KR20240168118A - Document Classification System and Method Using Entity Group Scores - Google Patents

Abstract

The present invention relates to a method and a system for performing document classification using an existing entity name recognition model. According to the present invention, an entity name recognition model generated in advance is used to recognize an entity name of a document, an entity name group score for calculating importance of entity names is calculated, an expression vector for the document is obtained by using the calculated entity name group score, and the documents are classified by analyzing a document expression vector.

Description

Document Classification System and Method Using Entity Group Scores

The present invention relates to a document classification system and method based on named entity recognition and unique named entity group scoring, and more particularly, to a document classification system and method for accurately classifying a document including multiple named entities.

The task of recognizing text information consists of various types of subtasks such as entity recognition, document classification, document summary, relationship analysis, and Q&A. In order to perform these tasks with supervised learning, time and effort must be invested in generating appropriate learning data for each subtask, which may delay the construction of a comprehensive document analysis system.

In most cases, researchers try to save the effort of data preparation by utilizing open data (such as benchmark datasets) suitable for each subtask. However, in the case of research literature in the field of science and technology, open general data that fits the characteristics of the field are often not provided. This is because many people are unable or cannot deal with the topic due to the scarcity and difficulty of the field of science and technology. This situation has been an unavoidable situation for many researchers in the field of science and technology for a long time.

In this case, if the results of performing one task can be used to perform another task, the time required for natural language processing can be saved.

In addition, document classification is an important task in various fields such as information retrieval, knowledge management, and natural language processing. As the amount of digital documents increases rapidly, automatic classification technology is essential for efficient document management and retrieval.

In this regard, various techniques including keyword-based methods, topic modeling, and deep learning techniques have been proposed for document classification, but these techniques often do not consider the relationships between named entities in the document, and thus have limitations in accurately classifying documents containing multiple entities.

Patent Document 1: US Registered Patent No. US 9760634 B1 Patent Document 2: Republic of Korea Registered Patent No. KR 2006646 B1

Accordingly, the present invention provides a method and device for classifying documents by using a named entity recognition model used in a named entity recognition process as a subtask of existing natural language processing in a method and system for document classification.

In order to solve the above-described problem, the present invention provides a document classification system for document classification, comprising: a. a control unit; b. a document input unit configured to receive document data for a plurality of documents included in a predetermined document pool; c. an entity recognition unit that recognizes predetermined entity names included in the document data of the plurality of documents to recognize entity names in the documents; d. a document entity group score calculation unit that groups the entities based on the entity recognition result for each document and calculates a score for each entity group; e. a document classification unit that classifies the plurality of documents based on the entity group scores calculated by the entity group score calculation module.

Meanwhile, the entity group score Sn for the above-mentioned nth entity group is calculated by the following (Formula 1).

(Formula 1)

Here, ecp represents the number of entities included in the entire document pool, ecd represents the number of entities in the target document for calculating scores by entity group, dfe represents the frequency of documents containing a specific entity group, and ld represents the length of the target document.)

At this time, the document classification unit is characterized by using the calculated entity group scores as feature values representing the document representation of each document to classify the plurality of documents.

In addition, the document classification system according to the present invention further includes a vector space analysis unit that maps the calculated named entity group score to a named entity group and performs dimension reduction and cluster analysis based on the same to generate an expression vector and cluster information for each document; and the document classification unit identifies and classifies documents with similar characteristics based on the expression vector and cluster information generated by the vector space analysis unit.

In addition, the present invention provides a document classification method, comprising: a. a document receiving step of receiving document data for a plurality of documents included in a predetermined document pool; b. an entity recognition step of recognizing predetermined entity names included in the document data of the plurality of documents to recognize entity names in the documents; c. a score calculation step of grouping entities based on the entity recognition result for each document and calculating a score for each entity group; d. a document classification step of classifying the plurality of documents based on the scores for each entity group.

According to the present invention, by utilizing a pre-learned named entity recognition model for document classification work, computer computational resources are saved, and further, by calculating named entity group scores indicating the importance of each named entity in a document pool and individual documents from the named entity recognition results and using these for document classification, a document classification method and system that are more effective than conventional methods are provided.

The following drawings attached to this specification illustrate preferred embodiments of the present invention and, together with the detailed description of the invention described above, serve to further understand the technical idea of the present invention, and therefore, the present invention should not be construed as being limited to matters described in such drawings.
FIG. 1 is a block diagram showing a document classification system according to an embodiment of the present invention, FIG. 2 is a flow chart showing a document classification method according to an embodiment of the present invention, FIG. 3 is a table showing examples of entity group scores calculated for recognized entity names in certain documents, FIG. 4 is an example diagram explaining vector dimensionality reduction analysis results, FIG. 5 is an example diagram explaining vector space analysis and document classification processes, and FIG. 6 is a block diagram of a computer system performing a document classification method according to the present invention.

FIG. 1 is a block diagram showing a document classification system according to an embodiment of the present invention. The document classification system includes a control unit 110, a document input unit 120, a named entity recognition unit 130, a document named entity group score calculation unit 140, a vector space analysis unit 150, and a document classification unit 160. The system according to the present invention may also include a memory device 170 for storing a named entity recognition model and a document classification model. The memory device 170 is for storing the named entity recognition model and the document classification model.

The document input unit 120 is configured to receive document data for a plurality of documents included in a predetermined document pool. The document data may include text, metadata, or other related information.

The entity name recognition unit 130 is configured to recognize a predetermined entity name included in document data of multiple documents. An entity is specific information related to a document and may include a name, date, location, organization, or other specific information.

The document entity group score calculation unit 140 calculates the entity group score for each entity group based on the entity recognition result of each document. The "entity group" refers to a set of entity names (nouns or phrases) related to a specific category in a document. The entity group includes a set of keywords found in a context related to the category and a set of words surrounding the set of keywords. For example, if the "entity group" category is 'medicine', the group may be composed of words and phrases related to medicine (e.g., disease names, treatments, drugs, etc.).

The present invention is characterized by classifying documents using such entity group scores. The present invention uses the analysis of the relationship between a set of keywords related to a specific category and a set of words appearing around them to determine whether a new document is related to a specific category.

In one embodiment, the entity name group score of a given entity name group n is calculated according to the following mathematical expression 1.

(Sn represents a named entity group score of a given named entity group n according to the first aspect of the present invention, ecp represents the number of entities in the entire document pool, ecd represents the number of entities in the target document for calculating the named entity group score, dfe represents the frequency of documents including a specific named entity group, and ld represents the length of the target document.)

In another embodiment, the entity group score can be calculated according to the following mathematical expression 2.

(Sn' is the score of the nth entity group according to the second aspect of the present invention, frequency = sum (number of word mentions per entity) / total document length, weight is a value set in advance by the user for each entity)

An example of weight in the above mathematical expression 2 is as follows.

biodegrad_poly=1, biodegrad_prop=1, mechanical_prop=0.7, poly_struc=0.7, rheological_prop=0.7

Next, the vector space analysis unit 150 maps the calculated named entity group scores to the corresponding named entity groups, generates a representation vector for each document based on this mapping, and performs dimensionality reduction analysis and clustering analysis based on this.

The document classification unit 160 classifies a plurality of documents based on the calculated named entity group scores, expression vectors, and cluster information. In one embodiment, the document classification unit 160 classifies a plurality of documents using the named entity group scores as a feature value representing the document expression.

The method by which the document classification unit 160 classifies documents can be performed by outputting the results of dimension reduction and cluster analysis performed in the vector space analysis unit 150 as visual graphic data, or by tagging each document data with a document classification category based on the results of the vector analysis for each document.

For each document, tagging the document classification category based on the results of the above vector analysis can be done by assigning a tagging value of the corresponding cluster to documents corresponding to expression vectors transcribed within a predetermined vector distance from the center point of each cluster as a result of the cluster analysis.

The memory device 170 stores an entity recognition model configured to recognize an entity name and a document classification model configured to classify a document using document representation feature values. The entity recognition unit 130 and the document classification unit 160 access the memory device 170 to use the entity recognition model and the document classification model, respectively.

Drawing 2 is a flowchart illustrating a document classification method according to an embodiment of the present invention. According to this method, it is configured to include a document receiving step S210, an entity recognition step S220, a score calculation step S230, a vector space analysis step S240, and a document classification step S250.

In the document receiving step S210, document data for a plurality of documents included in a predefined document pool is received. The document data may include text, metadata, or other related information.

In the entity name recognition step S220, a predetermined entity name included in document data of a plurality of documents is recognized. The entity name may include a name, date, location, organization, or other specific information related to the document.

In the score calculation step S230, the score of each entity group is calculated based on the entity recognition result of each document. This has been described in detail above with respect to FIG. 1 and the document entity group score calculation module 140.

In the vector space analysis step S240, the calculated named entity group scores are mapped to each named entity group, an expression vector for each document is generated based on this mapping, the expression vector of each document is transcribed to a vector space, and dimensionality reduction and/or cluster analysis is performed.

In the document classification step S250, a plurality of documents are classified based on the calculated named entity group scores, expression vectors, and cluster information. In one embodiment, the document classification step S250 classifies a plurality of documents using the named entity group scores as feature values representing document expressions.

Figure 3 shows a table showing the results of entity recognition for certain documents and the results of calculating entity group scores for each entity group of each document for the recognized entity names.

Figure 3 (a) shows entity names recognized from a document with the document title “How the composition and manufacturing parameters affect insulin release from polymeric nanoparticles.”

Fig. 3 (b) shows the case where five entity groups, biodegrad_poly, poly_struc, rheological_prop, biodegrad_prop, and mechanical_prop, are set, the entity group score is calculated for each entity in Fig. 3 (a) according to mathematical expression 1, each entity is mapped to each entity group, and each entity group score is calculated. The 'title' column in the table represents the titles of each document. For example, the calculated entity group score of the biodegrad_poly entity group of the title1 document is 11.

Figures 4 and 5 are exemplary diagrams showing the vector space analysis and document classification process based on the calculated named entity group scores. In this example, the named entity group scores are mapped to named entity groups, a representation vector for each document is generated based on this mapping, and the representation vector for each document is transferred to vector space and dimensionality reduction and cluster analysis are performed.

In the present invention, vector analysis such as dimension reduction analysis and clustering analysis applies known techniques, and the present invention is characterized by using expression vectors based on entity group scores for entity groups included in documents as inputs for clustering analysis in classifying documents rather than analysis techniques.

In Figs. 4 and 5, the expression vector for each document can be generated based on the entity group score. In addition, cluster information is also generated to identify clusters of documents with similar characteristics. Fig. 4 is a diagram showing the results of dimensionally reducing the expression vectors for each document into 1D, 2D, and 3D based on the calculated entity group score, and Fig. 5 is a diagram showing the results of classifying each document into cluster 0 to cluster 5 using the expression vector based on the calculated entity group score.

In the graphs of Figures 4 and 5, PCA means principal component analysis (PCA) in dimension reduction analysis, and Clustering KMeans indicates classification according to the "K-means clustering" technique. In addition, pca_vec1, pca_vec2, and pca_vec3, which indicate each axis, represent the three principal components with the highest explanatory power for the data among the axes of the reduced dimension. This is in accordance with the known principal component analysis technique, and the first principal component extracted from the principal component analysis is extracted in the direction with the greatest variance in the data, the second principal component is orthogonal to the first principal component and in the direction with the next greatest variance in the data, and the third principal component is extracted in the direction with the greatest variance in the direction perpendicular to the first and second principal components. Since the principal components extracted in this way have a great explanatory power for the data, they are mainly used when reducing the dimensionality of the data, and by projecting the data onto the three principal components and visualizing it, the pattern of the document data can be identified.

The table in (b) of Fig. 5 is a table that identifies and shows representative documents of each group from a document group that is clustered from 0 to 5 by vector analysis. There are known techniques for identifying representative objects in clustering analysis, and in this example, the K-means clustering technique is applied, and the document having the expression vector that becomes the center point among the expression vectors of each document becomes the representative document of each cluster. By identifying and showing representative documents in this way, the characteristics of each cluster group can also be grasped.

In summary, the present invention provides a document classification system and method that effectively classifies documents by recognizing entities in a document, calculating a named entity group score of each document based on the recognized entities, and using the named entity group score as a feature value representing each document. The vector space analysis module and the document classification module further enhance the classification process by generating expression vectors and cluster information based on the named entity group score. This enables efficient and accurate classification of documents with similar characteristics.

FIG. 6 is an exemplary block diagram showing a computer system for executing a document classification method according to the present invention. The computer system includes a processor 610, a memory 620, a document input module 630, a named entity recognition module 640, a named entity group score calculation module 650, a vector space analysis module 660, and a document classification module 670.

The processor 610 is configured to execute instructions stored in the memory 620, which may include an entity recognition model and a document classification model. The processor 610 controls the overall operation of the computer system and performs a document classification method according to the present invention.

Memory 620 stores data and instructions necessary for processor 610 to execute the document classification method. Memory 620 may include various types of memory devices, such as RAM, ROM, flash memory, and other non-volatile memory.

The document input module 630 is configured to receive document data of multiple documents included in the document pool, and the document data may be input from an external memory device through an input unit of a computer system or from an external device through a communication unit (not shown). The document data may include text, images, or other types of data related to the document.

The named entity recognition module 640 recognizes predefined entities included in document data of multiple documents. The named entity recognition module 640 can identify and recognize predefined entities using a named entity recognition model stored in the memory 620.

The named entity group score calculation module 650 calculates a named entity group score for each document based on the entity recognition results. The named entity group score can be calculated using mathematical expression 1 or mathematical expression 2 as described above.

The vector space analysis module 660 maps the calculated named entity group scores to named entity groups and performs dimension reduction and cluster analysis based on this mapping. The vector space analysis module 660 generates a representation vector and cluster information for each document.

The document classification module 670 classifies multiple documents based on the expression vector and cluster information generated by the vector space analysis module 660. The document classification module 670 identifies and classifies documents with similar characteristics. The document classification module 670 can generate output data to output the document classification results externally.

The above modules constituting the computer system of FIG. 6 may be a collection of computer-executable algorithms or programs configured to perform each of the above operations by being stored in the memory 620 and read by the processor 610.

In addition to the embodiments described above, the present invention can be implemented as a computer-readable recording medium storing a computer algorithm for performing a document classification method. The computer-readable recording medium can include various types of storage media such as optical disks, magnetic disks, flash memory devices, and other nonvolatile memory devices.

In addition, the present invention can be implemented as a computer program product including a computer-readable storage medium having computer-readable program code embedded therein. The computer-readable program code can be executed by a processor to perform a document classification method according to the present invention.

The present invention can be implemented as a computer system including a processor and a memory configured to perform the document classification method described above. The processor controls the overall operation of the computer system and performs the document classification method by executing instructions stored in the memory. The present invention can also be implemented as a method performed by a processor that executes instructions for performing the document classification method described above.

Below, examples according to the present invention are described in more detail.

<Example 1>

Embodiment 1 of the present invention provides a method for saving human and material resources by utilizing an existing entity name recognition model for document classification. The method includes the following steps.

First, for the field of interest, "biodegradable polymer materials", a named entity recognition model was created to recognize technical terms corresponding to the named entity groups covering five concepts in the literature: ① 'biodegradable polymer', ② 'polymer structure', ③ 'rheological properties', ④ 'biodegradation characteristics', and ⑤ 'mechanical properties'. The named entity recognition model used a known artificial intelligence model.

A large number of related documents were applied to the above-mentioned entity recognition model, and technical terms recognized as entities in the documents were assigned to each entity group. The entity group score was calculated using mathematical expression 1, and an abstracted "representation vector" corresponding to each document was obtained from each entity group score.

This is accomplished through the following process:

a. The calculated entity group score is calculated for each entity group. For example, entity group scores are calculated for all entities belonging to entity group ①_'biodegradable polymer', and entity group scores are calculated for all entities belonging to entity group ②_'polymer structure'.

b. For each entity group, an abstract "expressor vector" is generated based on the calculated entity group score. The expressor vector is a vector representation of the entity group score for each entity group in each document. For example, if there are a total of five entity groups, the expressor vector is represented as a five-dimensional vector, and each dimension of the vector represents the score of the corresponding entity group.

Afterwards, the descriptor vector for each document is transferred to a vector space, the vector space is interpreted to identify clusters among documents with similar characteristics, and the characteristics of the cluster are derived through the document located at the center of the vector space, and the documents can be classified accordingly.

<Example 2>

Example 2 is the same as Example 1 except that the entity name unique score is calculated using Equation 2. In Equation 2, the weight for each term is set as follows: biodegrad_poly=1, biodegrad_prop=1, mechanical_prop=0.7, poly_struc=0.7, rheological_prop=0.7.

<Comparative Example 1>

Comparative Example 1 is a method of additionally constructing a classification model that can classify documents into desired classes regardless of whether an existing named entity recognition model is constructed.

Comparative Example 1 proceeds as follows.

a. Manually select the literature to be used for training the classification model.

b. Create learning, validation, and test data by assigning desired class labels to the learning data.

c. Learning and testing of classification model

d. Perform document classification by inputting actual document data into the learned classification model.

The method of Comparative Example 1 can classify documents using the desired classification criteria regardless of whether an existing named entity recognition model is constructed, but since an additional named entity recognition model must be created to construct a classification model, there is a problem in that additional resources must be invested.

110 Control Unit 120 Document Input Unit
130 Entity Name Recognition Section 140 Document Entity Name Group Score Calculation Section
150 Vector Space Analysis Department 160 Document Classification Department
170 memory devices
610 Processor 620 Memory
630 Document Input Module 640 Entity Name Recognition Module
650 entity group score calculation module
660 Vector space analysis module 670 Document classification module

Claims (14)

As a document classification system for document classification,
a. Control unit;
b. A document input unit configured to receive document data for multiple documents included in a specified document pool;
c. An entity name recognition unit that recognizes predetermined entity names included in document data of the plurality of documents to recognize entity names within the document;
d. For each document, a document entity group score calculation unit groups the entity names based on the entity name recognition results and calculates the entity group score for each entity group;
e. A document classification unit that classifies the multiple documents based on the entity group scores for each entity group calculated above;
A document classification system comprising: In the first paragraph,
The entity group score Sn of the above entity n is calculated by the document classification system below (Formula 1).
(Formula 1)

(Here, ecp represents the number of entities included in the entire document pool, ecd represents the number of entities in the target document for calculating the entity group score, dfe represents the frequency of documents containing a specific entity group, and ld represents the length of the target document.) In the first paragraph,
The above document classification section is,
Classifying the multiple documents by utilizing the above calculated entity group scores as feature values representing the document representation of each document;
A document classification system featuring . In the first paragraph,
A vector space analysis unit that maps the above-mentioned calculated named entity group scores to named entity groups and performs dimension reduction and cluster analysis based on the same to generate expression vectors and cluster information for each document;
A document classification system that additionally includes: In paragraph 4,
The above document classification section is,
Identifying and classifying documents with similar characteristics based on the expression vector and cluster information generated by the above vector space analysis unit;
A document classification system featuring: In paragraph 5,
It additionally includes a memory device storing an entity recognition model set to recognize an entity name and a document classification model set to classify documents by utilizing features representing the document representation of each document;
The above entity name recognition unit and document classification unit are,
Set to access the above memory device and use the above entity name recognition model and the above document classification model respectively;
A document classification system featuring . As a document classification method,
a. A document receiving step for receiving document data for multiple documents included in a given document pool;
b. An entity name recognition step for recognizing certain entity names included in document data of the plurality of documents to recognize entity names within the document;
c. For each document, a score calculation step for grouping named entities based on the named entity recognition results and calculating the named entity group score for each named entity group;
d. A document classification step for classifying the multiple documents based on the above entity name group score;
A method of classifying documents that includes . In Article 7,
The above entity group score is a document classification method calculated by the following (Formula 2).
(Formula 2)

(Here, ecp represents the number of entities included in the entire document pool, ecd represents the number of entities in the target document for calculating the entity group score, dfe represents the frequency of documents containing a specific entity group, and ld represents the length of the target document.) In Article 7,
The above document classification steps are:
A document classification method characterized by utilizing the above-mentioned calculated entity group score as a feature value representing the document representation of each document. In Article 7,
A document classification method further comprising a vector space analysis step of mapping the above-mentioned calculated named entity group scores to named entity groups and performing dimension reduction and cluster analysis based on the same to generate an expression vector and cluster information for each document. In Article 10,
The above document classification steps are:
A document classification method characterized by identifying and classifying documents having similar characteristics based on expression vectors and cluster information generated in the above vector space analysis step. A computer system performing any one of the methods of claims 7 to 11. A processor performing the method of any one of claims 7 to 11. A recording medium having recorded thereon a computer algorithm for performing any one of the methods of claims 7 to 11.

Images