JPH06282587A - Automatic classifying method and device for document and dictionary preparing method and device for classification - Google Patents

Abstract

PURPOSE:To exactly attain automatic classification under the consideration of the meaning content of a text by extracting key words and simultaneously discriminating the classification of the key words, and making the key words same when the classification is same. CONSTITUTION:A document inputted through an inputting device 1 is transmitted to a key word extracting part 2, the kind(subject, object, and other) of an extracted noun is extracted, and the key words same up to the kind are treated as the same key words. The extracted key words are stored in a key word file 3, read, and applied to a character string matching key word extracting part 4. Then, the key words including the same characters in the sequence of the appearance of the characters constituting the basic key words are recognized to be the same, and stored in a character string matching key word file 5. Moreover, the number of documents in which the key word and a pair of key words appear is counted by count parts 6 and 7, transmitted to a featured value calculating part 8, and a feature for preparing a dictionary is searched.

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 documents and a method and apparatus for creating a dictionary for classification.

[0002]

BACKGROUND OF THE INVENTION When filing various documents and constructing a database, it is necessary to determine the classification and add keywords to the documents to be stored in the database. This is an important task because it leads to the convenience of using the database. However, if humans perform such classification, it is necessary to read and understand all the sentences before making a classification decision, which not only makes the work extremely complicated but also varies depending on the operator. (Reliability).

Therefore, various systems have been developed as systems for automatically classifying such documents. One of them is to create a semantic dictionary that represents a semantic relationship in a certain field, such as the intelligent search method disclosed in Japanese Patent Laid-Open No. 3-2009563, and to search for unknown data thereafter based on the dictionary. There is something to do.

However, in this method, it is necessary to form the semantic dictionary in advance by artificial analysis, and if the field is different, the analysis for the field is required again, and the processing concerned is complicated.

As another method, for example, as represented by "Fuzzy Document Retrieval System" (Information Processing Society of Japan 39th National Convention Proceedings, pages 1067-1068),
There is a search method that uses a dictionary that statistically processes the notational characteristics of documents in a specific field. According to this method, it is noted that the keyword pair commonly included in one document has some conceptual connection when the document is considered as an established concept, and the simultaneous appearance frequency value of the keyword pair is used for the document. The degree of familiarity between them is quantified (dictionaryized). The higher the degree of closeness, the more common the document can be certified. Therefore,
Create a dictionary based on documents whose fields are known in advance,
It seeks closeness between a new document and a document whose field is already known, and determines the field of the document with high closeness to be the field of the new document.

In this method, however, the dictionary can be constructed without human intervention as compared with the former method, and thus it is highly useful that the dictionary can be automatically created for different fields. Therefore, there is a problem that the generated affinity (fuzzy value) has low accuracy (the ambiguous range is wide). As a result, when attempting to classify more similar documents into appropriate fields, there is a possibility that they will fall within an ambiguous range and cannot be correctly classified.

Further, in the above-mentioned method that "the documents having a large number of common keyword pairs have a high degree of similarity between them (the higher the probability of belonging to a common field)", the same keyword for performing such determination is As it is based on perfect matching, for example, "○ x electric power" and its abbreviation "○ electric power" and "receipt" in the field related to electricity are used as they are.
Keywords that should be treated as having the same meaning, such as and “electricity receipt”, are treated as different keywords because of the difference in notation. As a result, the accuracy is further reduced.

The present invention has been made in view of the above background, and an object of the present invention is to perform a more accurate automatic classification in consideration of the semantic content of sentences and differences in notation, and to classify the same. The present invention provides an automatic document classification method and apparatus, and a dictionary creation method and apparatus for classification, which can automatically form a dictionary for performing.

[0009]

In order to achieve the above-mentioned object, in the dictionary creating method for automatic document classification according to the present invention, the subject and object are selected from the words and phrases respectively constituting a plurality of documents of known fields. Keyword extraction is performed while classifying such as, and the number of appearing documents of the extracted keywords and the number of appearing documents of a keyword pair in which two arbitrary keywords simultaneously appear are obtained. Then, the distance between the two keywords forming the keyword is calculated from the number of documents that appear the keyword and the number of documents that appear the pair of keywords. Next, the degree of dependency of each keyword pair in the field is calculated from the distance between the keywords, and at least the relationship between the keyword pair and the degree of dependence in the field is stored in the dictionary.

Preferably, when obtaining the number of appearing documents of the keyword pair, a group of keywords having the same meaning is obtained in advance, and if there is a keyword belonging to the same group, it is added to the appearing document number. is there.

Further, as a suitable device for manufacturing the above-mentioned dictionary, an input device for inputting a document known in the field, and words and phrases constituting the document given through the input device are used. Means for extracting a keyword while classifying the subject, object, etc., and a first counting means for receiving the output of the means for extracting the keyword and counting the number of documents in the same field having the same keyword A second counting unit that receives the output of the keyword extracting unit and counts the number of documents in the same field having the same keyword pair, and two keywords that form the keyword pair by receiving the outputs of both counting units. While calculating the distance between them, calculating the degree of dependence of the keyword pair on the field, and storing the determined degree of dependence of the keyword pair in the field in the dictionary. Further, by further comprising means for obtaining a group of keywords having the same meaning in response to the output of the means for extracting the keyword, the second counting means as a means for obtaining the group. It is preferable to receive the output of the keyword extracting unit and add the number of appearing documents of the keyword pair if there are keywords belonging to the same group.

On the other hand, in the method for automatically classifying documents according to the present invention, the keywords extracted while classifying the subject, the object, etc. from the words constituting the input document of the unknown field are extracted, and the extracted keywords are extracted. Determines whether or not matches with a keyword appearing in a predetermined field stored in the dictionary manufactured by the dictionary manufacturing method. Then, the relationship obtained is multiplied by the dependency of each keyword pair of the field stored in the dictionary to calculate the dependency of each keyword extracted from the input document in the field, The degree of ambiguity for each field is calculated from the degree of dependence, and the field having the minimum ambiguity is determined as the field to which the input document belongs.

As an apparatus for carrying out such a method, an input device for inputting a document whose field is unknown, and a subject selected from the phrases constituting the document given via the input device, A means for extracting a keyword while classifying objects such as an object, a dictionary in which data is stored by the dictionary manufacturing apparatus described above, a means for extracting the keyword and the dictionary are connected to the dictionary, and the extracted keyword is Means for obtaining a relationship as to whether or not it matches with a keyword appearing in a predetermined field stored in the dictionary, relationship output from the means, and dependency of each keyword pair of the field stored in the dictionary And a dependency degree calculating means for calculating the degree of dependency of each keyword extracted from the input document with respect to the field by multiplying the relationship with the degree of dependency, An ambiguity calculation unit that receives the output of the dependency calculation unit and obtains ambiguity for each field of the document, and an output of the ambiguity calculation unit that compares the ambiguity of each field to obtain the minimum ambiguity And a determination means for detecting

[0014]

[Function] The familiarity between documents is quantified (dictionary) by using the value of the frequency of simultaneous appearance of a pair of keywords in the document, and the fact that the higher the familiarity is, the more common the document is. A dictionary for each field is created based on the document, and the field of the dictionary that is most familiar to the new document is determined as the field of the document. At this time, when extracting the keywords,
The type (subject, object, etc.) is also determined at the same time,
Since the types are the same and the keywords are the same, the dictionary is created in consideration of the meaning and content of sentences. Therefore, a highly accurate dictionary close to the meaning and content of the document is created, and automatic classification is accurately performed for documents of unknown fields that are processed based on the dictionary thereafter.

Further, when the number of appearances of a keyword pair is calculated, a group of keywords having the same meaning is obtained, and if there is a keyword belonging to the same group, it is added to the number of appearing documents. Even if there is a difference, the correct dictionary can be created without being affected by it, and more accurate automatic classification is performed.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A preferred embodiment of an automatic document classification method and apparatus and a classification dictionary creating method and apparatus according to the present invention will be described in detail below with reference to the accompanying drawings. FIG. 1 shows an embodiment of a dictionary creating apparatus according to the present invention. As shown in the figure, a document of a known field is input via the input device 1 such as a keyboard or OCR. Then, the input document is the next keyword extraction unit 2
Sent to, to extract the nouns present in the document.

In the present invention, first, when extracting a keyword, rather than simply extracting a noun, focusing on the particle or case particle attached immediately after the noun, the type of the noun (subject, object) , Other) are also extracted. As a result, it is possible to understand how the extracted nouns (keywords) are used in the document, and understand the meaning content.

Therefore, the keyword extracting unit 2 in this example
Has a function to be processed by the flow shown in FIG. That is, first, the "divided writing process" for dividing the input document into words is performed (S101). Next, the “segment cutout” process is performed on the above-mentioned segmentation process. The phrase segmentation process is to find a word that is a segment of a segment according to a predetermined rule from among the segments segmented into words by segmentation, and segment the segmented segment again (S102).

Then, after the bunsetsu is cut out in this way, an important bunsetsu is extracted from the cut-out bunsetsu according to the following extraction rule (S103). The end of the phrase is "ga", "ha", "wo", "ni",
Extract the phrases "de,""ya," and "mo." As a result, a clause including the subject / object is extracted. Take out words whose endings are not "Hiragana mixed". For clauses such as verbs, "hiragana" is added at the end.
The focus is on the description of verbs, which eliminates verbs that are difficult to be keywords.

Further, when the important phrase is extracted by the above processing, the important phrase is again divided into words and divided into word units. And in each word, it is difficult to be a keyword, "verb", "particle", "auxiliary verb",
Furthermore, keyword (noun) candidates are extracted by removing "adjectives" such as "great" (S10).
4).

Next, by looking at the words that come next to the keyword candidates extracted by the processing up to S104, the types of keywords (subject, object, etc.) are classified (S10).
5). That is, the word that follows the keyword is "ga",
In the case of “ha”, the keyword is “subject”.
Also, in the case of "wa", "to", "de", "ya", and "mo", it is regarded as "object". Then, all the keywords that do not correspond to the above are “others”.

Further, since the keywords (with types) thus extracted may appear multiple times in the same document, duplicate keywords are deleted and only one is left in the same document. As a result, the keyword (with type) extraction processing ends.

An example of the extraction result (as a sample document, 45 in the same field (opinion about TVCM))
3 documents are shown in FIG. As is apparent from the figure, in the conventional method, the same keyword and the processed one are clearly separated. That is, for example, “○ Den (Other)” in Document 10 and “○ Den (Object)” in Document 22
Are treated as the same as in the past, but are treated as different in the present invention, and the same "○" as in document 2 is used.
Even in the case of "den", when they are used in different ways such as "subject", "object", and "other", they have been treated as one word in the past, but in the present invention, they are treated as three words. It will be treated separately. That is, the same keyword up to the type, for example, “○ den (other)” in documents 2, 3 and 10 and “○ den (object)” in documents 22 and 41 are treated as the same keyword.

In addition, in FIG. 3, the keywords are expressed as a specific character string, but in order to achieve high-speed processing, an ID number is assigned, and it is determined whether or not the following matches based on the ID number. Etc., and accordingly, the ID number management file is also received separately (the same applies hereinafter).

The thus-extracted keywords (with types) are stored in the next-stage keyword file 3 for each document. Further, in the present invention, in order to extract keywords that are different in terms of notation such as “○ den” and “○ × electricity” but should be treated as having the same meaning, the keywords stored in the keyword file 3 are read and It is given to the character string matching keyword extraction unit 4 as a means for obtaining a group, and here, grouping of keywords (with types) having the same meaning is obtained. That is, a table of the base keywords (comparative keywords) and the keywords (element keywords) belonging to the same group as the base keywords is created.

Specifically, in this example, paying attention to the appearance order of the character string of each keyword, the keywords including the same characters in the appearance order of the characters forming the base keyword are regarded as the same keyword (element keyword). I try to regard it. It should be noted that if the appearance order is the same, other characters may intervene between the characters.

That is, as shown in FIG.
If "Den" is the base (comparison keyword), the character string "○,
The keywords in the order of "Den" are the perfect match "○ Den"
Of course, “○ × power” also belongs to the same group (element keyword) and is counted. However, as shown in FIG. 2B, if the base is “◯ × electric power”, the abbreviation “◯ DEN” is not counted because “×” and “force” are not included. If this method is adopted, as shown in FIG. 7C, there is no relation between "PR" and "P".
"OINTER" is also counted, but even if such items are grouped as the same group, it is unlikely that such a keyword will appear as a keyword pair, and the influence on the later-described feature amount calculation processing and the like. There is not much problem because it is less frequent. On the other hand, the same figure (D)
As described above, when the group is “CI”, “CI” and “CI mark” are counted as the same group for the above-mentioned reason, but “IC” is not counted because their appearance order is different (reverse).

Then, it is actually obtained by comparing the permutations of the characters forming the read keyword with each other. Further, when extracting the character string matching keyword, all the documents that constitute the same field are collectively processed. Then, the character string matching keywords (comparison keyword: element keyword table) thus extracted are stored in the character string matching keyword file 5.

The "C" of the document 10 shown in FIG.
Based on “I object”, the “CI mark subject” of document 1 is the same as shown in FIG. 4D, but since the types such as “object” and “subject” are different from each other, different keywords are used. And is not stored in the character string matching file 5. If the total number of documents is eight as shown in FIG. 3, the character string matching keyword file will be stored in the state as shown in FIG.

Further, the output of the keyword file 3 is connected to the keyword counting section 6 as the first counting means and the keyword pair counting section 7 as the second counting means,
The keyword pair counting unit 7 is also connected to the output of the character string matching keyword file 5. Then, the former keyword counting unit 6 is adapted to obtain the number of documents in which each keyword appears (exists).

On the other hand, the keyword pair counting section 7 is adapted to obtain the number of documents in which the same keyword pair appears at the same time. Here, the keyword pair is a combination of two different pairs of keywords existing in one document, and the former combination is the reference keyword and the latter is the comparison keyword. Then, when counting, the reference word is an exact match, and the comparison keyword is a word (element keyword) that is the same as the obtained character string match keyword.

That is, a combination table of keywords existing in the same document is created, the data stored in both files 3 and 5 is accessed based on the created combination table of keywords, and comparison keywords Xj are set for each reference keyword Xi. If there is a document in which there is a combination in which the element keywords match, the document is counted up for that combination of keywords (keyword pair). As a result, the number of appearing documents of the keyword pair is obtained.

As an example, assuming that the total number of documents is eight as shown in FIG. 3, for example, "○ den (object)" (reference keyword) and "water heater (other)" (comparative keyword) In the case of a keyword pair consisting of, if the perfect match is found, the number of occurrence documents is once (only document 22), but in the present example in which the element keyword is also taken into consideration, the number of occurrence documents is twice (documents 22, 42). ). Also, similarly,
The number of appearance documents of the keyword pair consisting of "commercial (object) and Oden (other)" is two (documents 3 and 10). The commercial of document 4 is not counted because its type is “subject”.

Further, the number of appearing documents of each keyword and keyword pair obtained by each of the counting units 6 and 7 is sent to the feature amount calculating unit 8 to obtain each feature amount for creating a dictionary. That is, first, the distance L (xi, xj) between the reference keyword (xi) forming the keyword pair and the comparison keyword (xj) is obtained, and further the distance L (xi, xj) between the keywords is used. A fuzzy set f (x) of keywords is obtained. Then, specifically, it is calculated based on the following formula.

First, the distance L between the keywords is used to quantify the familiarity using the simultaneous appearance frequency value of the keyword pair, and is calculated by the following equation (1).

[0036]

## EQU1 ## L (xi, xj) = (N (xi) + N (xj) -N (xi, xj)) / N (xi, xj) (1) N (xi): Document containing the reference keyword xi Number (output from counting unit 6) N (xj): number of documents including comparison keyword xj (output from counting unit 6) N (xi, xj): number of documents simultaneously including xi and xj (output from counting unit 7) , The fuzzy set (the degree of dependence of the comparison keyword xj on the reference keyword xi) f (x) is obtained by substituting the calculated numerical value into the following equation (2).

[0037]

[Mathematical formula-see original document] f (x) = exp (-a | L (xi, xj) ² |) In the above formula (2), "a" indicates the strength of analogy and the distance L (xi, xj) The dependency of the longest keyword pair is 0.
Takes any value that is less than or equal to 05. Therefore, in this example, the distance L (xi, xj) of each keyword pair is obtained based on the above equation (1) for all documents in the same field, and the maximum value of the distance L (xi, xj) is given by the equation ( Substituting 2) and substituting 0.05 for dependency f (x) and solving for “a” determines “a” for that field.
The obtained dependency is a value of 0 to 1, which means that the closer it is to 1, the more markedly the conceptual features of the field are represented.

The output of the feature quantity extraction unit 8 is connected to the classification dictionary 9 so that the feature quantity obtained as described above is stored in the classification dictionary 9.
In order to perform automatic classification, it is sufficient that at least the degree of dependence on a keyword pair is stored, but as will be described later, the field of a new document is based on the data stored in the classification dictionary 9 concerned. If you have classified
In order to update the contents of the classification dictionary 9 based on the data of the new document, and to update the dictionary more accurately based on the latest multiple pieces of document information, in this example, in addition to the dependency degree, the keyword The number of appearances N (xi), N (xj), the number of appearance documents N (xi, xj) of the keyword pair, and the distance L (xi, xj) are also stored. An example of the stored data structure is shown in FIG. Further, the classification dictionary 9 and the keyword file 3 constitute a dictionary for automatic classification.

Next, one embodiment of the method of the present invention (dictionary creating method) will be described using the above-mentioned embodiment. As shown in FIG. 7, a plurality of sample documents whose fields are known are used, and each sample document (preparing a plurality of sample documents in the same field) together with the fields is input to the dictionary creating device via the input device 1 (S201). ). Next, a keyword (with type) is extracted from each document, and an element keyword representing the same meaning content as a certain keyword is obtained using the character string matching keyword extraction unit 5 (S202, 203).

Then, the number of documents having a keyword and the number of documents having a keyword pair (documents in which a pair of a reference keyword and a comparison keyword appear at the same time) are obtained in each counting section 6 and 7 (S20).
4). In this example, when the keyword pair is obtained, the comparison keyword includes the element keyword, but the comparison keyword may be an exact match like the reference keyword without considering the element keyword.

Then, the calculated keywords and the number of appearing documents of the keyword pairs are given to the feature amount calculating section 8 to calculate the distance between the keywords (between the reference keyword and the comparison keyword), and based on the distance, the keywords are related. Calculate the degree of dependence. Then, the calculated result is stored in the classification dictionary 9, and the process ends (S205 to 207).

Specifically, 45 documents in the same field (opinions about TVCM) are input to the dictionary creating device as sample documents, and a part of the obtained results (dictionaries) is as shown in FIG. Become. A part of the dictionary obtained by using the conventional method for the same sample document (a method that counts the number of documents that appear in a keyword pair based on complete matching of comparison keywords without classifying the keywords) Is shown in FIG. As is clear from FIGS. 6 and 8, the contents of the dictionaries obtained by both formulas are different. By the way, the accuracy of both dictionaries is not clear just by looking at this dictionary,
It is possible to judge pass / fail by inputting 45 sample documents used when the dictionary was created into the automatic classification device described below, and obtaining the degree (ambiguity) of the document to belong to the field, As a result, it was confirmed that the dictionary manufactured by the present invention has a higher degree of belonging (lower ambiguity).

FIG. 9 shows an embodiment of the automatic classification device according to the present invention. As shown in the figure, the dictionary creating apparatus shown in FIG. 1 includes an input device 1 ', a keyword extracting unit 2', a keyword file 3 created by the dictionary creating device, and a classification dictionary 9. Further, the outputs of the keyword extracting unit 2'and the keyword file 3 are input to the fuzzy relation creating unit 10. The fuzzy relation creating unit 10 compares the reference word of the fuzzy set of each field (obtained from the keyword file 3) with the extraction result of the unknown new document of the field (given by the keyword extracting unit 2) and matches them. If the fuzzy relation 1 is "1", and if they do not match, "0"
And Then, the fuzzy relation R thus obtained is stored in the fuzzy relation file 11.

For convenience of description, assume that the data of a certain field stored in the classification dictionary 9 is as shown in FIG. 10 (the number of keywords is 8), and from the new document, the data shown in FIG. If individual keywords are extracted, they are stored in the fuzzy relation file 11 in a state as shown in FIG. 12, for example.

Then, the fuzzy relation file 11 and the classification dictionary 9 are connected to the field-dependent dependency calculation unit 12, and the fuzzy set F stored in the classification dictionary 9 is input into the field-dependent dependency calculation unit 12. A new fuzzy set (dependency) is created by synthesizing the fuzzy relation R created based on the document (field unknown) (Equation (3) below).
As a result, the degree of dependence of each keyword of the document is obtained for each field.

[0046]

[Equation 3] Then, the degree of dependency of the keyword on the field shown in FIG. 10 is calculated according to the calculation formulas shown in FIGS. 13 and 14, and each value as shown in FIG. 14 is obtained. Here, “*” is a logical product (takes the minimum value), and “U” means to take the maximum value among the connected numerical values. And
Since such arithmetic processing is performed for all fields, in the end,
There are as many dependency relationships (tables) as there are fields as shown in FIG.

Then, the obtained field-specific dependency is sent to the field-specific ambiguity calculation unit 13 in the next stage, and the fuzzy set obtained above is substituted into the following equation (4) to obtain each category. Find the ambiguity d. The smaller the ambiguity d is, the less ambiguity is, that is, the document matches the field.

[0048]

D (category) = (Δ1 + Δ2 + ... + Δn) / n Δn = -ua (ai) log ₂ ua (ai)-(1-ua (ai)) log ₂ (1-ua (ai)) where , N indicates the number of keywords for each field, and ua (ai) indicates the degree of dependence of keyword pairs. 15 shows an example of the calculation (calculation of the ambiguity of the new document of FIG. 11 for a certain field of FIG. 10).

Then, the field-specific ambiguity calculation unit 13 thus obtained is sent to the determination / processing unit 14, where the field with the smallest ambiguity is determined as the field of the input document, and the result is determined. Output device (CRT, printer, etc.) 15
Output to. Note that, in general, in order to determine the field of the document to be stored in the database by carrying out the present invention when creating the database, the extracted keyword and the determined field may be stored in the predetermined database 16. The entire text of the document is stored in the database 16 as it is via the input device 1 '.

Further, in this example, the input new document to be processed is sent together with the determined field to the dictionary creating apparatus shown in FIG. Then, the dictionary can be updated based on the new document. Note that the count of the number of appearing documents of keywords and keyword pairs in each of the counting units 6 and 7 shown in FIG.
Actually, the process of incrementing the numerical value of the corresponding part (keyword, keyword pair existing in a new document) in each number of times stored in the classification dictionary 9 is performed.

When the classification dictionary is updated as described above, the keywords extracted by the keyword extraction unit 2'are already extracted from a known document and stored in the keyword file 3. If some of the stored keywords are not included, a keyword related to the subsequent automatic classification and data (element keyword, etc.) based on the keyword are created and stored in the predetermined files 3, 5
And so on.

Next, an example of the automatic classifying method according to the present invention using the above apparatus will be described. First, in this example, the classification dictionary 9 stores fuzzy sets for two fields. These two fields are the above-mentioned “TVC
Opinion about M (first field) "and" C using paper as medium "
Opinion about M (second field) ", and for the first field, create a dictionary based on 45 sample documents,
For the second field, a dictionary was created based on 69 sample documents.

In this state, an unknown document is input to the automatic classification device via the input device 1 ', and first the keyword (with type) is extracted by the keyword extraction unit 2'. At this time, only one duplicate keyword is left. And the result is shown in FIG.

Next, after the fuzzy relation creating unit 10 obtains the fuzzy relation, the keyword (xj) that constitutes the newly obtained fuzzy set that matches the keyword and that exists in the new document input for each field and its fuzzy Extract the set F. As a result, a fuzzy set for the first field (FIG. 17) and a fuzzy set for the second field (FIG. 18) are obtained. Based on this, the field-dependent dependency calculation unit 12 obtains a fuzzy set (dependency).

Further, the field-specific ambiguity calculator 13 calculates the ambiguity. The result is shown in FIG. 19 (note that the keywords for which the degree of dependence is “0” are omitted). Then, the determination / processing unit 14 compares the ambiguities of the respective fields and selects the field having the smallest value. Then, in this example, the first field is determined.

Further, as a result of inputting another document (having the keywords shown in FIG. 20) and performing the same processing as described above, the ambiguity as shown in FIG. 21 is obtained. It can be determined that the document belongs to two fields. FIG. 22 shows the processing flow described above (one embodiment of the automatic classification method according to the present invention).

Further, as a result of automatically classifying the above-mentioned document and the document based on the conventional method and obtaining the ambiguity with respect to the first field, the result is as shown in FIG. That is, as is clear from this figure, in the case of this sample document, the classification was correctly performed even in the conventional method (document → first field, document → second field). However, focusing on the ambiguity, the ambiguity of the document is “0.106 ...”, which is a larger value than the ambiguity “0.0841 ...” obtained by the present invention,
On the contrary, the ambiguity of the document is “0.40496 ...”, which is a numerical value smaller than the ambiguity of “0.5392 ...” obtained by the present invention. From this, according to the present invention, the ambiguity becomes smaller for the correct field, while the ambiguity becomes larger for the incorrect field, and the similar fields are accurately separated. It turns out that they can be classified.

Further, the ambiguity of the first field obtained as a result of inputting and processing all the keywords constituting the 45 sample documents of the first field used when creating the dictionary into the classifying device is: It became 0.08077751.
Similarly, in the conventional system, 0.08242087
Became. That is, it can be seen that the present invention has a lower ambiguity and the dictionary is created accurately.

[0059]

As described above, in the method and apparatus for automatically classifying documents according to the present invention and the method and apparatus for creating a dictionary for classification, when extracting a keyword, the type (subject, object, etc.) of the keyword is also calculated. Since it is determined that the type is the same and the keywords are the same, the dictionary is created in consideration of the semantic content of the sentence, so a highly accurate dictionary close to the content of the document can be created, and the document of unknown field The automatic classification can be performed accurately.

Further, when the number of appearances of the keyword pair is calculated, the character string coincidence is regarded as the same keyword. Therefore, even if there is a difference in notation,
A correct dictionary can be created without being affected by it, and automatic classification is possible.

[Brief description of drawings]

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

FIG. 2 is a flowchart showing the function of a keyword extracting unit.

FIG. 3 is a diagram showing an example of keywords extracted by a keyword extracting unit.

FIG. 4 is a diagram illustrating an operation of a character string matching keyword extraction unit.

FIG. 5 is a diagram showing an example of a keyword group constituting the same group extracted by a character string matching keyword extraction unit.

FIG. 6 is a diagram showing an example of a data structure of a classification dictionary.

FIG. 7 is a flowchart showing an embodiment of a dictionary creating method according to the present invention.

FIG. 8 is a diagram showing a part of a classification dictionary created by a conventional method.

FIG. 9 is a block diagram showing an embodiment of an automatic classification device according to the present invention.

FIG. 10 is a diagram showing an example of a dictionary stored in a classification dictionary.

FIG. 11 is a diagram showing an example of keywords extracted by a keyword extracting unit.

FIG. 12 is a diagram showing a relationship created by a fuzzy relationship creating unit.

FIG. 13 is a diagram for explaining an operation of a field-dependent dependency calculating unit.

FIG. 14 is a diagram illustrating an operation of a field-dependent dependency calculation unit.

FIG. 15 is a diagram illustrating an operation of a field-specific ambiguity calculation unit.

FIG. 16 is a diagram showing a process of an automatic classification method performed by using the automatic classification device.

FIG. 17 is a diagram showing a processing process of an automatic classification method performed using the automatic classification device.

FIG. 18 is a diagram showing a process of an automatic classification method performed using the automatic classification device.

FIG. 19 is a diagram showing a process of an automatic classification method performed using the automatic classification device.

FIG. 20 is a diagram showing a process of an automatic classification method performed by using the automatic classification device.

FIG. 21 is a diagram showing a process of an automatic classification method performed by using the automatic classification device.

FIG. 22 is a flow chart diagram showing an example of an automatic classification method according to the present invention, which is carried out using an automatic classification device.

FIG. 23 is a diagram showing a result of classification processing by a conventional method.

[Explanation of symbols]

 1, 1'Input device 2, 2'Keyword extracting device 3 Keyword file 4 Character string matching keyword extracting unit 5 Character string matching keyword file 6 Keyword counting unit 7 Keyword pair counting unit 8 Feature amount calculating unit 9 Classification dictionary 10 Fuzzy relations Creation unit 11 Fuzzy relation file 12 Dependency calculation unit by field 13 Ambiguity calculation unit by field 14 Judgment / processing unit 15 Output device 16 Database

Claims (6)

[Claims] 1. A keyword is extracted while classifying a subject, an object, etc. from words and phrases constituting a plurality of documents known in the field, and the number of documents in which the extracted keyword appears and the arbitrary two keywords are The number of appearing documents of a keyword pair that appears at the same time is calculated, the distance between two keywords forming the keyword is calculated from the number of appearing documents of the keyword and the number of appearing documents of the keyword pair, and each keyword is calculated from the distance between the keywords. A method for creating a dictionary for automatic classification of documents, in which the dependency of a pair on the field is calculated and at least the relationship between the keyword pair and the dependency in the field is stored in the dictionary. 2. The method according to claim 1, wherein, when obtaining the number of appearing documents of the keyword pair, a group of keywords having the same meaning is obtained in advance, and if there are keywords belonging to the same group, the number of appearing documents is added. Dictionary creation method for automatic classification of other documents. 3. An input device for inputting a document of a known field, and a keyword given to a document given via the input device while classifying the subject, object, etc. from the words and phrases constituting the document. An output means for extracting the keyword, a first counting means for receiving the output of the means for extracting the keyword and counting the number of documents in the same field having the same keyword, and an output for the means for extracting the keyword, Second counting the number of documents in the same field with keyword pairs
The counting means and the outputs of the both counting means are used to calculate the distance between two keywords forming the keyword pair, calculate the dependency of the keyword pair on the field, and calculate the keyword pair in the field. A dictionary creation device for automatic classification of documents, which is provided with a means for storing the degree of dependency in a dictionary. 4. The apparatus further comprises means for receiving an output of the means for extracting the keyword to obtain a group of keywords having the same meaning, wherein the second counting means comprises means for obtaining the group and means for extracting the keyword. The dictionary creating apparatus for automatic classification of documents according to claim 3, which receives the output and adds the keywords belonging to the same group to the number of appearing documents of the keyword pair. 5. The method according to claim 1, wherein a keyword is extracted while classifying a subject, an object, and the like from words and phrases that compose a document whose field is unknown, and the extracted keyword is the method. Then, a relationship is determined whether or not the keyword appears in a predetermined field stored in the dictionary manufactured by, and then the relationship obtained and the degree of dependence of each keyword pair of the field stored in the dictionary. And the degree of dependency of each keyword extracted from the input document in that field is calculated, the ambiguity for each field is obtained from the dependency, and the field having the minimum ambiguity is input as described above. A method of automatically classifying documents so that the field to which the document belongs is determined. 6. An input device for inputting a document of unknown field, and a keyword given to a document given through the input device while classifying the subject, object, etc. from the words and phrases constituting the document. Means for extracting, a dictionary data-stored by the device according to claim 3 or 4, means for extracting the keyword and connected to the dictionary,
A means for obtaining a relationship as to whether the extracted keyword matches a keyword appearing in a predetermined field stored in the dictionary, a relationship output from the obtaining means, and the field stored in the dictionary And a dependency degree calculating means for receiving the dependency degree of each keyword pair and calculating the dependency degree of each keyword extracted from the input document for the field by multiplying the dependency degree and the dependency degree. An ambiguity calculation unit that receives the output of the means and obtains the ambiguity for each field of the document, and an output of the ambiguity calculation unit that compares the ambiguity of each field, and detects the field having the minimum ambiguity. The automatic document classification device according to claim 5, further comprising a determination unit.