JP2000011004A - Automatic information classifying method, its device and recording medium recording its method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve classification precision by automatically preparing a new category (temporary category) for an erroneously classified document, assigning it there and classifying it through the use of an inputted category and a temporary category for it. SOLUTION: As of a document which is not correctly classified e.g. a document which is not classified into 'a category concerning 'life insurance'' erroneously, a temporary category (temporary category against 'the life insurance') is generated and the all are assigned there. When a newspaper article indicating 'the stock of a nonlife insurance company ascends and, in another insurance company...' exists if there are may documents concerning 'nonlife insurance', the value of a similarity degree of 'the temporary category for 'life insurance ''is larger than that of 'the category concerning 'life insurance'' when the similarity degree is calculated by using a feature vector of the document. And the result, the document which is erroneously classified into 'the category concerning 'life insurance'' is assigned to 'the temporary category for 'the life insurance'' this time and is not assigned to 'the category concerning 'life insurance''.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for automatically classifying a large amount of document data, and a recording medium on which the method is recorded.

[0002]

2. Description of the Related Art In recent years, it has become possible to exchange a large amount of electronic documents through a computer network such as the Internet. For this reason, services that allow individuals to search for information they need are implemented on networks. However, the amount of information acquired by the user becomes large, and it becomes difficult to extract characteristics of individual information. Therefore, a technique for classifying and organizing the acquired information is required.

Conventionally, research has been conducted on a method of automatically classifying document information. A typical method is to classify information into categories that are considered appropriate for new information, with known breaks (called categories) for classifying like a library (" Automatic text classification "Yamamoto, Masuyama (Toyohashi University of Technology)
Naito (NTT), 1995 Natural Language Processing Society of Japan 3
Month) or a method in which the category to be classified is unknown, and similar documents are collected from a set of documents to create and assign a classification category ("Automatic division by competitive learning neural network" Kikuchi, Matsuoka et al. (Utsunomiya Univ.), 1995 IEICE Transactions October). With these technologies,
A large number of documents are sorted and arranged.

[0004] The classification method to which the present invention is directed is the former method in which a category for classification is known in advance. Given a category to classify in advance and a sample document or word to be included in the system, the system calculates the importance of the word from the information, and as a feature of the category, the word and the importance for the category Produces a pair of vectors. Similarly, for the document to be classified, the importance is calculated for the word and the document, and a vector is generated ("Automatic TextProcessin
g "Gerard Salton, 1989 Addison-Wesley pub.co.).

The feature vector Wi (vector) generated for each category i is: Wi (vector) = (wi ₁ ,..., W _ij ,..., Wi _N ) (1) w _ij. The importance of the word j with respect to i is (2). Incidentally, N represents the number of dimensions. In calculating the value of each element of the feature vector Wi (vector), the frequency of occurrence of words belonging to each category is compared, and the value of a word existing in many categories is small, and the value of a word appearing biased in a certain category is small. The value is calculated so as to increase.

The classification defines a distance between a feature vector of a category and a feature vector of a document, and assigns each document to a similar category using the value. If it is determined that the distance is very large, that is, it is not similar to any category, no assignment is made to any category.

[0007]

In the conventional method, when a homonymous word is included, a feature vector including an appropriate word cannot be generated, and classification is performed using such a feature vector. In some cases, documents that were not appropriate for the category were classified.

According to the present invention, a document that is correctly classified is assigned to a correctly classified category by using the result of classification, and a document that is incorrectly classified is automatically created and assigned to a new category (temporary category) and input. It is an object of the present invention to improve the classification accuracy by performing classification using the determined category and the provisional category corresponding thereto.

[0009]

[Means for Solving the Problems] Some documents are classified into predetermined categories. as a result,
Some documents may not fall into the appropriate category.

In the present invention, a temporary category is automatically created and assigned to a document that has not been correctly classified. This temporary category is generated for a category that has not been correctly classified. When performing the classification process, the similarity between a category, a provisional category, and a document is calculated, and assigned to one of the similar categories. One temporary category may be generated for a plurality of categories.

[0011] The cause of incorrect classification is that similar words are used but the contents are different. Even if a document contains a group of similar words by this operation,
By generating another category and generating a feature vector there, words other than the similar word group can be extracted as feature vectors of the category, thereby improving classification accuracy.

[0012]

Embodiments of the present invention will be described below with reference to the drawings. In all the drawings describing the embodiments, the same reference numerals are given to the same elements, and the repeated description thereof will be omitted.

FIG. 1 is a block diagram showing a schematic configuration of an automatic information classification apparatus according to an embodiment of the present invention. Reference numeral 101 denotes a document input unit; 102, a document storage unit;
Is a category information input unit, 104 is a category information storage unit,
105 is a classification result input unit, 106 is a classification result storage unit, 1
Reference numeral 07 denotes a provisional category generation unit, and reference numeral 108 denotes a classification calculation unit.

(Description of Each Part) In the automatic information classification apparatus according to the embodiment of the present invention, a document to be processed is input to the document input unit 1.
Enter with 01. The documents to be processed include all the documents input to the computer. For example, there are newspaper articles, HTML files on the Internet, net news, text broadcasting, FM multiplex broadcasting, and broadcast manuscripts on television. Further, the document storage unit 102 stores the information input in the document input unit 101 so that the document of each information transmission medium, some of the information transmission media, or all the information transmission media can be retrieved.

In the category information input unit 103, a category to be classified by the system user and document information and keywords to be sampled into the category are input. For example, to collect information on sports related to high school baseball, a category related to “high school baseball” and documents and keywords related to high school baseball collected from newspaper articles and the Internet are input.

The category information storage unit 104 stores information input by the category information input unit 103, calculates a feature vector from the input information, and stores the result.

The classification result input unit 105 inputs the results of some classifications. Enter the correctly classified document and the category name to which it was assigned, and the incorrectly classified document and the category name to which it was assigned. For example, if there is a result of classification into a category relating to "high school baseball", and a document relating to "professional baseball" is mistakenly classified as high school baseball, a document relating to "professional baseball" will be regarded as a document that has not been correctly classified. And that it was not correctly classified is a category relating to "high school baseball".

A temporary category generation unit 107 generates a new temporary category for a document that has not been correctly classified,
Assign there. The tentative category and document information thus generated are stored in the category information storage unit 104. Thus, the category information storage unit 104 generates and stores the feature vector of the temporary category. At this time,
The category name that is not correctly classified and that is paired with the temporary category is also stored. When a correctly classified document is input from the classification result input unit 105, the document is stored in the category information storage unit 104 in combination with the category information input by the category information input unit 103.

In the classification calculation unit 108, the document input unit 101
A process of allocating the input document to the category and the temporary category is performed. If there is a category in which a temporary category exists, the classification process does not classify into both the temporary category and the category, but classifies into either the category or the temporary category corresponding thereto. The classification for the category other than the temporary category may be a classification criterion that is assigned to a plurality of categories.

(Explanation of Processing Using Specific Example) The flow of the processing of the present invention will be specifically described using an example. FIG. 2 shows an example of a classification process when a temporary category is not generated, and FIG. 3 shows an example of a classification process when a temporary category is generated.

The document to be input to the system is a newspaper article of a newspaper company for the past ten years. The document set is sent to the document input unit 101
Enter more. Next, from the category information input unit 103,
Enter the category you want to collect and the sample documents and words that go into it. For example, enter "categories related to" life insurance "" or "categories related to" typhoon "", and as a sample document to be included in each category, some articles related to "life insurance" in the past
In the “category related to“ typhoon ”, the user inputs information on past“ typhoon ”and articles on damage caused by the typhoon to“ category related to “typhoon” ”.

The category information storage unit 104 creates and stores a feature vector of each category from the information input by the category information input unit 103. For example, in the “category about“ life insurance ””, ((life insurance, 0.23), (insurance, 0.13), (damage, 0.10), (liberalization, 0.08), (Company A, 0.05), ...) (3) Generate and store. In the “category about“ typhoon ””, ((typhoon, 0.25), (damage, 0.21), (landing, 0.18), (strength, 0.14), (roof) , 0.10), ...) (4) Generate and store.

Next, past results classified into the category input by the category information input unit 103 are input from the classification result input unit 105. For example, "life insurance"
In the "category about", words such as "life insurance", "life" and "insurance" are acquired as feature vectors.
Due to the word “insurance” in this feature vector, for example, as shown in FIG. 2, a document relating to “life insurance” may be classified into a “category relating to“ life insurance ””. That is, in FIG. 2, it is classified into a “category relating to“ life insurance ””. If a past example or a document example that could not be classified correctly by classifying some samples could be extracted, enter such a document as an example that was not correctly classified in the “category related to“ life insurance ”” .

The provisional category generation unit 107 generates a new provisional category based on the input from the classification result input unit 105, and assigns a document that has not been correctly classified. For example, for a document that has not been erroneously classified as a “category relating to“ life insurance ””, a temporary category is generated as shown in FIG. 3 (this will be referred to as a “temporary category for“ life insurance ”). All assigned to.

The temporary category generated by the temporary category generation unit 107 and the document information are stored in the category information storage unit 104, and a feature vector is generated. For example, in the “temporary category for“ life insurance ””, if there are many documents related to “non-life insurance” as documents that were not correctly classified, the feature vector is ((non-life insurance, 0.23), (insurance, 0.22), ( (Guarantee, 0.11), (life, 0.09), (disaster, 0.09), ・ ・ ・) (5)

Thereafter, the classification calculation unit 108 performs a classification process using the categories and their feature vectors stored in the category information storage unit 104. At this time, in the classification process into the “category of“ life insurance ””, the “temporary category for“ life insurance ”” and the similarity of the document are calculated and classified into one of the categories. Now, for example, if there is a newspaper article such as "The stock price of a non-life insurance company has risen and other insurance companies ...", the feature vector of this document is ((General Insurance, 0.18), (Insurance, 0.16 ), (Stock price, 0.12), (performance, 0.11), ...) (6) Assuming that the inner product of the vectors is used as a function of the similarity between the category and the document, in this case, the “temporary category for“ life insurance ”” has a larger similarity value. As a result, as shown in FIG. 3, when there is no provisional category, which is the conventional method (in the case of FIG. 2), a document incorrectly classified as “category related to“ life insurance ”” And is no longer assigned to the "category for" life insurance "".

By extracting information relating to the provisional category, it is possible to grasp the category that conflicts with the category to be classified, and it is also possible to artificially input the category to be collected from the category information input unit 103. Also, using the information of the provisional category and the result of the classification,
By manually assigning the category name, the category can be handled in the same manner as the category input by the category information input unit 103 instead of the temporary category. Thereby, new document information and a new category can be obtained.

FIG. 4 shows a processing flowchart of one embodiment of the present invention. Step (S1): Create a feature vector for a document to be classified. Step (S2): Calculate the degree of similarity with a certain category A.

Step (S3): It is checked whether or not it is similar to category A. If they are similar, step (S5)
If not, the process proceeds to step (S4). Step (S4): Return to step (S2) for calculation with the next category.

Step (S5): It is checked whether or not there is a temporary category for category A. If there is a step (S
Go to 7), otherwise go to step (S6). Step (S6): Assign to category A, and go to step (S2) for calculation with the next category.

Step (S7): It is checked whether or not it is similar to the temporary category. If they are similar, step (S9)
If not, the process proceeds to step (S8). Step (S8): Assigned to category A, return to step (S2) for calculation with the next category.

Step (S9): Assign to a temporary category (do not assign to category A). Step (S10): Check whether there is another category. If there is, the process returns to step (S2). If not, END
Becomes

In the above, the automatic information classification method and the classification device have been described. However, the processing relating to the classification can be recorded in a recording medium in the form of a program. For this reason, the present invention includes the recording medium.

[0034]

According to the present invention, when a large set of documents is classified into a large number of classification categories, classification accuracy can be improved by giving a result of the classification processing as a sample.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a schematic configuration of an information automatic classification device according to an embodiment of the present invention.

FIG. 2 is an example of a classification process of the information automatic classification device when a temporary category is not generated.

FIG. 3 is an example of a classification process of the information automatic classification device when a temporary category is generated.

FIG. 4 shows a processing flowchart of an embodiment of the present invention.

[Explanation of symbols]

 101: document input unit 102: document storage unit 103: category information input unit 104: category information storage unit 105: classification result input unit 106: classification result storage unit 107: temporary category generation unit 108: classification calculation unit

Continuing on the front page (72) Inventor Kazuhiro Hayakawa 3-19-2 Nishi Shinjuku, Shinjuku-ku, Tokyo Nippon Telegraph and Telephone Corporation (72) Inventor Kazuo Tanaka 3-192-2 Nishi-Shinjuku, Shinjuku-ku, Tokyo Nippon Telegraph and Telephone Telephone Co., Ltd. F term (reference) 5B009 SA12 VA02 VA09 5B075 ND03 NK12 NK35 NR03 NR12 PR06 QM08 QS01 UU06

Claims (6)

[Claims] 1. An automatic information classification method for automatically classifying a large amount of information into a large number of categories, a first step of inputting a document described in a natural language, and storing input data obtained in the first step. A second step of inputting a category for classification and its characteristics; a fourth step of storing the category and its characteristics obtained in the third step; a correctly classified document A fifth step of inputting a document that has not been correctly classified, a sixth step of storing the document obtained in the fifth step, and a correct classification using the document obtained in the fifth step. The classified document is added as information of the category input in the third step, and the document that is not correctly classified is
The 7th to automatically generate a new temporary category and assign it to it
And an eighth step of classifying the document obtained in the first step using the information on the category obtained in the third step and the information on the temporary category obtained in the seventh step. An automatic information classification method, comprising: 2. The method according to claim 1, wherein in the seventh step, the fifth
Using the document obtained in the step, the document that is not correctly classified is assigned to a temporary category, and the document obtained in the first step in the eighth step is divided into the category obtained in the third step and the category obtained in the third step. 2. The information automatic classification method according to claim 1, wherein the information is classified into any one of provisional categories. 3. An automatic information classification apparatus for automatically classifying a large amount of information into a large number of categories, comprising: a document input unit for inputting a document described in a natural language; and a document for storing input data obtained by the document input unit. A storage unit; a category information input unit for inputting a category for classification and its characteristics; a category information storage unit for storing the categories and the characteristics obtained by the category information input unit; A classification result input unit that inputs a document that has not been classified, a classification result storage unit that stores a document obtained by the classification result input unit, and a document that is correctly classified using the document obtained by the classification result input unit Is added as information on the category input in the category information input section, and for documents that have not been correctly classified, a temporary category is automatically generated and a new temporary category is assigned thereto. A glue generation unit, a classification calculation unit that classifies the document obtained by the document input unit using the information on the category obtained by the category information input unit and the information on the temporary category obtained by the temporary category generation unit, An automatic information classification apparatus, comprising: 4. The provisional category generation unit assigns a document that has not been correctly classified to a provisional category using a document obtained by the classification result input unit, and a document obtained by the classification calculation unit by the document input unit. 4. The automatic information classification apparatus according to claim 3, wherein the information is classified into one of a category obtained by the category information input unit and a temporary category corresponding to the category. 5. A first step of receiving a document described in a natural language on a recording medium recording an information automatic classification method for automatically classifying a large amount of information into a large number of categories; A second step of storing input data, a third step of inputting a category for classification and its characteristics, a fourth step of storing the category obtained in the third step and its characteristics, A fifth step of inputting a classified document and a document that has not been correctly classified; a sixth step of storing the document obtained in the fifth step; and a document processing apparatus using the document obtained in the fifth step. The correctly classified documents are added as information of the category input in the third step, and the documents that are not correctly classified are
The 7th to automatically generate a new temporary category and assign it to it
And an eighth step of classifying the document obtained in the first step using the information on the category obtained in the third step and the information on the temporary category obtained in the seventh step. A recording medium recording an information automatic classification method characterized by recording a described processing program. 6. The method according to claim 7, wherein in the seventh step, the fifth
Using the document obtained in the step, the document that is not correctly classified is assigned to a temporary category, and the document obtained in the first step in the eighth step is divided into the category obtained in the third step and the category obtained in the third step. 6. A recording medium recording the information automatic classification method according to claim 5, wherein a processing program adapted to classify the information into one of the provisional categories is recorded.