JP2978044B2 - Document classification device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a document classification device used for a document automatic classification machine for storing / automatically classifying documents and a word processor / filing system.

[0002]

2. Description of the Related Art Conventionally, it has been difficult to automatically classify documents, and a user manually classifies documents or extracts keywords in a document and classifies them using a thesaurus created in advance. Even in a system called automatic classification, basic data for classification had to be manually input in the form of a basic example sentence or the like.

[0003]

However, in such a classification, manual work becomes a bottleneck.
Classifying a large number of documents is very difficult.

[0004] The present invention has been made in view of the above circumstances, and has as its object to provide a document classification device that automatically classifies documents without manual intervention.

[0005]

The invention according to claim 1 is
In the document classification device, a storage unit that stores the document data, a document analysis unit that analyzes the document data, and automatically generates a feature vector that expresses a feature of each word using a co-occurrence relationship between words in the document. Word vector generation unit, a word vector storage unit that stores the feature vector, a document vector generation unit that generates a document feature vector from the word feature vector included in the document, and stores the feature vector A document vector storage unit, a classification unit for classifying documents using similarity between the feature vectors of the documents, a result storage unit for storing the results of the classification, and words used when generating feature vectors are registered. A dictionary for feature vector generation, and automatically generates feature vectors representing the features of each word using co-occurrence relationships between words in a large number of document files, Automatically, characterized in that can be classified.

The invention according to claim 2 further comprises, in addition to the above configuration, a useful word selecting section for selecting useful words at the time of classification by using the classification result stored in the result storage section. After classifying the document file, by examining the occurrence rate of words for each classified group, words useful for classification are selected, and classification accuracy is improved by using only words useful for classification. It is characterized by being able to.

According to a third aspect of the present invention, in addition to the above configuration, a representative vector generating unit for obtaining a feature vector representing each taxonomic group using the classification result stored in the result storage unit, The apparatus further includes a representative vector storage unit that stores the representative vector generated by the generation unit, and after classifying a large number of document files, uses the feature vector of each classified group and the feature vector of the document to classify the field. It is characterized in that a representative feature vector can be obtained.

[0008]

The operation at the time of learning the feature vector of a word according to claim 1 will be described. The contents of a large amount of document files stored in the document storage unit are passed to the document analysis unit and sentence analysis (morphological analysis, etc.) is performed, and the word vector generation unit analyzes the co-occurrence relationship and appearance frequency of words. To generate a feature vector for each word. The feature vector of the word thus generated is stored in the word vector storage. In this way, the learning of the word feature vector is performed. By limiting the words for generating feature vectors to words registered in the feature vector generation dictionary, the storage space for feature vectors is prevented from becoming too large.

The operation at the time of classifying documents according to claim 1 will be described. When classifying sentences, the contents of the document file stored in the document storage unit are passed to the document analysis unit to analyze the sentence (morphological analysis, etc.), and the document vector generation unit analyzes the sentence in the document analysis unit. The feature vector of the word that appears when the analysis is performed is obtained by referring to the word vector storage unit, and the feature vector of the document is generated from the feature vector of the word included in the document. The document feature vector generated in this way is stored in the document vector storage unit, and the classification unit classifies the document based on the similarity between the feature vectors of the document. This classification result is stored in the result storage unit.

According to the second aspect of the present invention, after classifying a large number of documents, a useful word selecting unit selects useful words at the time of classification by using the classification results stored in the result storage unit. I do. Registering only words selected by the useful word selection unit in the dictionary for generating feature vectors, and then learning the word feature vectors again, and re-classifying using the obtained word feature vectors Accordingly, it is possible to reduce the storage space of the feature vector as compared with the configuration of the first aspect, and to increase the accuracy of classification.

According to a third aspect of the present invention, after a large number of documents are classified, the representative vector generation unit uses the classification results stored in the result storage unit to represent each classification group. Find a vector. The representative vector generated by the representative vector generation unit is stored in the representative vector storage unit. Once you have generated a representative vector for each taxon,
When classifying new document data, it is possible to determine to which classification group the document belongs simply by comparing the feature vector of the document with the representative vector of each classification group.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 shows an embodiment of the first aspect of the present invention. The document classification device includes a storage unit 10 for storing document data.
1, a document analysis unit 102 that analyzes document data, and a word vector generation unit 1 that automatically generates a feature vector that expresses a feature of each word using a co-occurrence relationship between words in the document.
03, a word vector storage unit 104 that stores the feature vector, a document vector generation unit 105 that generates a document feature vector from the word feature vector included in the document, and a document vector that stores the feature vector A storage unit 106, a classification unit 107 for classifying documents using similarity between the feature vectors of the documents, a result storage unit 108 for storing the results of the classification, and words used when generating feature vectors are registered. Dictionary for generating feature vectors 1
09.

In general, the number of words used in an ordinary document is very large, and it is more realistic to limit the words used when creating a feature vector. The feature vector generation dictionary 109 is used for this purpose. By creating a feature vector of a word using only the words registered in the dictionary 109, it is possible to suppress an increase in the storage space of the feature vector.

FIG. 2 shows a system configuration at the time of learning a word feature vector. At the time of learning a word feature vector, a large amount of document data for learning is stored in the document storage unit 101 and read out from the document storage unit 101. The document data is read into the document analysis unit 102 for each appropriate chunk such as an article, a paragraph, and a sentence, and the document analysis unit 102 analyzes the document data to extract words. A word feature vector is generated by the word vector generation unit 103 based on the word string extracted here, and the word feature vector generated by the word vector generation unit 103 is stored in the word vector storage unit 104. Thus, the feature vector of the word is learned.

FIG. 3 shows a system configuration at the time of document classification.
When classifying a document, the data of the document to be classified is stored in the document storage unit 101, and the document data read from the document storage unit 101 is stored in the document analysis unit 102 for each unit (eg, article unit) to be classified. And the document analysis unit 102 analyzes the document data to extract words. The extracted feature vector of the word is 10
4 with reference to the contents of the word vector storage unit. Usually, a plurality of words are extracted from one unit (eg, one article) of the document data. At this time, the feature vector of the document is calculated by averaging the values of the feature vectors of all the words obtained. At this time, instead of simply averaging, each feature vector is weighted according to the reciprocal of its appearance frequency (for example, the number of articles in which the word appears from a large number of articles is investigated, and log A better value may be obtained by averaging (after multiplying the total number of articles / the number of articles in which the word appears) by the feature vector of the word.

When the feature vector of the document is obtained, the document can be classified by applying the conventional clustering technique. For example, it may be considered that documents having a short distance between the feature vectors of the documents belong to the same field.

Further, a human selects a typical document for each taxonomy, generates a temporary representative vector of the taxonomy from a feature vector of a word extracted from the document, and reads it from the document storage unit 101. Documents can also be classified according to which taxonomic feature vector of the document to be read is close to a temporary representative vector of which taxonomy. Even with such a classification method, the document storage unit 10
If a large amount of document data is read from one, the influence of an error that a human selects a temporary representative vector is reduced, and a fairly general representative vector for each field can be finally generated.

Now, a method for generating a word feature vector will be specifically described. The feature vector of a word is obtained by adding a value obtained by multiplying the frequency of appearance of a word contained in a piece of document data by the frequency of occurrence of the word in the piece of document data. . A specific example will be described.

Example A: "The US government has proposed a drastic review of Cocom regulations to major developed nations." Example B: Provided that regulated countries regulate the export of industrial products that lead to the manufacture of weapons, It seems that Kocom intends to significantly reduce the list of regulated items. " Here, the document data is read in units of “one sentence”, but this may be another unit such as one article.

The number of dimensions of the feature vector is 21 (the number of words registered in the feature vector generation dictionary is 21)
Each element is "U.S., government, advanced, major, national, cocom, regulation, drastic, review, proposal, target, weapons, manufacturing,
Industry, product, export, condition, item, significant, reduction, intention ".

Under these conditions, when the example sentence A is read from the document storage unit 101, the document analysis unit 102 analyzes the sentence A and reads "America, Government, Advanced, Major, Country, Cocom,
Regulations, drastic, reviews, proposals "are extracted. At this time, the word vector generation unit 103 adds 1 to the elements corresponding to these words in the feature vectors of all these words. Then, the feature vectors of words appearing in the example sentence A, such as “America” and “Government,” are (1,1,1,1,1,1,1).
1,1,1,1,0,0,0,0,0,0,0,0,
(0,0,0). FIG. 8 illustrates this.

Next, when the example sentence B is read from the document storage unit 101, it is analyzed by the document analysis unit 102 and "regulation, target, country, weapon, manufacture, industry, product, export, regulation, condition,
COCOM, regulation, item, significant, reduction, intention "are extracted.

The word appearance frequency distribution obtained from this is (0,0,0,0,1,1,3,0,0,0,1,1,1).
1,1,1,1,1,1,1,1,1,1). Since "Regulation" appears three times, it is a vector that triples this word appearance frequency distribution (0, 0, 0, 0, 3, 3, 3).
9,0,0,0,3,3,3,3,3,3,3,3
(3,3,3) is added to the feature vector of “regulation”, and the feature vectors of words that appear only once in example sentence B, such as “target” and “country”, are (0,0,0,0,1,1, 1,3,0,0,
0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1). FIG. 9 illustrates this.

Since the size of a vector to be added changes depending on the length of a sentence in the method of always adding integers, the absolute value of the vector to be added is normalized to 1 or the frequency A method of normalizing the absolute value of the vector to 1, multiplying the value by a value proportional to the number of appearances, and then adding the values is also conceivable.

The finally obtained feature vector is
The absolute value is normalized to 1.

The word feature vector thus obtained is stored in the word vector storage unit 104, and is used at the time of document classification.

Next, the process of generating a feature vector of a document at the time of document classification will be described as a specific example when the following example sentence C is read.

Example sentence C "The US government has proposed reduction of weapons." When the example sentence C is read from the document storage unit 101, it is analyzed by the document analysis unit 102 and "USA, government, weapons, reduction, proposal" is extracted. Is done. Then, the document vector generation unit 1
In step 05, the feature vectors of words appearing in the example sentence C, such as "America" and "government", are added with reference to the contents of the word vector storage unit 104, and (3, 3, 3, 3 , 5,5,9,3,3,3,2,2
2, 2, 2, 2, 2, 2, 2, 2, 2). FIG. 10 illustrates this. Although vector normalization is not performed in FIG. 10 in order to give priority to clarity, in actual processing, the absolute value of the feature vector of each word is normalized to 1 before addition, and then addition is performed. The obtained feature vector is stored in the document vector storage group 106.

Next, a description will be given of how the classifying unit 107 uses the document feature vector at the time of classification. In brief, first, the absolute value of the obtained feature vector of the document is normalized to 1 and then classified using a conventional method such as the K-means method, or the similarity of the taxonomic group to the (temporary) representative vector. It is sufficient to classify by the degree (obtained by calculating the distance or calculating the inner product), but the feature vector obtained by this method has the characteristic that the value of the element corresponding to the word that appears frequently becomes very large. For this reason, it is often better to devise such a feature so that it does not adversely affect the classification. For example, when calculating the distance, calculate the distance so that the difference between the elements does not expand (usually calculate the square root of the sum of the squares of the differences between the elements. For example, calculate the sum of the absolute values of the differences between the elements) It is better to use the distance obtained by performing a calculation or the like, and before calculating the inner product, log all the elements or take the root of a power and then normalize the values to calculate the values.

As a specific example of classification, there are three classification groups.
Suppose that the representative vector of each taxon was obtained as follows.

The representative vector (1,1,1,1,
1,0,0,0,0,0,0,0,0,0,0,0,
0, 0, 1, 1, 1, 1) The representative vector (1,1,1,1,1,1,1,
1,1,1,1,5,5,5,5,5,5,5,5
5,5,5) Representative vector of taxon 3 (4,4,4,4,6,6,6)
6,3,3,3,1,1,1,1,1,1,1,1,1,
1,1,1) As a measure of the degree of similarity, the absolute value of both the feature vector of a document and the representative vector of a taxonomic group is normalized to 1 and then the inner product of the two is calculated. If the degree is high, the feature vector of the example sentence C

[0033]

(Equation 1)

(3,3,3,3,5,5,9,3,3
3,2,2,2,2,2,2,2,2,2,2,2) Representative vector of taxon 1

[0035]

(Equation 2)

(1,1,1,1,0,0,0,0,0,
0,0,0,0,0,0,0,0,1,1,1,1) Representative vector of taxon 2

[0037]

(Equation 3)

(1,1,1,1,1,1,1,1,1,1,
1,5,5,5,5,5,5,5,5,5,5,5) Representative vector of taxon 3

[0039]

(Equation 4)

(4,4,4,4,6,6,6,3,3,3
3,1,1,1,1,1,1,1,1,1,1,1,1) Therefore, the inner product of the feature vector of the example sentence C and the representative vector of each taxon is the dot product of the taxon1.

[0041]

(Equation 5)

Dot product with taxon 2

[0043]

(Equation 6)

Dot product with taxon 3

[0045]

(Equation 7)

The feature vector of the example sentence C is
, The example sentence C is classified into the classification group 3. This is illustrated in FIG. In FIG. 11, as in FIG. 10, the vectors are not normalized in order to give priority to simplicity. However, in the actual processing, the absolute value of each vector is normalized to 1 before comparison, and then the comparison is performed. The result of the classification is stored in the result storage unit 108.
Is stored.

Next, a second embodiment of the present invention will be described with reference to FIG.
Shown in Here, those represented by reference numerals 201 to 209 are the same as those represented by reference numerals 101 to 109 in FIG.

The document classification device includes a storage unit 201 for storing document data, and a document analysis unit 202 for analyzing document data.
A word vector generation unit 203 that automatically generates a feature vector that expresses a feature of each word using a co-occurrence relationship between words in a document, a word vector storage unit 204 that stores the feature vector, A document vector generation unit 205 that generates a document feature vector from a feature vector of a word included in a document, a document vector storage unit 206 that stores the feature vector, and a similarity between the feature vectors of the document. A classification unit 207 for classifying documents, a result storage unit 208 for storing the classified results, a feature vector generation dictionary 209 in which words used at the time of generating feature vectors are registered, and a result storage unit 208 And a useful word selecting unit 210 for selecting a useful word at the time of classification using the classified result.

FIG. 5 is a diagram showing a system configuration at the time of learning and at the time of classification. First, a feature vector of a word is learned by the same method as in the first embodiment, and a large amount of document data is classified based on the learned feature vector. The result of the classification is stored in the result storage unit 208. Based on the result, the useful word selection unit 210 selects a useful word. this is,
The frequency of each word is obtained for each taxon, and words included in every taxon at a similar ratio are removed (method 1:
Remove those that have a ratio of the highest frequency to the lowest frequency below a certain threshold) or select those that are contained only in a certain group at a high rate (Method 2: There is a ratio between the highest frequency and the second highest frequency) Those with a threshold value or more are selected). In addition, useful word selection unit 2
The words to be selected in 10 need not necessarily be words from the words registered in the feature vector generation dictionary 209,
Selection can be made from a wider range of words.

As a specific example, it is assumed that there are three classification groups a, b, and c, and that the words registered in the feature vector generation dictionary 209 are three, "politics, Japan, and international". Then, for each taxonomic group, each word (in addition to the words registered in the feature vector generation dictionary 209, "election", "problem"
The frequency is also determined as follows.

Taxa a Political 30%, Japan 5%, International 35%, Election 10%, Problem 20% Taxa b Politics 3%, Japan 55%, International 35%, Election 2%, Problem 5% Taxa c Politics 3%, Japan 30%, international 35%, election 2%, problem 30% If you use method 1, "international" is included in every taxon in the same ratio, so it is used for feature vector generation. Will be removed from the dictionary. Since "politics", "Japan", "elections" and "problems" are uneven in the frequency of each taxon, they are selected as useful words, and the feature vector generation dictionary 209 is used.
(If it is desired to reduce the number of registered words at this time, it is only necessary to take the number of words to be registered in the order of the total appearance frequency among words having a biased frequency). When the method 2 is used, only “politics” and “election” are selected and registered in the feature vector generation dictionary 209, and “Japan”, “international” and “problem” are registered.
Are not registered in the feature vector generation dictionary 209. As an intermediate method between Method 1 and Method 2,
A method of selecting useful words depending on whether or not the ratio of the frequency of the rank (n is 3 or more, and the number of taxon groups minus 1) is equal to or greater than a certain threshold is also conceivable. In addition, a method is also conceivable in which a value having a large frequency variance instead of a frequency ratio is selected.

The words selected in this manner can be considered to have importance according to the frequency ratio (or frequency variance). Therefore, when calculating the feature vector of a document, Is weighted according to this ratio (or variance) and then averaged (for example, after multiplying the log (frequency ratio) by that feature vector), a better document feature vector location is obtained. May be

In this way, the feature vector generation dictionary 209
By registering only words that are useful for classification, learning the feature vectors of words again, and classifying documents using them, the dictionary for generating feature vectors can be made smaller, and the accuracy of classification can be improved. .

FIG. 6 shows a third embodiment of the present invention. Here, those represented by reference numerals 301 to 310 are shown in FIG.
Are the same as those represented by reference numerals 201 to 210, respectively.

The document classification device includes a storage unit 301 for storing document data and a document analysis unit 302 for analyzing the document data.
A word vector generation unit 303 that automatically generates a feature vector that expresses a feature of each word using a co-occurrence relationship between words in a document, a word vector storage unit 304 that stores the feature vector, A document vector generation unit 305 that generates a document feature vector from a feature vector of a word included in a document, a document vector storage unit 306 that stores the feature vector, and a similarity between the feature vectors of the document. A classifying unit 307 for classifying documents, a result storage unit 308 for storing the classified results, a dictionary 309 for generating feature vectors in which words used at the time of generating feature vectors are registered, and a result storage unit 308 A useful word selection unit 310 for selecting a useful word at the time of classification using the classification result, and a classification result stored in the result storage unit 308. A representative vector generation section 311 for obtaining a feature vector representative of each taxon, representatives from vector storage unit 312. storing the representative vectors generated by the representative vector generation unit 311.

When the system according to the third aspect is configured using the embodiment of the first aspect, the system does not include the useful word selecting unit 310.

FIG. 7 is a diagram showing a system configuration at the time of learning and at the time of classification. At first, a feature vector of a word is learned by the same method as in the embodiment of the first and second embodiments, and a large amount of document data is classified based on the learned feature vector. The classified result is stored in the result storage unit 308, and a representative vector is generated by the representative vector generation unit 311 based on the result. This can be generated, for example, by calculating the frequency of each word for each taxonomy, selecting words that are included only in a certain taxonomy at a high rate, and averaging the feature vectors of such words. As specific examples, the taxon is a,
Assuming that there are three, b and c, the dictionary for feature vector generation 3
Suppose that the words registered in 09 are three, “politics, parliament, and international”. Assume that the frequency of each word for each taxon is as follows.

Taxa a Politics 40%, Diet 50%, International 10% Taxa b Politics 10%, Diet 10%, International 80% Taxa c Politics 20%, Diet 10%, International 70% Then, Taxa a Is given as an average of the feature vector of "politics" and the feature vector of "diet." In addition, it is also conceivable to assign a weight not based on an average but on an appearance ratio. For example, if the frequency of occurrence of “politics” is twice the frequency of occurrence of “diet”, twice the feature vector of “politics” and the feature vector of “diet” are added, and the result obtained by dividing by 3 is the classification group a. And so on as a representative vector.

Similarly, a method in which the average of the feature vectors of the documents classified into the classification group a is taken as the representative vector of the classification group a can be considered.

In this manner, when the representative vector is generated, it is stored in the representative vector storage unit 312.
By referring to this representative vector at the time of subsequent document classification, the document read from the document storage unit 301 can be classified into a classification group corresponding to the representative vector most similar to the feature vector of the document. Become.

The present invention can be used not only for document classification, but also for automatically classifying e-mails and e-news, and selecting e-mails and e-news that are likely to be of interest to the user ( Fuzzy search (a document whose similarity between the feature vector of the search keyword and the feature vector of the document to be searched exceeds a certain threshold) Can be used to search for related keywords even if they do not exactly match the search keyword), or to select homophonic meaning words in kana-to-kanji conversion (from the converted content up to that point). Select the same meaning word based on the similarity to the feature vector
Select a conversion result that is most suitable for past contexts in speech recognition, handwritten character recognition, etc. (select recognition results based on similarity with feature vectors obtained from the content recognized so far) ), Or narrows the search space for words and the like at the time of recognition, etc. (Only words corresponding to elements that are above a certain threshold among the elements of the feature vector obtained from the content recognized so far) Search)).

[0062]

According to the present invention, a feature vector of a word can be automatically created, and a document can be automatically classified. In addition, the feature vector of the word created by this method can be used not only for document classification but also for fuzzy search and selection of homophonic meaning words in kana-kanji conversion, and in speech recognition, handwritten character recognition, etc. It can also be used when selecting a recognition result that best fits the context of.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a basic configuration of one embodiment of the invention according to claim 1;

FIG. 2 is a block diagram showing a system configuration at the time of learning of the system shown in FIG. 1;

FIG. 3 is a block diagram showing a system configuration when the system shown in FIG. 1 is classified.

FIG. 4 is a block diagram showing a basic configuration of one embodiment of the invention according to claim 2;

FIG. 5 is a block diagram showing a system configuration at the time of learning and classification of the system shown in FIG. 4;

FIG. 6 is a block diagram showing a basic configuration of one embodiment of the invention according to claim 3;

FIG. 7 is a block diagram showing a system configuration at the time of learning and classification of the system shown in FIG. 6;

FIG. 8 is a diagram illustrating generation of a word feature vector.

FIG. 9 is a diagram illustrating generation of a word feature vector.

FIG. 10 is a diagram illustrating generation of a feature vector of a document.

FIG. 11 is a diagram illustrating classification of documents.

[Explanation of symbols]

 101, 201, 301 Document storage unit 102, 202, 302 Document analysis unit 103, 203, 303 Word vector generation unit 104, 204, 304 Word vector storage unit 105, 205, 305 Document vector generation unit 106, 206, 306 Document vector Storage unit 107, 207, 308 Classification unit 108, 208, 308 Result storage unit 109, 209, 309 Feature vector generation dictionary 210, 310 Useful word selection unit 311 Representative vector generation unit 312 Representative vector storage unit

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-2-158871 (JP, A) JP-A-2-98778 (JP, A) JP-A-2-105973 (JP, A) JP-A-2-158 235176 (JP, A) JP-A-5-54037 (JP, A) Ikuo Akutoshi, Takao Inui, Kenichiro Ishikura, "Associative search from large-scale document database", IEICE Technical Report ( AI-92-99), Vol. 92, No. 426, 1993 (Heisei 5-1-22), pp. 73-80 Ikuo Akutoshi, Takao Inui, Toshiyuki Okunishi, "An attempt to construct a self-organizing encyclopedia database using semantic vectors", 1993 National Convention of the Japanese Society for Artificial Intelligence (7th) Papers, 1993 (Heisei 5-7-20) p. 317-320 (58) Field surveyed (Int.Cl. ⁶ , DB name) G06F 17/30 JICST file (JOIS)

Claims (3)

(57) [Claims] In a document classification device, a storage unit for storing document data, a document analysis unit for analyzing the document data,
A word vector generation unit that automatically generates a feature vector that expresses a feature of each word using a co-occurrence relationship between words in the document, a word vector storage unit that stores the feature vector, and a word vector storage unit that is included in the document. A document vector generation unit that generates a document feature vector from a feature vector of a word, a document vector storage unit that stores the feature vector, and a classification unit that classifies documents using similarity between the document feature vectors And a result storage unit for storing the classified results, and a dictionary for generating feature vectors in which words to be used when generating feature vectors are registered, using a co-occurrence relationship between words in a large number of document files. And a document classification apparatus capable of automatically generating a feature vector expressing the feature of each word and automatically classifying documents. 2. The document classification apparatus according to claim 1, further comprising: a useful word selecting section for selecting a useful word at the time of classification by using a classification result stored in a result storage section. After classifying a file, by examining the frequency of occurrence of words for each classified group, select words that are useful for classification, and improve the accuracy of classification by using only words that are useful for classification. A document classification device characterized by being able to do. 3. A representative vector generation unit for obtaining a feature vector representing each classification group using a classification result stored in a result storage unit, in addition to the configuration of the document classification device according to claim 1 or 2. And a representative vector storage unit for storing the representative vector generated by the representative vector generation unit, and after classifying a large number of document files, using the feature vectors of the words and documents of each of the classified groups. A document classification device capable of obtaining a feature vector representative of the field.