KR101335540B1 - Method for classifying document by using ontology and apparatus therefor - Google Patents

Abstract

An ontology-based document classification method and apparatus suitable for using the class of each entity name modeled in the ontology and the attributes of the class for classification are disclosed.
The ontology-based document classification method extracts the types, properties, and property relations of objects that commonly occur in documents belonging to the same category in order to express the characteristics of the documents belonging to each category as ontology. An ontology modeling process of modeling the ontology by abstraction; An entity name recognition process of recognizing an entity name from documents to be classified; An entity relationship extraction process of extracting relationship information between entity names through sentence analysis for all categories; An instance peeling process of mapping object names and relationship attributes extracted from the document to object-class and object-attribute-object by comparing the ontology of each category; And a document field determination process of selecting an ontology most suitable for the document to be classified in consideration of the filled class, the relationship between the instances, and the weight of the class attribute, and determining the field of the ontology as the field of the document to be classified. .

Description

Ontology-based document classification method and apparatus {Method for classifying document by using ontology and apparatus therefor}

The present invention relates to a document classification method and apparatus, and more particularly, to an ontology-based document classification method and apparatus.

As the amount of information has increased, various techniques have been developed for effectively searching and classifying information. Existing studies have calculated the similarity between documents and classification by comparing features selected through POS tagging of documents and used the results in classification. In addition, the search engine uses the selected qualities as indexes of documents and compares the feature values with the relevance of the documents, and uses them to prioritize search.

Ontology is a method of knowledge representation and is used to model various holding resources through classes and relations between classes. Unlike general domains, in case of complaint documents, there are various types of complaints, contents of complaints, types of complaints, and related organizations.

An object of the present invention is to provide an ontology-based document classification method that utilizes a class and an attribute of a class of each individual modeled in an ontology for classification.

Another object of the present invention is to provide an apparatus suitable for the above document classification method.

Ontology-based document classification method according to the present invention for achieving the above object

In order to express the characteristics of the documents belonging to each category with ontology, the ontology is extracted by abstracting and abstracting the kinds, properties and property relations of objects that are common in the documents belonging to the same category. Ontology modeling process for modeling;

An entity name recognition process of recognizing an entity name from documents to be classified;

An entity relationship extraction process of extracting relationship information between entity names through sentence analysis for all categories;

An instance peeling process of mapping object names and relationship attributes extracted from the document to object-class and object-attribute-object by comparing the ontology of each category; And

A document field determination process of selecting an ontology most suitable for the document to be classified and considering the field of the ontology as the field of the document to be classified in consideration of the weight of the filled class, the relationship between the instances, and the class attribute. It features.

Here, the document classification method according to the present invention preferably further includes a representative document determination process of determining the documents that have the largest number of entities and relationships among entities as representative documents in documents determined in each field.

In the document classification method according to the present invention, it is preferable to further include a step of filtering the document by a keyword before the entity name recognition process.

In the object relationship extraction process, it is preferable to extract the relationship between objects based on the requirements expressed in the document.

In the entity relationship extraction process, it is preferable to refer to a general sentence pattern for extracting a sentence representing a requirement.

The document classification process

Calculating a similarity degree for each ontology based on the type and number of filled classes, weights of key attributes filled in each class, and weights of relationships among instances; And

It is preferable to include the process of selecting the ontology having the highest similarity by comparing the similarity for each ontology calculated.

The modeled ontology is

-Discriminatory factors by sector, topic, and ontology can be derived.

-Avoid ontology modeling that has a common class / property

-It is desirable to be designed under the condition that the whole ontology can be linked to the top level ontology.

Ontology-based document classification apparatus according to the present invention to achieve the above another object

An entity name extraction module for reading a document to be classified and extracting an entity included in the document;

An entity relationship extraction module for extracting a relationship between entities extracted by the entity name extraction module;

An instance peeling module for mapping the entity name extracted by the entity name extraction module and the entity relationship extraction module and relationship attributes between entities to entity-class and entity-property entities by comparing the ontology modeled for each field;

A database storing ontology for each field referred to by the instance filling module; And

The document field determination module calculates the similarity between the document and the ontology for each field by referring to the results filled in by the instance filling module for the ontology of each field, and outputs the field of the ontology having the largest similarity as the field of the document. It is characterized by including.

Here, the ontology-based document classification apparatus according to the present invention is an ontology modeling modeling the ontology for each document field based on the keyword, the feature having the selected feature through part-of-speech tagging, thesaurus, the entity name and relationship information extracted from the sampled document. It is preferable to further provide a module.

Various researches have been conducted to classify documents as a way to secure management efficiency due to the proliferation of documents, and various features representing documents have been used for classification.

The present invention proposes a method of determining a classification by comparing the ontology structure by expressing the structure of a document belonging to a classification for each classification by an ontology, extracting an object-to-object relationship attribute from a document.

If the structure of the document is expressed as an ontology, it can be used not only for classification, but also for improving the retrieval method using the attributes of the relationship between the entities and the parent-child relationship, and selecting the representative documents for each classification.

As described above, the ontology-based document classification apparatus and method according to the present invention can obtain an effect of improving the accuracy of document classification by 10-15% by using the structure information of the document on the basis of the ontology.

1 is a flowchart illustrating an ontology-based document classification method according to the present invention.
2 shows an example of modeled ontology applied to the national pension field.
3 shows an example of utilizing an ontology-based document classification apparatus according to the present invention.
4 shows a detailed configuration of a document classification apparatus according to the present invention shown in FIG.
5 illustrates the results filled by the instance filling module.
6 shows an example of a civil petition in the national pension field.
FIG. 7 shows that the main concept / relationship is derived based on the requirements of the complaint document shown in FIG. 6.
8 shows an example of ontology built for the national pension sector.
9 shows an example of ontology built for the field of fair trade.

Hereinafter, the configuration and operation of the present invention will be described in detail with reference to the accompanying drawings.

The present invention is a method for supplementing the existing classification that calculates the similarity between the document and the classification by using the terms appearing in the document as a feature, and utilizes the class and the property of each entity name modeled in the ontology.

1 is a flowchart illustrating an ontology-based document classification method according to the present invention. Referring to FIG. 1, the document classification method according to the present invention includes an ontology modeling process for each field (s102), a named entity recognition process (s104), an entity relation extraction process (s106), an instance filling process ( instance filling, s108), a field determination process s110, and a representative document selection process s112.

The ontology modeling process (s102) extracts the types, properties, and property relations of objects that commonly occur in documents belonging to the same field in order to express the characteristics of documents belonging to each field as ontology. Abstraction

Ontology refers to the fields and techniques that deal with human knowledge, and ontology, in particular, as a computer-based expression of knowledge, expresses both explicit and logical terms to express the knowledge of a certain field with a calculator. To enable the sharing and reuse of that knowledge. In ontology, classes provide the ability to group resources with the same properties and to logically express common properties. The nature of a class can be expressed by specifying the condition of the property of the class. On the other hand, an instance refers to an entity belonging to the concept.

Ontology construction is accomplished by analyzing patterns and components from sampled documents in each field and modeling the ontology based on the analysis results.

2 shows an example of modeled ontology applied to the national pension field.

It is important to design the ontology modeled by each field to show the difference in the type and number of classes and the type and number of attribute relationships with the ontology of other fields. For example, in the field of national pensions, classes and attributes related to pensions and classes and attributes related to transactions between companies in the field of fair trade between companies should be reflected in the ontology.

Once the ontology is built, it is applied to the documents to be classified and analyzed the fields of the documents.

The entity name recognition process s104 recognizes the entity name from the documents to be classified.

The process of identifying the classes and instances of classes that constitute the ontology is very important. Various entities such as human names, institution names, insurance names, legal names, etc. are determined during the entity name recognition process, and authority data and thesaurus (authority data) for the entity name are improved to improve the accuracy of entity name recognition. thesaurus) is used together.

Authority data refers to data for two notations referring to the same entity to the same entity. Authority data basically has the same record structure as bibliographic data. In the data field, various notations of the same person, general signature, and subject name are formed into one group, and the headings among them are selected based on the list rule.

A thesaurus is a lexical tool that provides information about the usage of terms and the relationships between them. Terms are generally classified into broader terms, narrower terms, use for or synonymous, related terms, and alternatives. It is mainly used to expand the meaning of the terms included in the query when searching using these relationships.

As a preliminary step of the entity name recognition process s104, keyword-based primary filtering may be performed on the document. For this purpose, a positive / negative keyword may be used. Here, the positive keyword refers to searching for a document including the corresponding keyword, and the negative keyword refers to searching for a document not including the corresponding keyword.

The entity relationship extraction process (s106) extracts relationship information between entity names through sentence analysis for all fields. In order to extract the entity relationship, the relationship information is extracted through sentence analysis based on the entity name recognized in the previous step. In this process, the relationship between various entities (name-relationship-person name, name-relationship name-organization name) is generated by considering domain and range. Since the field of the document has not yet been determined in this process, all relationships that can be extracted from all attribute names of the entire field should be extracted from the document.

The instance filling process s108 maps the object name and the relationship attribute extracted from the document into object-class and object-property-object by comparing the ontology of each field.

The classification of complaints requires tagging for insurance, pensions, transactions, etc., as well as deriving the type or additional attributes of each institution, in addition to tagging of PLO (person, location, organization) through general entity name recognition. For example, in case of a medical disaster or a health insurance related complaint, the name of the medical facility, treatment, illness, etc. are shown as the name of the individual in the complaint document. In addition, in the field of fair trade between companies, there is a civil dispute between the company and the entity, and the relationship between the entity and the company must be filled with an instance of the class.

After performing the instance filling process (s108) with the object name and the object relationship extracted from the document, the document field determination process (s110) selects the most suitable ontology in consideration of the weighted class, the relationship between the instances, and the class attributes. Then, the field of the ontology is determined as the field of the document to be classified.

In detail, the final similarity between the document and the ontology is determined in consideration of the type of the class in which the instance is filled, the weight of the main attribute of each class, the relationship between the instances, and the weight of the main relationship. In the embodiment of the present invention, all weights are set to the same for the ontology attribute, but the weight for each attribute may be adjusted by analyzing the ontology-based classification results.

The representative document determination process for each field (s112) selects the documents that have the largest number of entities and the relationships among them in the documents determined for each field as representative documents. In the practice of the present invention, the attribute name weights are applied in the same way, but it is also possible to give different attribute weights important for determining the representative document for each field.

3 shows an example of utilizing an ontology-based document classification apparatus according to the present invention. Referring to FIG. 3, based on a keyword and a feature dictionary 302, a thesaurus dictionary 304, and an entity name and relationship information 306 extracted from a sampled document having a selected feature through keyword and part-of-speech tagging, each document field may be selected. Ontologies 308 are modeled. The modeled ontology 308 is stored in a database and provided to the document classification device 310.

The document classification apparatus 310 extracts the entity name and the relationships between the entities by performing entity name recognition on the document, and determines the document field by instance filling and similarity comparison.

4 shows a detailed configuration of the document classification device 310 according to the present invention shown in FIG. Referring to FIG. 4, the document classification apparatus 310 according to the present invention includes an entity name extraction module 402, an entity relationship extraction module 404, an instance filling module 406, and a document field determination module 408. .

The entity name extraction module 402 reads a document to be classified and extracts an entity, that is, a noun, a verb, a search, an adjective, and the like from a sentence.

The entity relationship extraction module 404 extracts a relationship between entities in a sentence.

The instance filling module 406 maps the object name and the relationship attribute between the objects to object-class and object-property objects by comparing the ontology of each field.

The instance filling module 406 fills the extracted objects to each class in a plurality of ontology models (e.g., national pension, corporate regulation, fair trade, etc.) modeled using the entity name and the relationship attribute between the entities. For example, individuals extracted from 'Document No. 1', 'Document No. 2' and 'Document No. 3' are classified into seven classes (National Complaints, Pension Information, Contact Person, Disability Information, Retirement Pay Information, Wage Pay and Company) within the National Pension Ontology. The peeling result may be illustrated as shown in FIG. 5.

Each sector ontology referenced by the instance filling module 406 is stored in the database 410. The database 410 stores ontology for each field modeled by the ontology modeling module 412.

The ontology modeling module 412 includes a keyword and feature dictionary 202, a thesaurus dictionary 204, and entity names and relationship information extracted from the sampled document having the selected feature through keyword and part-of-speech tagging as shown in FIG. Model ontology-specific ontology 208 based on 206.

The document field determination module 408 calculates the similarity between the document and the ontology of each field by referring to the results filled by the field filling module 406 for the ontology of each field, and documents the field of the ontology having the largest similarity among them. Output as the field of.

In detail, the civil document field determination module 408 determines the field of the input document with reference to the result filled by the instance filling module 406.

5 illustrates the results filled by the instance filling module 406.

Document number 1, document number 2 and document number 3 of Figure 5 is the number of classes filled in each class in the national pension ontology '4', '2' and '6', respectively. In this case, since the class of the complaint is a person, if the property of the filled class is a company name such as 'International Co., Ltd.', the number of filled classes is not recognized.

Assuming that the standard number of classes for classifying a document is set to '3' or more, 'document number 1' and 'document number 3' are regarded as 'national pension'.

On the other hand, as described above, if the weights of the classes filled in each class in the National Pension Ontology are not the same, and the weights of the classes filled in the 'Citizen' class are 10% higher (weight can be increased according to the importance of the class). The number of classes filled in each class in the National Pension Ontology is '4.4', '2' and '6.6', respectively. Therefore, although the result of the civil document field determination is the same as compared with the above-described case where the weight is the same, it is a matter of course that the document field determination result may be different due to the weight.

In addition, the representative complaint document determination module 412 also refers to the result of the instance peeling, and determines the document number 3, the document having the largest number of filled classes, as a representative complaint document in the field of the national pension complaint document. ). As a result, when a document related to the 'National Pension Service' is searched, the representative complaint document 'Document No. 3' may be displayed at the top. Alternatively, the ontology may be reconstructed using the representative document.

Example

The following describes an example in which the present invention is applied to civil complaint documents in three fields, namely, "national pension", "corporate regulation" and "fair trade", in which civil complaint activity is active. A civil complaint is something a citizen applies for an administrative agency. The citizens at this time are called civil complaints, the contents of the application are called civil complaints, and the administrative agencies are responsible for handling them. The administrative agency handles the civil affairs and provides the result to the civil affairs service. This whole process is called civil affairs administration. In other words, a civil administration is an administration of activities in which an administrative body responds to a citizen's request for a specific action from an administrative body.

In order to expedite complaint services, it is necessary to automatically classify the fields of complaint documents received.

First, in order to construct ontologies for each sector, the body of the sampled complaint documents was analyzed manually.

Of the civil affairs documents, we analyzed 94/123 cases where the keyword "national pension" was found in the field of national pension, and 21/23 cases where the keyword "corporate & (and) regulation" was found in the field of corporate regulation. For the fair trade field, 76/115 cases with the keyword "corporate & (and) domineering" were found. The reason for the difference in the number of cases, such as 94/123, is that the duplicated, misclassified, and unclear requirements are excluded from the searched keywords.

FIG. 6 shows an example of a complaint document in the field of the national pension, and FIG. 7 shows a main concept / relationship derived from the requirements of the complaint document shown in FIG. 6.

The analysis results of the civil documents in the field of national pension are as follows.

* Complaint

Personal Name: Hong Gil-dong, ...

-Relationships: father, mother, ...

* Contact Person

-National Pension Service + Name + Position / Title:

 -Branch of National Pension Service

* company

Hospitals, courts and the IRS

* Contents

-Payment, late payment, payment, payment, overdue

-Payments, payments

-Pensions (age pension, disability pension, retirement pension, national pension)

* Time

-Retirement, loss

Here, the thesaurus constructed for each field / topic can be utilized.

The ontology of each field is modeled based on the data analyzed for each field.

In terms of class, the ontology

-Discriminatory factors by sector, topic, and ontology can be derived.

-Avoid ontology modeling that has a common class / property

-It is designed to link the whole ontology with the highest ontology.

From an instance point of view, the ontology

-Subjects and requirements should be extracted from the same sentence as much as possible.

-If the sentence can't be found, extract it from the text based on TF.

Instances are processed at the object name extraction level,

-Extract object name through ontology class.

Here, it is preferable to construct a general sentence pattern for extracting a sentence including a requirement. An example of a general sentence pattern is as follows.

Example 1) "Please answer ... half of the amount of the national pension."

Example 2) "Please review and punish the person in charge."

8 shows an example of ontology built for the field of "national pension".

9 shows an example of ontology built for the field of fair trade.

The results of analysis of complaint documents in the field of fair trade are as follows.

* Complaint

-Individuals, meetings, groups

-SME

* Contact Person

-Large and small companies

 Committees, Divisions

     - Public institutions

-Public Institution + Name + Person in Charge

* Target

-Contracts, land, apartments, buildings, services

* company

Hospitals, courts and the IRS

* Contents

-Profits, exploitation, unfairness, damage, tyranny, costs and payment

-Abuse, reimbursement, compensation, disaster, guarantee, bond

-Termination, default, dismissal, selection, industrial accident

-Evaluation, delivery

*area

-Place name

After doing so, the document field determination module 408 may determine the field of the inputted civil document based on the structural information of the civil document, that is, the number of peeled classes that are the object peeling results.

That is, 'document number 1', 'document number 2' and 'document number 3' of FIG. 5 have the number of classes filled in each class in the National Pension Ontology '4', '2' and '6', respectively ( Since the 'Complaint' class is a person, if the property of the filled class is a company name such as 'International Co., Ltd.', the number of filled classes is not recognized.)

Therefore, assuming that the standard number is set to '3' or more, 'Document No. 1' and 'Document No. 3' are regarded as the 'National Pension' in the field of complaint documents (see Fig. 5).

On the other hand, as described above, if the weights of the classes filled in each class in the National Pension Ontology are not the same, and the weights of the classes filled in the 'Citizen' class are 10% higher (weight can be increased according to the importance of the class). The number of classes filled in each class in the National Pension Ontology is '4.4', '2' and '6.6', respectively. Therefore, although the result of the civil document field determination is the same as compared with the above-described case where the weight is the same, it is a matter of course that the result of the civil document field determination may be different due to the weight.

In addition, the representative complaint document determination module 130 also uses the structure information of the complaint document, that is, the document number 3, which is the largest complaint document corresponding to the object peeling result, the representative complaint in the field of the National Pension Complaint Document. The document is determined and stored in the database 112. As a result, when a 'national pension' related complaint document is searched for in the search box of FIG. 5, the representative complaint document 'document number 3' may be displayed at the top.

310 ... Document sorting device
404 ... Object Name Extraction Module 404 ... Object Relationship Extraction Module
406 ... instance filling module ...

Claims (9)

In order to express the characteristics of the documents belonging to each field with ontology, the ontology is extracted by abstracting and abstracting the kinds, properties, and property relations of objects common to the documents belonging to the same field. Ontology modeling process for modeling;
An entity name recognition process of recognizing an entity name from documents to be classified;
An entity relationship extraction process of extracting relationship information between entity names through sentence analysis for all fields;
An instance filling process of mapping object names and relationship attributes extracted from a document to object-classes, object-attributes-objects by comparing the ontology for each field;
The document field determination process of referring to the instance filling process and determining the field of the ontology selected as the field of the document according to the similarity between the document to be classified and the ontology for each field in consideration of the weight of the class, the relationship between the instances, and the class attribute. ; And
Ontology-based document classification method comprising; a representative document determination process of determining the documents that have the largest number of relationships between entities and the objects constituting the ontology in the documents determined in each field. delete The method of claim 1,
Ontology-based document classification method further comprising the step of filtering the document by the keyword before the entity name recognition process. The method of claim 1, wherein the entity relationship extraction process
Ontology-based document classification method characterized by extracting the relationship between objects based on the requirements expressed in the document. The ontology-based document classification method of claim 4, wherein the entity relationship extraction process refers to a general sentence pattern for extracting a sentence representing a requirement. The method of claim 1, wherein the document field determination process is performed.
Calculating a similarity degree for each ontology based on the type and number of filled classes, weights of key attributes filled in each class, and weights of relationships among instances; And
Ontology-based document classification method comprising the step of selecting the ontology having the highest similarity by comparing the similarity for each ontology. delete An entity name extraction module for reading a document to be classified and extracting an entity included in the document;
An entity relationship extraction module for extracting a relationship between entities extracted by the entity name extraction module;
Instance filling that maps the entity name extracted by the entity name extraction module and the entity relationship extraction module and relationship properties between entities to object-class and object-property entities by comparing with ontology modeled for each field. module;
A database storing ontology for each field referred to by the instance filling module;
A document field determination module configured to calculate a similarity between the document and the ontology of each field by referring to the results of filling the ontology of each field by the instance filling module, and to output the field of the ontology having the largest similarity among them as the field of the document; And
And a representative document determination module for determining, by a document field determination module, a document that has the largest number of entities and relationships among entities in a document determined to each field as a representative document. 9. The method of claim 8,
Ontology-based document further comprises an ontology modeling module for modeling the ontology for each document field based on the feature, thesaurus, and object information extracted from the sampled document, the thesaurus having a selected feature through keyword and part-of-speech tagging Sorting device.

Images