JP2001312501A - Automatic document classification system, automatic document classification method, and computer-readable recording medium recording automatic document classification program - Google Patents

Abstract

(57) [Summary] [Problem] There is a problem that a classification category of a document to be classified must be created in advance, and a large number of learning documents set for each classification category must be prepared. SOLUTION: Based on a classification reference sentence provided corresponding to a classification field, a document structure analysis means for analyzing a sentence structure of a classification target document to be classified into the classification field, and a result of analysis by the document structure analysis means. A similarity calculating means for clustering the classification reference sentence according to a classification rule and calculating a similarity relating to the point of interest between the analysis result of the document to be classified and the clustered classification reference sentence according to a classification rule defining a point of interest of classification; Document classification means for determining a classification field of the classification target document based on the similarity calculated by the degree calculation means.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention analyzes the contents of a freely described answer among document data in which various questions are described in a specific field, such as a questionnaire, with options and free descriptions. The present invention relates to an automatic document classification system, an automatic document classification method, and a computer-readable recording medium on which an automatic document classification program is recorded.

[0002]

2. Description of the Related Art As a conventional automatic document classification system, an automatic document classification method based on the learning result of a semantic attribute, which automatically classifies using the appearance frequency of words appearing in a document.
(Information Processing Society of Japan Vol.33 No.9, pp.
1114-1122).

FIG. 10 is a block diagram showing the configuration of the above-described conventional document automatic classification system. In the figure, 1 is a dictionary referred to in the analysis of documents, 3 is a classification target document storage unit for storing documents to be classified, and 101 is a classification field to which each classification category (for example, document contents such as economy, sports, culture, etc. belongs) A) a learning document storage unit for storing a learning document, which is sample data that has been classified, and 102 performs morphological analysis on the classification target document and the learning document to extract nouns, and the classification target document noun storage unit 103 or Noun extraction processing means 103 stored in learning document noun storage unit 104
Is a classification target document noun storage unit that stores the noun extracted from the classification target document by the noun extraction processing means 102;
A learning document noun storage unit 4 stores the noun extracted from the learning document by the noun extraction processing unit 102.
Reference numeral 5 denotes a classification vector creating unit that counts the appearance frequency of the semantic attribute assigned to the noun stored in the learning document noun storage unit 104 for each classification category already assigned to each learning document. Further, the result of calculating the weight of the semantic attribute of each noun for each classification category is stored in the classification vector storage unit 106 as a classification vector.

[0006] Reference numeral 106 denotes a classification vector storage unit for storing classification vectors, and 107 denotes a classification target document noun storage unit 10.
A document classifying unit that counts the appearance frequencies of the semantic attributes assigned to the nouns stored in No. 3 and converts them into a vector in the same manner;
Is a vector similarity calculation unit that calculates the similarity between the classification vector stored in the classification vector storage unit 106 and the vector created by the document classification unit 107, and outputs the result to the document classification unit 107. The classification result storage unit stores the classification category having the highest similarity as the classification destination of the classification target document.

Next, the operation will be described. First, dictionary 1
In advance, a classification item to be classified (hereinafter referred to as a classification category) is created in advance, and a learning document as sample data classified for each classification category is stored in the learning document storage unit 101. Next, the noun extraction processing means 102 extracts a noun from each of the learning documents in the learning document storage unit 101 using morphological analysis or the like.
04.

[0006] Thereafter, the classification vector creating means 105
Calculates the appearance frequency of the semantic attribute assigned to the noun in the learning document noun storage unit 104 for each classification category.
The result of calculating the weight of the semantic attribute of each noun for each classification category is stored in the classification vector storage unit 106. In the calculation of the weight, the conventional document automatic classification system uses the chi 2
The calculation is performed by a method using the power statistics. The classification vector is a vector representation of the weight of the semantic attribute of the noun for each classification category.

Next, automatic classification will be described. First, the classification target documents to be automatically classified are taken out one by one from the classification target document storage unit 3, and the noun extraction processing means 102 extracts the noun from the classification target document like the learning document, and the classification target document noun storage unit 103 To be stored. Next, the document classifying unit 107 counts the appearance frequencies of the semantic attributes assigned to the nouns stored in the classification target document noun storage unit 103, and expresses them as a vector like the learning document. Further, the vector similarity calculating means 108 calculates the similarity between the above-mentioned vector and the classifying vector in the classifying vector storage unit 106, and determines the classification category having the highest similarity (that is, the classification category having the highest similarity). The classification category set in the corresponding learning document) is stored in the classification result storage unit 109 as the classification destination of the classification target document.

[0008]

Since the conventional automatic document classification system is configured as described above, a classification category of a document to be classified is created in advance, and a large number of learning documents set for each classification category are prepared. There was a problem that it had to be prepared.

More specifically, the above-mentioned problem will be described. In order to set a learning document in which a classification category is set in advance and compare the document with a document to be classified and perform automatic classification, as described in a questionnaire result, It is difficult to set a classification category in advance when it is not known whether the contents are described.

[0010] In addition, it is necessary to prepare a large number of learning documents set for each classification category. Preparing a large number of training documents is cost-effective, but preparing a large number of learning documents is cost-effective when automatically classifying documents for a one-time analysis like a questionnaire. Do not fit.

Furthermore, the conventional automatic document classification system has a problem that the method of calculating the similarity is a single method, so that it is not possible to perform classification using a point of interest suitable for the purpose of the user.

[0012] To explain the above problem in detail, the open-ended answer of the questionnaire is often described in terms of the nature of the document, from one sentence to at most several sentences, and the questionnaire questionnaire has a limited number of classification categories. Therefore, there is a high possibility that the appearance tendency of words freely described in the questionnaire will be similar. In addition, the free description content of the questionnaire may include the special intention or judgment of the respondent. When automatically classifying such a questionnaire, for example, paying attention to "what objects are answered by many people" or "whether users think they are good or bad" We must focus on and classify. In other words, as you categorize your questionnaire,
The focus of classification as a linguistic expression is different. In a classification method that simply treats the presence or absence of words as in a conventional automatic document classification system, it is not possible to classify by distinguishing the noted points as described above.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and a classification for defining a point to be focused on at the time of classification for a document set such as a questionnaire in which a target classification category is limited to some extent. Rules are given in advance, and when a document serving as a classification reference sentence set for each classification category is given as sample data, clustering is automatically performed, so that automatic classification that requires a large amount of learning documents can be performed. An object of the present invention is to obtain an automatic document classification system, an automatic document classification method, and a computer-readable recording medium on which an automatic document classification program is recorded, which can automatically perform classification with a small amount of sample data.

[0014]

An automatic document classification system according to the present invention provides a classification reference sentence corresponding to a classification field and a document for analyzing the sentence structure of a classification target document to be classified into the classification field. Clustering the classification standard sentence based on the analysis result by the structure analysis means and the document structure analysis means, and paying attention to the analysis result of the document to be classified and the clustered classification standard sentence according to the classification rule which specifies the point of interest of the classification And a document classifying unit that determines the classification field of the document to be classified based on the similarity calculated by the similarity calculating unit.

An automatic document classification system according to the present invention comprises:
The classification rule defines the point of interest in the classification based on the feature amount of a natural language document.

An automatic document classification system according to the present invention comprises:
A clause similarity indicating the similarity of each clause constituting the classification reference sentence and the document to be classified, and a dependency similarity indicating the similarity of dependency between each clause constituting the classification reference sentence and the document to be classified. At least one of the degree and the modal similarity indicating the similarity of the modal expression between the classification reference sentence and the classification target document is used as a feature amount of a natural language document.

An automatic document classification system according to the present invention comprises:
The classification rule defines a plurality of points of interest at the time of classification, sets a threshold according to the purpose of classification for the similarity of each point of interest, and the document classification means classifies the document to be classified based on the similarity equal to or more than the threshold. Determine the field.

An automatic document classification system according to the present invention comprises:
The classification target document is composed of a plurality of sentences, and the similarity calculation means calculates the similarity of the classification target document in units of sentences.

An automatic document classification system according to the present invention comprises:
The similarity calculating means sets the highest similarity among the similarities between the sentence constituting the classification target document and the classification reference sentence as the similarity between the classification target document and the classification reference sentence.

An automatic document classification system according to the present invention comprises:
When the classification rule uses the phrase similarity indicating the similarity of each of the phrases constituting the classification reference sentence and the document to be classified as the feature amount of the document of the natural language, only the phrases consisting of the nominative words,
Alternatively, the similarity calculating means calculates the phrase similarity only from a phrase consisting of words of the declinable word.

An automatic document classification system according to the present invention comprises:
The document classifying means classifies the documents to be classified hierarchically according to the points of interest in the classification, and the classification rules are preset for each layer according to the points of interest in the classification.

The automatic document classification system according to the present invention comprises:
When the degree of similarity between the classification target document and the plurality of classification reference sentences is equal to or greater than a predetermined threshold, the document classification means classifies the classification target document into a classification field corresponding to each of the plurality of classification reference sentences.

An automatic document classification system according to the present invention comprises:
A plurality of classification criteria sentences are set for one classification field,
The similarity calculating means calculates the similarity between each classification standard sentence and the document to be classified, and the document classifying means classifies the class having the highest similarity among the similarities calculated by the similarity calculating means. This classifies the document to be classified into a classification field corresponding to the reference sentence.

According to the automatic document classification method of the present invention, there is provided a document structure analysis step of analyzing a classification reference sentence provided corresponding to a classification field and a sentence structure of a classification target document to be classified into the classification field. Based on the analysis result in the document structure analysis step, the classification reference sentence is clustered, and the similarity between the analysis result of the document to be classified and the clustered classification reference sentence is determined according to the classification rule that specifies the point of interest for classification. The method includes a similarity calculation step to be calculated, and a document classification step of determining a classification field of a document to be classified based on the similarity calculated in the similarity calculation step.

In the automatic document classification method according to the present invention, the classification rule specifies a point of interest in classification based on the feature amount of a document in a natural language.

In the automatic document classification method according to the present invention, the classification rule indicates the similarity of each of the clauses constituting the classification reference sentence and the classification target document, and the clause similarity indicating the similarity between the classification reference sentence and the classification target document. At least one of the dependency similarity indicating the similarity of the dependency between the phrases and the modal similarity indicating the similarity of the modal expression between the classification reference sentence and the document to be classified is used as the feature amount of the document in the natural language. To use.

In the automatic document classification method according to the present invention, the classification rule defines a plurality of points of interest at the time of classification, sets a similarity degree for each point of interest to a threshold value according to the purpose of classification, and performs a document classification step. That is, the classification field of the document to be classified is determined based on the similarity equal to or larger than the threshold value.

In the automatic document classification method according to the present invention, in the similarity calculation step, a similarity is calculated for each classification target document composed of a plurality of sentences.

In the automatic document classification method according to the present invention, in the similarity calculation step, the similarity indicating the highest value among the similarities between the sentence constituting the classification target document and the classification reference sentence is determined by the classification target document and the classification target document. This is a similarity to the reference sentence.

In the automatic document classification method according to the present invention, in the similarity calculation step, the classification rule indicates the similarity of each of the clauses constituting the classification reference sentence and the document to be classified by the clause similarity of the natural language document. When used as a feature amount, the phrase similarity is calculated from only phrases consisting of nominative words or only phrases consisting of verbal words.

In the automatic document classification method according to the present invention, in the document classification step, the documents to be classified are classified hierarchically according to the points of interest of the classification, and the classification rules are set in advance for each layer corresponding to the points of interest of the classification. It is set.

In the automatic document classification method according to the present invention, in the document classification step, when the degree of similarity between the classification target document and the plurality of classification reference sentences is equal to or greater than a predetermined threshold value, the document classification step corresponds to each of the plurality of classification reference sentences. This classifies the document to be classified into the classification field to be classified.

In the automatic document classification method according to the present invention, a plurality of classification reference sentences are set for one classification field, and a similarity calculation step calculates a similarity between each classification reference sentence and a document to be classified. Then, in the document classification step, the classification target document is classified into the classification field corresponding to the classification reference sentence having the similarity indicating the highest similarity value among the similarities calculated in the similarity calculation step. .

A computer-readable recording medium on which the automatic document classification program according to the present invention is recorded includes a classification reference sentence provided corresponding to a classification field and a sentence structure of a document to be classified to be classified into the classification field. The document structure analysis function to be analyzed and the classification standard sentence are clustered based on the analysis result by the document structure analysis function, and the analysis result of the document to be classified and the clustered classification standard sentence are And a document classification function for determining the classification field of the document to be classified based on the calculated similarity.

[0035]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a block diagram showing a configuration of an automatic document classification system according to Embodiment 1 of the present invention. In the figure, reference numeral 1 denotes a dictionary referred to when analyzing a sentence structure of a classification reference sentence or a document to be classified, 2 denotes a classification reference sentence storage unit that stores a classification reference sentence associated with a classification category (classification field), A classification target document storage unit 3 stores the classification target document. Reference numeral 4 denotes a document analysis unit (document structure analysis unit) that analyzes the classification reference sentence of the classification reference sentence storage unit 2 and the sentence structure of the classification target document in the classification target document storage unit 3 using the dictionary 1.

Reference numeral 5 denotes a classification reference sentence analysis result storage unit for storing the result of analyzing the sentence structure of the classification reference sentence in the classification reference sentence storage unit 2 by the document analysis means 4. The classification target document analysis result storage unit stores the result of analyzing the sentence structure of the classified target document by the document analysis unit 4. Reference numeral 7 calculates the similarity between the classification reference sentence stored in the classification reference sentence analysis result storage 5 and the classification target document stored in the classification target document analysis result storage 6 which has been subjected to the sentence structure analysis. This is a classification rule storage unit that stores the classification rules used when performing.

Reference numeral 8 denotes a classification standard sentence having a sentence structure analyzed stored in the classification standard sentence analysis result storage unit 5 and a classification target document analysis result storage unit 6 stored in accordance with the classification rule stored in the classification rule storage unit 7. Document similarity calculating means (similarity calculating means) for calculating the similarity to the classified sentence structure analyzed document. Reference numeral 9 denotes a document classifying unit that determines the classifying destination of the document to be classified according to the similarity calculated by the document similarity calculating unit 8. Reference numeral 10 denotes a classification result storage unit that stores the classifying destination determined by the document classifying unit 9.

Next, the operation will be described. FIG. 2 is a flowchart showing the operation of the automatic document classification system according to the first embodiment of the present invention. The operation will be described with reference to FIG. First, the document analysis unit 4 performs a morphological analysis process or the like on the classification target document stored in the classification target document storage unit 3 to perform a sentence structure analysis. Specifically, for example, “morphological analysis of Japanese sentence including unregistered words” (Information Processing Society Transactions Vol.
30 No. 3, pp. 294-301 (1989))
Performs the morphological analysis processing disclosed in (1). This will be briefly described below.

First, the document analysis means 4 analyzes each morpheme of the document to be classified using the dictionary 1. Next, the presence / absence of an unknown word is checked with respect to the analysis result of the morpheme. If there is an unknown word, a range constituting the unknown word is estimated, and a plurality of morphemes are put together. Then, the phrase structure is analyzed from the morpheme analysis result. Here, attribute information, independent word information, and adjunct word information of each clause constituting the classification target document are extracted. The attribute information is composed of information such as the dependency of the word constituting the phrase as a grammatical property, the independent word information is information on the independent word constituting the phrase, and the auxiliary word information is the independent word constituting the phrase. This is information on adjuncts such as particles and adverbs attached to. When the phrase structure is analyzed, dependency analysis is performed on morphemes constituting the phrase based on the extracted attribute information, independent word information, and attached word information. As a result, the dependency structure of each clause constituting the classification target document is determined.

When the document analyzing means 4 analyzes the sentence structure of the document to be classified in this way, it stores the result in the document analysis result storage unit 6 for classification. In the same manner as described above, the document analysis unit 4 performs a sentence structure analysis on the classification standard sentence stored in the classification standard sentence storage unit 2 and stores the analyzed result in the classification standard sentence analysis result storage unit 5. . The processing so far corresponds to step ST2-1 (sentence structure analysis step).

FIG. 3 is a diagram showing a freely described questionnaire as an example of a document to be classified by the automatic document classification system according to the first embodiment. In this example, the contents are freely described in response to a question such as "Please describe below if you have any information about your personal computer". In the figure, the questionnaire number is a serial number given to each freely described questionnaire. For example, the answer to the questionnaire number 1 corresponds to "memory cannot be added" in the figure. As shown in the figure, the answer content is 1
The sentence is described in a few sentences at most, and the classification category is limited to some extent from the questionnaire, and the appearance tendency of words freely described in the questionnaire is similar.

FIG. 4 is an example showing the result of analyzing the sentence structure of the answer to question number 1 in FIG. As shown in the figure, "memory" and "expansion" are extracted as independent words,
As a result of the analysis, a sentence structure in which “memory” is dependent on “expansion” having an attribute of “possible or negative”, that is, “not possible” is obtained.

Next, the document classifying means 9 sequentially reads the sentence-structure-analyzed classification reference sentence and the classification target document from the classification reference sentence analysis result storage unit 5 and the classification target document analysis result storage unit 6 for each single sentence. The document similarity calculating means 8 calculates the similarity between the classification reference sentence and the classification target document according to the classification rule stored in the classification rule storage unit 7 for the read classification standard sentence and the simple sentence constituting the classification target document. Calculate (Step S
T2-2, similarity calculation step).

The document classifying means 9 assigns a document to be classified to a classification category having the most similar classification standard sentence based on the similarity between the classification standard sentence calculated by the document similarity calculating means 8 and the document to be classified. The classification result is stored in the classification result storage unit 10 (step ST2-3, document classification step).

Next, the similarity calculation processing and the document classification processing will be described in detail. FIG. 5 is a flowchart showing a similarity calculation process and a document classification process in the automatic document classification system according to the first embodiment.
This corresponds to the operation of T2-2 and step ST2-3. First, before proceeding to step ST3-1, the document analysis unit 4 analyzes the sentence structure of the classification target document stored in the classification target document storage unit 3 in advance. Next, a classification category, a classification standard sentence corresponding thereto, and a classification rule applied at the time of classification are input (step ST3-1). Thereafter, the document analysis means 4 analyzes the sentence structure of the classification reference sentence in the same manner as described above, and stores the analysis result in the classification target document analysis result storage unit 6 (step ST3-2, document structure analysis step).

When the sentence structure of the classification reference sentence is analyzed in step ST3-2, the document classifying means 9 determines whether or not the analysis result of the classification target document is stored in the classification target document analysis result storage unit 6. If there is a document to be classified that has been analyzed, the document to be classified is extracted for each single sentence, output to the document similarity calculation means 8, and the process proceeds to step ST3-4. Is performed, and if there is no analyzed document to be classified, the process proceeds to step ST3-6, where the classification result is stored in the classification result storage unit 10 (step ST3-3, similarity calculation step).

In step ST3-4, the similarity comparison between the classification reference sentence corresponding to each classification category and the document to be classified is calculated according to the classification rule to be applied. The similarity calculation process between the classification reference sentence and the document to be classified will be described in detail after a specific example of a classification rule is described.

FIG. 6 is a diagram showing an example of a classification rule used by the automatic document classification system according to the first embodiment. In the figure, reference numeral 31 indicating a region surrounded by a broken line indicates a classification rule in rule number 1. The classification rule 31 is a phrase similarity (see FIG. 6) indicating the similarity of each of the phrases forming the classification reference sentence and the document to be classified, as a feature amount of a natural language document that defines a point of interest at the time of classification. Dependency similarity (corresponding to “Dependency” in FIG. 6) indicating the similarity of dependency information between the clauses constituting the classification reference sentence and the document to be classified. And a modal similarity indicating the similarity of the modal expression between the classification reference sentence and the classification target document (corresponding to “modal” in FIG. 6 and a modal expression attached to a word such as “negation” is referred to as “modal”) You are using Here, in the illustrated example, the parameter of the phrase similarity is 0.8, the parameter of the dependency similarity is 0.2, and the modal similarity related to the modal expression is 1.0 and the similarity. These parameters are set according to weighting (threshold) for each similarity, that is, which similarity the user classifies with importance. The similarity with the parameter set to 1.0 is not used for calculating the similarity between the classification reference sentence and the classification target document.

When the phrase similarity parameter is increased, sentences having the same word can be collected. That is, when classifying a plurality of classification target documents (questionnaire numbers 1 to 2000) as shown in FIG. 3, the classification target documents having the same word are collected from the phrase similarity between the classification reference sentence and each classification target document. be able to. Thus, for example, when attention is paid to "what is answered in the questionnaire" as the purpose of classification, this can be easily known.

Furthermore, if the parameter of the dependency similarity is increased, documents to be classified having similar expressions can be collected. As a result, it is possible to perform classification by paying attention to "whether the same thing is being done". Furthermore, if the parameter of modal similarity related to modal similarity is reduced, documents to be classified can be classified based on differences in detailed expressions with respect to adjectives. Can be classified. As described above, by changing the parameter of the degree of similarity, it is possible to change a point of interest at the time of classification, and it is possible to perform classification that matches the purpose of the user.

Here, the process of calculating the similarity between the classification reference sentence and the document to be classified in step ST3-4 will be described in detail. As a method of calculating the similarity between the classification reference sentence and the classification target document, for example, a “similar document search system and method, and a computer-readable recording medium recording a similar document search program” (Japanese Patent Application No. 11-2)
No. 57167). In the similar document search system, the similarity is calculated by the following equation. Similarity = α × Phrase similarity + β × Dependency similarity Equation (A) α: Parameter of phrase similarity, β: Parameter of dependency similarity Here, the phrase similarity is common to two sentences. The similarity is calculated based on the number of independent words, and the dependency similarity is calculated based on the number of common dependency structures.

Further, in the calculation of the phrase similarity, the similarity of each word is weighted according to whether or not the modal is different from the expression attached to the word such as “negation” (specifically, the similarity of the modal is the same). In this case, the similarity between words is set to 1.0, and when they do not match, it is set to 0.1). The weight related to the modal is treated independently as another parameter, and the modification is calculated so that the similarity is calculated by the following equation. Similarity = α × Phrase similarity (γ) + β × Dependency similarity Equation (B) γ: Parameter at the time of modal mismatch (when the modal does not match, the similarity of the phrase is γ (0 ≦ γ ≦ 1),
When the modal matches, the similarity of the phrase is set to 1.0). That is, assuming that the value of γ is 0.1, Japanese Patent Application No. Hei 11-2
The expression (A) used in the similar document search system of Japanese Patent No. 57167 for similarity calculation is the same. In the following, the formula (B)
And the document similarity calculating means 8 calculates the similarity.

In the similar document search system disclosed in Japanese Patent Application No. 11-257167, the target of similarity calculation is calculated between an input search sentence and a cluster representative structure. The classification reference sentence is calculated as an input search sentence, and the classification target document is calculated as a cluster representative structure.

Next, an example of application of the classification category, classification reference sentence, and classification rule given in step ST3-1 will be described, and specific classification processing will be described. FIG. 7 is a diagram illustrating an example in which the above-described classification rules are applied to the classification target documents classified hierarchically. In the figure, reference numeral 36 indicating an area surrounded by a broken line indicates that the rule number of the classification rule applied to the classification target document classified into the first hierarchy is 1, that is, FIG.
This indicates that the classification rule indicated by reference numeral 31 in the middle is applied. The hierarchy column indicates the hierarchical order in the classification of the documents to be classified, where 1 indicates the document to be classified in the first layer, and 2 indicates the document to be classified in the second layer subordinate to the first layer.

FIG. 8 is a diagram showing classification categories used by the automatic document classification system according to the first embodiment and classification reference sentences corresponding thereto. In the figure, reference numeral 41 indicating a region surrounded by a broken line has a classification category C1 or C2 as a classification destination of the parent classification category C0, and a sentence "memory" or "printer" (or a word) as a classification reference sentence. Indicates that is clustered. Similarly, the reference numeral 42 surrounded by a broken line indicates the classification category C3 or C3 as the classification destination of the parent classification category C1.
4 and the classification criteria sentence is "insufficient memory"
The sentences "out of memory" and "expansion of memory" are clustered. In addition, the classification category C0
Include the uncategorized document to be classified, that is, the entire document to be classified.

Here, it is assumed that two classification categories C1 and C2 indicated by reference numeral 41 in FIG. 8 have been given to the automatic document classification system according to the first embodiment in step ST3-1. Thereafter, in step ST3-2, the document analyzing unit 4 analyzes the sentence structure of the classification reference sentences "memory" and "printer" given to the two classification categories.
In this case, since the classification reference sentences in the two classification categories C1 and C2 are only words (“memory” and “printer”), the words are the analysis results as they are.

Next, since the parent classification category is C0 for these two classification standard sentences, the result of analyzing the free answer contents in FIG. 4 and the analysis of "memory" and "printer" Based on the result, the similarity is calculated in accordance with the weight of rule number 1 indicated by reference numeral 36 in FIG. 7 (parameter set for each similarity of rule number 1 indicated by reference numeral 36 in FIG. 6), and the classification destination is determined. decide. That is, step ST
In step 3-3, the document classification unit 9 extracts the documents to be classified whose sentence structures have been analyzed from the classification target document analysis result storage unit 6 one by one. Since the document is not classified at first, the classification category C0 including the entire document to be classified is set. after this,
The document similarity calculation means 8 selects the analysis result of the questionnaire number 1 from the document classification means 9 and calculates the similarity in step ST3-4.

To explain this similarity calculation specifically, for example, the similarity between the analysis result of “memory” and the questionnaire number 1 “memory cannot be added” has a common independent word of “memory”. Since the phrase similarity parameter in rule number 1 is 0.8, the similarity 0.8
Is calculated. On the other hand, the similarity between the analysis result of "Printer" and the questionnaire number 1 "Cannot add memory"
Since there is no common independent word, the similarity is 0. Therefore, in step ST3-5, the document classifying means 9 classifies the document to be classified with the questionnaire number 1 into the classification category C1 that is most similar based on the similarity.

For the other questionnaire numbers 2 to 2000, the processing of steps ST3-3 to ST3-5 is similarly executed to calculate the similarity. Here, the classification target documents having a similarity of 0 for all the classification categories are classified into the “other” category. When all the documents to be classified have been processed, in step ST3-6, the final set of (category category, questionnaire number list) is (C1, [1,.
3, 7, 8, 12,. . . ]), (C2, [1
1,. . . ]), (C0 ′, [2, 10,...]) Are obtained. Here, Ci ′ indicates an “other” category that does not match any of the specified classification categories.

Next, the classification category C1 is similarly classified according to the processing flow of FIG. This classification category C1
Are two classification categories C3, indicated by reference numeral 42 in FIG.
Classified as C4. Here, the classification category C3 has two classification reference sentences “insufficient memory” and “insufficient memory”.
Are clustered. 1 for classification category C4
One of the classification criteria sentences “memory is added” is clustered.

When classifying such a classification category C1, the classification rule of layer 2 (rule number 2) shown in FIG.
Is applied. Rule number 2 calculates the similarity using α = 0.2, β = 0.8, and γ = 1.0 in equation (B) as shown in FIG. It is considered important as a point. As a result of the analysis, as a result of the analysis, the classification target documents of questionnaire number 7 and questionnaire number 8, which have similar dependency structures to the classification standard sentence “insufficient memory” and “insufficient memory” corresponding to classification category C3, are obtained. It is classified into the classification category C3.

Further, questionnaire numbers 1 and 3 having a similar dependency structure to the classification reference sentence "addition of memory" corresponding to the classification category C4 are classified into the classification category C4.
At this time, the classification target sentences of questionnaire numbers 1 and 3 have different modalities but have γ = 1.0, and therefore have the same similarity to the classification reference sentence “memory is added”. Therefore, (C3, [7, 8,...]), (C4, [1,.
3,. . . ]), (C1 ′, [12,...]).

Finally, the classification standard sentence "I want to add memory" corresponding to the classification category C5 and the classification standard sentence "Memory cannot be added" corresponding to the classification category C6 are compared with the questionnaire numbers 1 and 3, respectively. I do.
Since the third-level classification is performed on the classification category C4, rule number 3 is applied as shown in FIG. Therefore, α = 0.2, β = 0.8, γ = 0.1, and the difference in modality becomes the difference in similarity.
Is classified into the classification category C6, and the questionnaire number 3 is classified into C5.

FIG. 9 is a diagram showing classification results of the automatic document classification system according to the first embodiment. As described above, the documents to be classified are hierarchically classified according to the point of interest of the classification, and the classifications C1, C2, and C0 are set as the first layer from the “unclassified” “whole” (classification category C0). "Memory", "Printer", and "Other" corresponding to
"Insufficient memory", "Additional memory", and "Other" corresponding to the classification categories C3, C4, and C1 'as the second hierarchy from the classification category C1.
Finally, "want to add memory" and "cannot add memory" corresponding to classification categories C5 and C6 as the third hierarchy from classification category C4.
Are classified into two categories.

In the above embodiment, the similarity calculation between the classification standard sentence and the document to be classified is performed by simply calculating the similarity between the result of analyzing the classification standard sentence and the result of analyzing the document to be classified. When the classification target document is composed of multiple sentences, the similarity is calculated between the analysis result of each sentence of the classification target document and the analysis result of the classification reference sentence, and the maximum similarity is calculated based on the classification reference sentence. And the similarity between the documents to be classified. As a result, classification can be performed by using a point of interest that is more suitable for the purpose of the user, and classification can be further speeded up.

Further, in the above embodiment, the similarity calculation between the classification reference sentence and the classification target document is performed by calculating the similarity between the analysis result of each sentence of the classification target document and the analysis result of the classification reference sentence. If the degree is equal to or more than a certain threshold, the data may be classified into a plurality of classification categories. As a result, it is possible to perform classification using a point of interest that is more suitable for the purpose of the user.

In the present embodiment, the classification category and its classification reference sentence are set in advance, and the automatic classification is performed.
An interface may be provided to the user of the system of the present invention for setting a lower hierarchy to be classified in more detail for an arbitrary classification category, and the classification may be performed interactively and stepwise. As a result, it is possible to perform classification using a point of interest that is more suitable for the purpose of the user.

Further, in the above embodiment, the calculation of the phrase similarity of the formula (B) is performed on the classification reference sentence and all the independent words appearing in the document to be classified. It may be possible to limit to only the word or only the word. Thereby, classification can be speeded up.

In the embodiment, the threshold value of the similarity is not dependent on the classification rule, and the processing is performed assuming a constant value. However, the threshold value may be set for each classification rule.
As a result, it is possible to perform classification using a point of interest that is more suitable for the purpose of the user.

As described above, according to the first embodiment, the document analysis means for analyzing the classification reference sentence provided corresponding to the classification category and the sentence structure of the classification target document to be classified into the classification category. 4 and clustering of the classification reference sentence based on the analysis result by the document analysis means 4, and according to the analysis result of the document to be classified and the point of interest of the clustered classification reference sentence according to the classification rule that defines the point of interest of the classification. Document similarity calculating means 8 for calculating similarity
And a document classifying means 9 for determining the classification category of the document to be classified based on the similarity calculated by the document similarity calculating means 8, so that the classification reference sentence as sample data is clustered into the classification categories C0 to Cn. Being
By using the similarity between these and the documents to be classified,
Automatic classification that requires learning with a large amount of learning documents can be automatically classified with a small amount of sample data. Also,
Since the similarity is calculated according to the classification rule that defines the point of interest in the classification, the classification can be performed using the point of interest that matches the purpose of the user.

Further, according to the first embodiment, the same effect as described above can be obtained because the classification rule defines the point of interest in the classification by the feature amount of the document in the natural language.

Further, according to the first embodiment, the classification rule forms the phrase similarity indicating the similarity of each of the phrases forming the classification reference sentence and the document to be classified, and the classification reference sentence and the document to be classified. At least one of the dependency similarity indicating the similarity of the dependency between each clause and the modal similarity indicating the similarity of the modal expression between the classification reference sentence and the classification target document is a feature amount of the document in the natural language. , So that the same effects as above can be obtained.

Further, according to the first embodiment, the classification rule defines a plurality of points of interest at the time of classification, sets a threshold value according to the purpose of classification for the similarity of each point of interest, 9 determines the classification category of the document to be classified based on the similarity greater than or equal to the threshold value, so that the classification can be performed using a point of interest more suited to the purpose of the user.

Further, according to the first embodiment, the document to be classified is composed of a plurality of sentences, and the document similarity calculating means 8 calculates the similarity to the document to be classified in units of sentences. Speed can be increased.

Further, according to the first embodiment, the document similarity calculating means 8 determines the highest similarity among the similarities between the sentence constituting the classification target document and the classification reference sentence, by using Since the degree of similarity to the classification reference sentence is used, classification can be performed using a point of interest that more closely matches the purpose of the user, and classification can be further speeded up.

Further, according to the first embodiment, the document classifying means 9 classifies the documents to be classified hierarchically according to the point of interest of the classification, and the classification rule is set in advance for each layer corresponding to the point of interest of the classification. Since it is set, classification can be performed using a point of interest more suited to the purpose of the user.

[0077]

As described above, according to the present invention, the classification reference sentence provided corresponding to the classification field and the sentence structure of the classification target document to be classified into the classification field are analyzed, and the analysis result is obtained. Based on the classification rules, the classification reference sentence is clustered, and the similarity relating to the point of interest between the analysis result of the document to be classified and the clustered classification reference sentence is calculated in accordance with the classification rule that defines the point of interest in the classification. Since the classification field of the document to be classified is determined by using the similarity between the clustered classification reference sentence as sample data and the document to be classified, automatic classification that requires learning with a large number of learning documents can be performed. There is an effect that automatic classification can be performed with a small amount of sample data. Further, since the similarity is calculated according to the classification rule that defines the point of interest of the classification, there is an effect that the classification can be performed using the point of interest that matches the purpose of the user.

According to the present invention, since the classification rule defines the point of interest of the classification by the characteristic amount of the document in the natural language, the same effect can be obtained by applying to the configuration of paragraph 0077.

According to the present invention, the classification rule indicates the similarity between the phrases forming the classification reference sentence and the document to be classified, and the relation between the phrases forming the classification reference sentence and the document to be classified. Since at least one of the dependency similarity indicating the similarity of the subject and the modal similarity indicating the similarity of the modal expression between the classification reference sentence and the classification target document is used as the feature amount of the natural language document, A similar effect can be obtained by applying to the configuration of paragraph 0078.

According to the present invention, the classification rule defines a plurality of points of interest to be focused on at the time of classification, sets a threshold according to the purpose of classification for the similarity related to each point of interest, and sets the similarity based on the similarity not less than the threshold. Since the classification field of the document to be classified is determined, there is an effect that the classification can be performed by using a point of interest more suited to the purpose of the user.

According to the present invention, the similarity is calculated for each document to be classified composed of a plurality of sentences in units of sentences, so that there is an effect that the classification can be speeded up.

According to the present invention, the similarity indicating the highest value among the similarities between the sentence constituting the classification target document and the classification reference sentence is determined as the similarity between the classification target document and the classification reference sentence.
There is an effect that the classification can be performed by using a point of interest more suitable for the purpose of the user, and the classification can be further speeded up.

According to the present invention, when the classification rule uses the phrase similarity indicating the similarity of each of the phrases constituting the classification reference sentence and the document to be classified as the feature amount of the document in the natural language, Since the phrase similarity is calculated from only phrases that are composed of words or words composed of words of declinable words, there is an effect that classification can be speeded up.

According to the present invention, the documents to be classified are classified hierarchically according to the points of interest in the classification, and the classification rules are preset for each layer according to the points of interest in the classification. There is an effect that it is possible to perform classification using a point of interest that matches with.

According to the present invention, when the degree of similarity between the classification target document and the plurality of classification reference sentences is equal to or more than a predetermined threshold, the classification target document is classified into a classification field corresponding to each of the plurality of classification reference sentences. Therefore, there is an effect that classification can be performed by using a point of interest more suitable for the purpose of the user.

According to the present invention, a plurality of classification reference sentences are set for one classification field, the similarity between each classification reference sentence and the document to be classified is calculated, and among the calculated similarities,
Since the classification target document is classified into the classification field corresponding to the classification reference sentence having the highest similarity, the classification can be performed using a point of interest more suitable for the user's purpose. Has the effect that the speed can be further increased.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an automatic document classification system according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing an operation of the automatic document classification system according to the first embodiment of the present invention;

FIG. 3 is a diagram illustrating a freely described questionnaire as an example of a classification target document classified by the automatic document classification system according to the first embodiment;

FIG. 4 is an example showing the result of analyzing the sentence structure of the answer to questionnaire number 1 in FIG.

FIG. 5 is a flowchart showing a similarity calculation process and a document classification process in the automatic document classification system according to the first embodiment.

FIG. 6 is a diagram showing an example of a classification rule used by the automatic document classification system according to the first embodiment.

FIG. 7 is a diagram illustrating an example in which a classification rule is applied to documents to be classified that are classified hierarchically.

FIG. 8 is a diagram showing classification categories used by the automatic document classification system according to the first embodiment and classification reference sentences corresponding thereto.

FIG. 9 is a diagram showing a classification result of the automatic document classification system according to the first embodiment.

FIG. 10 is a block diagram showing a configuration of a conventional automatic document classification system.

[Explanation of symbols]

1 dictionary, 2 classification standard sentence storage section, 3 classification target document storage section, 4 document analysis means (document structure analysis means), 5 classification standard sentence analysis result storage section, 6 classification target document analysis result storage section, 7 classification rule storage Section, 8 document similarity calculation means (similarity calculation means), 9 document classification means, 10 classification result storage section, 31 classification rules, 36 layers and rule numbers, 41 parent classification category, classification category and classification reference statement, 42 parent Classification category, classification category and classification standard statement, C0 to C6 classification category (classification field).

Continuation of front page (72) Inventor Katsushi Suzuki 2-3-2 Marunouchi, Chiyoda-ku, Tokyo F-term in Mitsubishi Electric Corporation (reference) 5B075 ND03 NK32 NR12 PR06 QM08

Claims (21)

[Claims] 1. A document structure analyzing means for analyzing a classification reference sentence provided corresponding to a classification field and a sentence structure of a document to be classified to be classified in the classification field, and the analysis by the document structure analyzing means. Based on the result, the classification reference sentence is clustered, and a similarity between the analysis result of the document to be classified and the clustered classification reference sentence is calculated according to a classification rule that defines a point of interest for classification. An automatic document classification system, comprising: a similarity calculating unit; and a document classifying unit that determines a classification field of the classification target document based on the similarity calculated by the similarity calculating unit. 2. The automatic document classification system according to claim 1, wherein the classification rule defines a point of interest in the classification by a feature amount of a natural language document. 3. The classification rule includes a phrase similarity indicating the similarity of each of the phrases constituting the classification reference sentence and the document to be classified, and a dependency between the phrases constituting the classification reference sentence and the document to be classified. At least one of a dependency similarity indicating the similarity of a document and a modal similarity indicating a similarity of a modal expression between the classification reference sentence and the classification target document is used as a feature amount of a natural language document. 3. The automatic document classification system according to claim 2, wherein: 4. A classification rule defines a plurality of points of interest at the time of classification and sets a threshold value according to the purpose of classification for the similarity of each point of interest. 2. The automatic document classification system according to claim 1, wherein a classification field of the classification target document is determined based on the degree of similarity. 5. The classification target document is composed of a plurality of sentences, and the similarity calculation means calculates a similarity between the classification target document and each sentence. The document automatic classification system according to any one of the above. 6. The similarity calculating means determines a similarity indicating the highest value among similarities between a sentence constituting a document to be classified and a classification reference sentence by calculating a similarity between the document to be classified and the classification reference sentence. 6. The automatic document classification system according to claim 5, wherein 7. A similarity calculating means, wherein the classification rule uses a phrase similarity indicating the similarity of each of the phrases constituting the classification reference sentence and the classification target document as a feature amount of the natural language document. 4. The automatic document classification system according to claim 3, wherein the phrase similarity is calculated only from a phrase composed of words or only a phrase composed of words of declinable words. 8. The document classification means classifies documents to be classified hierarchically according to a point of interest of the classification, and the classification rule is preset for each layer corresponding to the point of interest of the classification. Claim 1 to Claim 7
The document automatic classification system according to any one of the above. 9. When the degree of similarity between a document to be classified and a plurality of classification reference sentences is equal to or greater than a predetermined threshold, the document classification means adds the classification target document to a classification field corresponding to each of the plurality of classification reference sentences. The automatic document classification system according to any one of claims 1 to 8, wherein the classification is performed. 10. A plurality of classification reference sentences are set for one classification field, a similarity calculation unit calculates a similarity between each of the classification reference sentences and a document to be classified, and the document classification unit sets the similarity. 2. The classification target document according to claim 1, wherein the classification target document is classified into a classification field corresponding to a classification reference sentence having a similarity indicating the highest similarity value among the similarities calculated by the degree calculation means. Automatic document classification system. 11. A document structure analyzing step for analyzing a classification reference sentence provided corresponding to the classification field and a sentence structure of a document to be classified to be classified into the classification field, and the analysis in the document structure analysis step. Based on the result, the classification reference sentence is clustered, and a similarity between the analysis result of the document to be classified and the clustered classification reference sentence is calculated according to a classification rule that defines a point of interest in the classification. An automatic document classification method, comprising: a similarity calculation step; and a document classification step of determining a classification field of the classification target document based on the similarity calculated in the similarity calculation step. 12. The classification rule according to claim 1, wherein a point of interest of the classification is defined by a feature amount of a document in a natural language.
1. The document automatic classification method according to 1. 13. The classification rule is a phrase similarity indicating the similarity between the phrases forming the classification reference sentence and the document to be classified, and the dependency between the phrases forming the classification reference sentence and the document to be classified. At least one of a dependency similarity indicating the similarity of a document and a modal similarity indicating a similarity of a modal expression between the classification reference sentence and the classification target document is used as a feature amount of a natural language document. Claim 1 characterized by the following:
2. Automatic document classification method described in 2. 14. A classification rule defines a plurality of points of interest at the time of classification, sets a threshold value according to the purpose of classification for the similarity of each point of interest, and, in the document classification step, The method according to claim 11, wherein a classification field of the document to be classified is determined based on the similarity. 15. The method according to claim 11, wherein in the similarity calculation step, the similarity is calculated for each classification target document composed of a plurality of sentences. Item 1. The automatic document classification method according to Item 1. 16. A similarity calculating step, wherein the similarity indicating the highest value among the similarities between the sentence constituting the classification target document and the classification reference sentence is determined by the similarity between the classification target document and the classification reference sentence. The automatic document classification method according to claim 15, wherein 17. In the similarity calculation step, when the classification rule uses the phrase similarity indicating the similarity of each of the phrases forming the classification reference sentence and the classification target document as a feature amount of the natural language document, 14. The automatic document classification method according to claim 13, wherein the phrase similarity is calculated only from a phrase composed of words of the above or only a phrase composed of words of the declinable word. 18. In the document classification step, the classification target documents are classified hierarchically according to the point of interest of the classification, and the classification rule is set in advance for each layer corresponding to the point of interest of the classification. The document automatic classification method according to any one of claims 11 to 17, characterized in that: 19. When the similarity between a document to be classified and a plurality of classification reference sentences is equal to or greater than a predetermined threshold value in the document classification step, the classification target is added to the classification field corresponding to each of the plurality of classification reference sentences. 19. The automatic document classification method according to claim 11, wherein the documents are classified. 20. A plurality of classification reference sentences are set for one classification field, a similarity calculation step calculates a similarity between each of the classification reference sentences and the document to be classified, and a document classification step. Wherein the classification target document is classified into a classification field corresponding to a classification reference sentence having a similarity indicating a value having the highest similarity among the similarities calculated in the similarity calculating step. 11. The document automatic classification method according to item 11. 21. A document structure analysis function for analyzing a sentence structure of a classification reference sentence provided corresponding to a classification field and a classification target document to be classified in the classification field, and the analysis by the document structure analysis function. Based on the result, the classification reference sentence is clustered, and a similarity between the analysis result of the document to be classified and the clustered classification reference sentence is calculated according to a classification rule that defines a point of interest in the classification. A computer-readable recording medium that records an automatic document classification program having a similarity calculation function and a document classification function of determining a classification field of the classification target document based on the calculated similarity.