
40.0), and similarly, the thirtieth feature element is “finance” (frequency: 12.0).

FIG. 12 shows a feature element appearance frequency order list R ₂ (feature element ranking) in which feature elements are ranked for each classification category by adopting a statistical method called the Kullback-Leibler method among the three ranking methods described above. (Corresponding to extraction result data). The basic configuration of the feature elements ranking list R ₂ shown in this figure is the same as that of the feature element appearance frequency order list R ₁ (see FIG. 11).

However, the feature element ranking list R
_{In 2} , a statistical method is adopted for lowering the ranking of feature elements that also appear in other classification categories and improving the ability to discriminate the classification category from other classification categories. For example, "Economic" shown in FIG.
"Dollar" ranked third in the classification category
(Feature element) is ranked 31st or lower (not shown) in the classification category of “Economic” shown in FIG.

Returning to FIG. 8, in step SA5, the classification method determining section 50 determines whether or not to automatically determine the classification method to be applied to the new text group 80. If there is no instruction from the user, the classification method determining unit 50 proceeds to step SA
The determination result of No. 5 is “Yes”. On the other hand, when the user instructs the classification method by manual operation, the classification method determination unit 50 sets the determination result in step SA5 to “No”, and determines the classification method based on the instruction from the user in step SA7.

In this case, in step SA6, the classification method determining unit 50 performs, for example, cross validation.
A classification method determination process for automatically determining a classification method is executed. That is, in step SC1 shown in FIG. 10, the classification method determination unit 50 randomly divides the classification samples (texts) into N pieces for each classification category in the classification sample data 30. In step SC2, the classification method determination unit 50 applies a plurality of learning algorithms (classification methods) to the (N-1) classification samples, respectively, and outputs a feature element or classification learning information corresponding to each learning algorithm. Create

In step SC3, the classification method determining section 50
Performs a classification test and calculates classification accuracy by applying the learning algorithm method to the remaining (1 / N) classification samples using the feature elements and the classification learning information created in step SC2. This classification accuracy is individually calculated for each of the plurality of learning algorithms. In step SC4, the classification method determination unit 50 determines whether or not the classification test has been performed N times.
The determination result is “No”. Thereafter, in step SC2 and step SC3, the classification accuracy for each of the N classification samples is calculated for each of the plurality of learning algorithms by changing the classification samples one by one.

Then, the determination result of step SC4 is "Y
When "es" is reached, in step SC5, the classification method determination unit 50 calculates an average value of the classification accuracy for the N classification samples for each of the plurality of learning algorithms. Step SC
In 6, the classification method determination unit 50 selects the highest one of the average values of the plurality of classification accuracies respectively corresponding to the plurality of learning algorithms (classification methods), and then selects the learning algorithm (classification method) corresponding to the selected classification accuracy. Method). Further, the classification method determining unit 50 determines a learning algorithm (classification method) having the highest classification accuracy by the classification learning information generation unit 6.
0 and notify the automatic classification unit 90.

Returning to FIG. 8, in step SA8, the classification learning information generation unit 60 converts the learning algorithm (classification method) notified by the classification method determination unit 50 and the feature element extraction result data from the feature element extraction unit 40 into the learning algorithm. Based on this, the classification learning information 70 is generated. In step SA9, the classification learning information generation unit 60 registers the classification learning information 70 in a database (not shown). In step SA10, the automatic classification unit 90 determines whether or not the new text group 80 to be classified has been input. In this case, the determination result is “No”.
And repeat the same judgment.

When the new text group 80 is input to the automatic classification unit 90, the automatic classification unit 90
The determination result of No. 10 is “Yes”. Step SA11
Then, the automatic classifying unit 90 sets the new text TX ₁ and the new text TX constituting the new text group 80 (see FIG. 2).
₂ ,... It is determined whether or not all the automatic classifications of the new texts TX ₁₀ ,... Have been completed.
o ". Hereinafter, Step SA15 to Step SA2
In 1, the automatic classification unit 90 executes an automatic classification process based on the classification method determined by the classification method determination unit 50.

In the following, as an example of the classification method, a case where the new text group 80 is classified based on the vector space method will be described. In this case, it is assumed that the classification learning information 70 includes 30 feature elements for each category, and includes vectors of all feature elements and vectors of each category. In this state, step SA1
In 5, the automatic classification unit 90 performs a morphological analysis on the new text TX ₁ (see FIG. 2) in the new text group 80 and extracts a feature element (word). Step SA1
In 6, the automatic classification unit 90 executes a synonym unification process of unifying synonyms in the extracted feature elements.

In step SA17, the automatic classifying section 90
Counts the extracted feature elements. Step SA1
In 8, automatic classification section 90, the same feature element and the feature element in the classification learning information 70, obtains a plurality of feature elements included in the new text TX _1. Next, the automatic classification unit 90
Acquired feature element, i.e., generates a document vector for new text TX _1.

At Step SA19, the new text TX
The similarity (cosine value) between the document vector related to ₁ and the vector of each classification category in the classification learning information 70 is calculated. The similarity (cosine values) is a vector of classification category A, When the document vector for the new text TX ₁ B is expressed by the following. Similarity (cosine value) = Inner product of vector A and document vector B / (size of vector A × size of document vector B)

[0063] That is, in step SA19, in relation to the new text TX _1, a few minutes of the similarity of the classification category (cosine value) is calculated. At step SA20,
The automatic classifying unit 90 normalizes the plurality of calculated similarities (cosine values) (to a value from 0 to 100). In step SA21, the automatic classification unit 90 selects a similarity that is equal to or greater than a threshold value (for example, 70) among a plurality of similarities (cosine values), and then newly assigns a classification category corresponding to the selected similarity. to classify the text TX _1. In the case where all of the plurality of similarity is less than the threshold value, automatic classification section 90, a text that can not be classified the new text TX _1. Thereafter, Step SA15 to Step SA
By repeating the processing up to 21, the new text is sequentially classified into the classification categories.

When the classification of all new texts is completed, the automatic classification unit 90 sets the result of the determination in step SA11 to "Yes". In step SA12, the automatic classification unit 90 outputs the classification result data 100 shown in FIG. In this figure, the text names text1 to text20, ... correspond to the text names text1 to text10, ... shown in Fig. 2 and "AUTOMOTIVE_INDUSTRY"
And so on indicate a classification category, and a number on the right side of the classification category indicates a score (for example, similarity). In other words, the new text TX ₁ shown in FIG. 2, "AUTOMOTIVE
_INDUSTRY ”, and the score (similarity) is“ 85 ”.

Returning to FIG. 8, in step SA 13, clustering section 110 clusters new text group 80 using classification result data 100. FIG.
9 is a diagram showing a clustering result C in the clustering unit 110. FIG. This figure shows a case where a new text group 80 consisting of 1000 new texts is classified, and is classified into a classification category “Economic”.
A breakdown (the number of texts, characteristic elements) of the 26 new texts when 6 new texts are classified is illustrated.

In step SA14, the display unit 120 displays, for example, the clustering result C shown in FIG. Thereby, the user can confirm what kind of content is classified into the classification category (Economic in this case).

In one embodiment, FIG.
The indicated feature element ranking list R_TwoIs displayed on the display unit 120.
Feature list according to the user's request.
R _Two Is edited, and the feature element ranking list shown in FIG.
R_Three May be used to perform the classification. This place
In this case, the user uses the input unit 130 to
Grist R_Two Delete feature elements that are judged unnecessary in
Edit that. As a result, the feature element ranking squirrel
R_Three (See FIG. 13) is created, and this feature element Rankin
Grist R_Three The above-described processing is executed on the basis of.

In one embodiment, when the structure of the classification sample data 30 and the new text group 80 is defined in advance, the input unit 130 specifies the classification target portion in the classification sample data 30 and the new text group 80. You may make it.

In the above description, an example was described in which the result (classified sample data 30) clustered by the clustering unit 20 shown in FIG. 1 was used as it is in the feature element extraction unit 40. The clustered result will be verified. You may do so. Hereinafter, this case will be described as a first modification of the embodiment with reference to FIGS.

In step SD1 shown in FIG.
The clustering unit 20 performs clustering on the sample text group 10 included in the classification sample data 30 (correct answer data) shown in FIG. in this case,
The assignment of the classification category in the classification sample data 30 is ignored. FIG. 14 is a diagram illustrating a result of clustering performed by the clustering unit 20 (clustering result distribution data CB). The figure shows seven classification categories ("Economic", "Foreign",
... “Sport”) are clustered in 277 texts.

According to this figure, “Spor” of the A record
"ts", "Politics" of C and E records
Indicates that the classification categories are clearly assigned. On the other hand, D record "Economi
c ”and“ Industry ”and“ Foreign ”,“ Industry ”,“ Po ”
liters "," Science "," Societ "
It can be seen that the distinction of "y" is ambiguous. In this case, the process of step SD4 described later is executed. In step SD2, the result of the clustering (assignment of the classification category) is compared with the assignment of the classification category originally assumed by the user by the comparison unit (not shown).

In step SD3, the comparing section determines whether or not the comparison result in step SD2 is the same. If the determination result is "No", the comparing section displays the comparison result on display section 120.
To be displayed. Thus, in step SD4, the user uses the input unit 130 to review the clustered result (assignment of the classification category) and edit the classification category. On the other hand, if the determination result of step SD3 is “Ye
s ", that is, since the assignment of the classification category in the classification sample data 30 is the same as that initially assumed by the user, in step SD5, the classification category and the classification sample (text) are set as the learning information. . In step SD6, the comparison unit (not shown) outputs the classification sample data 30 to the feature element extraction unit 40. Thereby, the above-described processing is executed.

In the above, an example has been described in which the classification result data 100 classified by the automatic classification unit 90 is output as it is. However, after the classification is performed by the automatic classification unit 90, the classification result data 100 is a desired one. It is also possible to verify whether or not there is, and if the verification result is NG, feed back the verification result to the classification learning information 70 and re-learn to improve the classification accuracy. Hereinafter, this case will be referred to as a second modification of the embodiment shown in FIG.
This will be described with reference to FIG. In the figure, portions corresponding to the respective portions in FIG. 1 are denoted by the same reference numerals. In this figure, a relearning processing unit 140 is newly provided. The relearning processing unit 140 creates the classification learning information 70A in response to the feedback described above.

When a new text group 80 composed of 20 new texts is input to the information classification device 200, the new text group 80 is generated based on the classification learning information 70 and a predetermined classification method in the same manner as the above-described operation. Are automatically classified. As a result, the classification result data 100 is output from the information classification device 200. This classification result data 10
0 is displayed on the display unit 120. Here, in the classification result data 100, the new texts (5) and (6) allocated to the classification category B should be allocated to the classification category A, and the new texts (9) allocated to the classification category C. Should be assigned to the classification category D, the user
Is used to edit to the desired assignment.

Thus, the re-learning processing unit 140 executes the re-learning process based on the edited classification result data 100 by the same operation as the classification learning information generating unit 60 (see FIG. 1). The information 70A is reconstructed. In this state, when the new text group 80 is input to the information classification device 200, the new text group 80 is generated based on the reconstructed classification learning information 70A and the predetermined classification method in the same manner as the operation described above. Automatically classified. In this case, the classification accuracy of the classification result data 100 output from the information classification device 200 is extremely high due to the effect of relearning.

In one embodiment, the table shown in FIG.
The screen G shown in FIG.₁Is displayed and the classification
Various kinds of information generated in the process may be displayed.
Screen G₁ Has a classification category called “User Claim Classification”.
Folder H corresponding to category ₀ , Distribution of this classification category
"Initial failure" under ... "Inquiry" and
Classified as "Unclassified documents"
Corresponding folder H₁~ H₇ Are displayed respectively
You.

Further, screens G _{2 to} G ₄ are displayed on the screen G ₁ by window control. The screen G _2,
Title K ₁ and text content K ₂ of the sample document corresponding to the classification category of "inquiry" (corresponding to the classification sample data 30) is displayed as shown in FIG. 6. Also, the screen G ₃ shown in FIG. 7, the keyword corresponding to the classification category of "Query" (feature element) is displayed. Screen G ₄ shown in FIG. 5,
The contents display screen J ₂ is displayed on the contents of the list screen J ₁ and the new text of the new text, which is classified as a classification category of "inquiry". Here, the icon I ₁ ~I ₄ in the list screen J ₁ of the new text
Is the score (similarity) before re-learning according to Modification 2 described above.
Represents a change in score after re-learning with respect to.

That is, the icon I ₁ means that the score (similarity) is higher than the previous time, and the icon I ₂ means that the score (similarity) is lower than the previous time. ing. Icon I ₃ is the last time, the new text, which has been classified into the classification category (in this case "inquiry") is, this time, which means that it has not been classified into the classification category. Also, the icon I ₄ was the last time,
This means that a new text that has not been classified into the classification category (in this case, “inquiry”) is now classified into the classification category.

As described above, according to one embodiment, a plurality of classification methods are set to be usable, and the classification method determining unit 50 sets a plurality of classification methods based on the classification sample data 30. After determining the classification method with the highest classification accuracy from, the new text group 80 is classified according to the classification method according to this classification method. , Classification accuracy can be improved.

According to one embodiment, a plurality of feature element extraction methods can be used by the feature element extraction unit 40, and feature elements corresponding to the plurality of feature element extraction methods are extracted. In particular, since a feature element corresponding to a feature element extraction method having high discrimination ability between classification categories is used as an extraction result, the classification accuracy of the classification result corresponding to this feature element can be further improved.

According to one embodiment, the input unit 13
0 and the display unit 120 (editing means) are provided so that the extracted feature elements can be edited (deleted, added, etc.), so that flexible feature element setting can be performed for the classification category.

According to one embodiment, the input unit 13
0 and the display unit 120 (designating means) designate a classification target portion in the classification sample data 30 and the new text group 80. Therefore, a portion unnecessary for classification is eliminated, and a portion essentially required is targeted. Since the classification can be performed, the classification accuracy can be further improved.

Further, according to the embodiment, the classification sample data 30 is generated by the clustering unit 20, so that the efficiency is significantly improved as compared with the case where the classification category is manually generated from a plurality of sample texts. Can be improved, and the work load on the user can be reduced.

Further, according to one embodiment, the clustering result of the clustering unit 20 is compared with a desired clustering result. Since the data 30 can be changed, the new text group 80 can be classified based on the more complete classification sample data 30, so that the classification accuracy can be extremely increased.

According to one embodiment, the display unit 12
Clustering result distribution data CB to 0 (see FIG. 14)
Is displayed, so that the user can easily grasp the distribution of the classification results.

Further, according to one embodiment, as described in the second modification, the classification learning information 70A is regenerated based on the optimized classification result, and the classification learning information 70A is generated.
, The new text group 80 is classified again, so that the classification accuracy can be further improved.

[0087] According to one embodiment, to display the difference of the classification result in the before and after optimization icon I ₁ to 1 ₄ as (difference recognition information), so the user has to be aware of the difference at a glance In addition, it is possible to promptly respond to the user based on the difference, and as a result, the classification accuracy can be improved.

Although the embodiment of the present invention has been described in detail with reference to the drawings, the specific configuration is not limited to this embodiment and does not depart from the gist of the present invention. Even if there is a change in the design, the present invention is included in the present invention. For example, in the above-described embodiment, an information classification program for realizing the function of the information classification device is stored in the computer-readable recording medium 4 shown in FIG.
00, the information classification program recorded in the recording medium 400 may be read by the computer 300 shown in FIG.

The computer 300 shown in FIG. 17 stores a CPU 301 for executing the information classification program, an input device 302 such as a keyboard and a mouse, a ROM (Read Only Memory) 303 for storing various data, and an arithmetic parameter. RAM (Random Access Memory) 3 for storing
04, a reading device 305 that reads the information classification program from the recording medium 400, an output device 306 such as a display or a printer, and a bus BU that connects each unit of the device.

The CPU 301 reads the information classification program recorded on the recording medium 400 via the reading device 305, and executes the information classification program to perform the information classification described above. The recording medium 400
Include optical disks, floppy (registered trademark) disks,
In addition to a portable recording medium such as a hard disk, a transmission medium such as a network for temporarily recording and holding data is also included.

Further, in the embodiment, the case where the cross-validation method is adopted as an example of the method of determining the classification method in the classification method determining section 50 shown in FIG. 1 has been described, but the present invention is not limited to this method. Instead, the classification method may be determined by using the values such as the recall (the ratio of correct answers included in the results) and the precision (the number of mistakes in the results) as keys. In short, as long as a plurality of classification methods can be used and a classification method having the highest classification accuracy can be selected from these classification methods, the present invention can be applied to any method. included.

[0092]

As described above, according to the first aspect of the present invention, a plurality of classification methods are set in a usable state, and a plurality of classification methods are determined by the classification method determining means based on the classification sample information. After determining the classification method with the highest classification accuracy from among, the new text group is classified according to the classification category according to this classification method. Therefore, an effect that the classification accuracy can be improved can be achieved.

According to the second aspect of the present invention, a plurality of feature element extraction methods can be used by the feature element extraction means, and feature elements respectively corresponding to the plurality of feature element extraction methods are set. Since extraction is performed, in particular, a feature element corresponding to a feature element extraction method having a high ability to discriminate between classification categories is used as an extraction result, the classification accuracy of the classification result corresponding to this feature element can be further improved. It works.

According to the third aspect of the present invention, an editing means is provided so that the extracted feature element can be edited (deleted, added, etc.). This has the effect that it can be performed.

According to the fourth aspect of the present invention, a plurality of classification methods are set in a usable state, and the classification method determining means determines the most accurate classification among the plurality of classification methods based on the classification sample information. After determining the classification method with high cross-validation method, the new text group is classified for each classification category according to this classification method, so compared to the past, regardless of the content and amount of classification target information, There is an effect that classification accuracy can be improved.

According to the fifth aspect of the present invention, the designation means designates the classification sample information and the classification target position in the new text group. Therefore, unnecessary portions for classification are eliminated, and essentially, Since classification can be performed for a necessary portion as an object, there is an effect that classification accuracy can be further improved.

According to the sixth aspect of the present invention, since the classification sample information is generated by the clustering means, the efficiency is significantly improved as compared with the case where the classification category is manually generated from a plurality of sample texts. And the burden on the user can be reduced.

According to the seventh aspect of the present invention, the clustering result of the clustering means is compared with a desired clustering result, and when the comparison result is, for example, inconsistent, the classification sample information can be changed by the changing means. Since it is possible to classify a new text group based on more complete classification sample information,
There is an effect that the classification accuracy can be made extremely high.

According to the eighth aspect of the present invention, since the clustering result is displayed by the clustering result display means, the user can easily grasp the distribution of the classification result.

According to the ninth aspect of the present invention, classification learning information is regenerated based on the classification result optimized by the optimizing means, and the new text group is classified again according to the classification learning information. As a result, there is an effect that the classification accuracy can be further improved.

According to the tenth aspect of the present invention,
The difference between the classification results before and after the optimization is displayed as difference recognition information so that the user can recognize the difference at a glance, so that the user can be promptly dealt with based on the difference, and as a result, the classification accuracy is improved. It has the effect of being able to increase.

According to the eleventh, twelfth, and thirteenth aspects of the present invention, a plurality of classification methods are set to be usable, and in the classification method determining step, a plurality of classification methods are selected based on the classification sample information. After determining the classification method with the highest classification accuracy from, the new text group is classified according to the classification method according to this classification method, so compared to the past, regardless of the content and amount of information to be classified, There is an effect that classification accuracy can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment according to the present invention.

FIG. 2 is a diagram showing an example of a new text group 80 shown in FIG.

FIG. 3 is a diagram showing an example of classification result data 100 shown in FIG.

FIG. 4 is a diagram showing a clustering result C in the clustering unit 110 shown in FIG.

FIG. 5 is a diagram showing a display example of a display unit 120 shown in FIG.

FIG. 6 is a diagram showing a display example of a display unit 120 shown in FIG.

FIG. 7 is a diagram showing a display example of a display unit 120 shown in FIG.

FIG. 8 is a flowchart illustrating the operation of the same embodiment.

FIG. 9 is a flowchart illustrating a feature element extraction process illustrated in FIG. 8;

FIG. 10 is a flowchart illustrating a classification method determination process illustrated in FIG. 8;

11 is a diagram illustrating a feature element appearance frequency order list R ₁ in the same embodiment.

12 is a diagram showing the feature elements ranking list R ₂ in the same embodiment.

13 is a diagram showing the feature elements ranking list R ₃ in the same embodiment.

FIG. 14 is a diagram showing clustering result distribution data CB in the same embodiment.

FIG. 15 is a flowchart illustrating a first modification of the same embodiment.

FIG. 16 is a diagram illustrating a second modification of the same embodiment.

FIG. 17 is a block diagram showing a third modification of the same embodiment.

FIG. 18 is a block diagram showing a configuration of a conventional information classification device.

[Explanation of symbols]

 Reference Signs List 20 clustering unit 40 feature element extraction unit 50 classification method determination unit 60 classification learning information generation unit 90 automatic classification unit 110 clustering unit 120 display unit 130 input unit 300 computer 301 CPU 400 recording medium

Continued on the front page (72) Inventor Norihiko Sakamoto 18-1, Minamicho, Shizuoka City, Shizuoka Prefecture Inside Fujitsu Infosoft Technology Co., Ltd. (72) Inventor Ryu Shibata 18-1, Minamicho, Shizuoka City, Shizuoka Prefecture Fujitsu Infosoft Co., Ltd. F term in technology (reference) 5B075 ND03 NK06 NK32 NK46 NR12 PQ02 PQ46 QM05 UU06 5B082 GA08

Claims (13)

[Claims] A feature element extracting unit configured to extract a feature element for each classification category from each of a plurality of sample texts included in classification sample information in which a plurality of sample texts and a plurality of classification categories are associated in advance; A classification method determining unit that determines a classification method with the highest classification accuracy from a plurality of classification methods based on the classification sample information; and an extraction by the feature element extraction unit according to the classification method determined by the classification method determination unit. A classification learning information generating unit that generates classification learning information representing a feature for each classification category based on the obtained feature element; and a classification target according to the classification method determined by the classification method determination unit and the classification learning information. An information classification device, comprising: a classification unit that classifies a new text group for each classification category. 2. The feature element extraction means extracts feature elements by a plurality of feature element extraction methods, respectively, and based on the extraction results, a discriminating ability between classification categories from the plurality of feature element extraction methods. The information classification apparatus according to claim 1, wherein a high feature element extraction method is selected, and a feature element corresponding to the selection result is set as an extraction result. 3. The information classification apparatus according to claim 1, further comprising an editing unit that edits the feature element extracted by the feature element extraction unit. 4. The classification method according to claim 1, wherein the classification method determination unit determines a classification method having the highest classification accuracy from a plurality of classification methods by a cross-validation method. Described information classification device. 5. The information classification apparatus according to claim 1, further comprising a specification unit that specifies a classification target portion in the classification sample information and the new text group. 6. A clustering means for clustering a plurality of sample texts to generate the classification sample information in which the plurality of sample texts are associated with a plurality of classification categories. The information classification device according to any one of Items 1 to 5, 7. A clustering means for clustering the classified sample information, a comparing means for comparing a clustering result of the clustering means with a desired clustering result, and, if necessary, based on a comparison result of the comparing means. The information classification device according to claim 1, further comprising a change unit configured to change the classification sample information. 8. The information classification according to claim 1, further comprising a clustering result display unit that clusters a new text group in the classification result of the classification unit and displays the clustering result. apparatus. 9. An optimizing unit for optimizing a classification result of the classification unit, wherein the classification learning information generating unit regenerates classification learning information based on the optimized classification result,
The method according to claim 1, wherein the classification unit classifies the new text group to be classified for each classification category according to the classification method determined by the classification method determination unit and the regenerated classification learning information. 8. The information classification device according to any one of 8. 10. A difference recognition information display means for displaying a difference between the classification result before optimization and the classification result after optimization as difference recognition information that can be visually recognized. 9. The information classification device according to item 9. 11. A feature element extracting step of extracting a feature element for each classification category from each of a plurality of sample texts included in classification sample information in which a plurality of sample texts and a plurality of classification categories are associated in advance, Based on the classification sample information, a classification method determination step of determining the classification method with the highest classification accuracy from among a plurality of classification methods, according to the classification method determined in the classification method determination step,
A classification learning information generating step of generating classification learning information representing a feature for each classification category based on the characteristic element extracted in the characteristic element extraction step; a classification method determined in the classification method determining step; and the classification learning A classification step of classifying a new text group to be classified according to the information for each classification category. 12. A feature element extracting step of extracting a feature element for each classification category from each of a plurality of sample texts included in classification sample information in which a plurality of sample texts and a plurality of classification categories are associated in advance, Based on the classification sample information, a classification method determining step of determining the classification method with the highest classification accuracy from a plurality of classification methods, according to the classification method determined in the classification method determination step,
A classification learning information generation step of generating classification learning information representing a feature for each classification category based on the feature element extracted in the characteristic element extraction step; a classification method determined in the classification method determination step; and the classification learning A computer-readable recording medium that records an information classification program for causing a computer to execute a classification step of classifying a new text group to be classified into classification categories according to information; 13. A feature element extraction procedure for extracting a feature element for each classification category from each of a plurality of sample texts included in classification sample information in which a plurality of sample texts and a plurality of classification categories are associated in advance; Based on the classification sample information, a classification method determination procedure for determining the classification method with the highest classification accuracy from among a plurality of classification methods, and according to the classification method determined in the classification method determination procedure,
A classification learning information generation step of generating classification learning information representing a feature for each classification category based on the feature element extracted in the characteristic element extraction step; a classification method determined in the classification method determination step; and the classification learning An information classification program characterized by causing a computer to execute a classification procedure for classifying a new text group to be classified according to information according to classification categories.