KR20230063441A - System for recommending document using dynamic change of time-series pattern information - Google Patents

Abstract

A system for personalized document recommendation using dynamic change of time-series pattern information according to the present invention is characterized by comprising: a document object database part in which saved is a document object, a collected document through document centralization; a history information generation part which saves, when the document object is used by a user, history information including information on a user who used the document object and information on hours of use during which the document object is used in conjunction with the document object; a request information input part into which entered are requester information and a request signal including at least one request period information among a plurality of different setting periods; a first selection part selecting a used document object which is a document object used by the user corresponding to the requester information from the document object based on the history information; a recommendation document selection part selecting a recommendation document which is at least one document object having the information on hours of use relatively close to the reference point of the request period information from the used document object; and a list-providing part which generates a recommendation list information corresponding to the recommendation document and provides the generated recommendation list information to the requester. According to the present invention, it is possible to provide a recommendation document list necessary to a user by analyzing and tracing multifacetedly various patterns of a user which dynamically change over time.

Description

Customized document recommendation system using dynamic change of time-series pattern information {SYSTEM FOR RECOMMENDING DOCUMENT USING DYNAMIC CHANGE OF TIME-SERIES PATTERN INFORMATION}

The present invention is a system for recommending a document to a user, and specifically, it dramatically reduces the time required to search for a document by effectively combining analysis of time-series changes in pattern information in which a user browses or accesses a document and statistical utilization. In addition, it relates to a customized document recommendation system using dynamic changes in time-series pattern information that can maximize the convenience of users' document access and use by inducing document omissions to be effectively prevented.

In the case of the traditional method of creating documents based on a personal PC within the company and storing the documents on the PC, it is impossible to fundamentally prevent the leakage of documents due to private use, and information such as virus infection, hacking, and hard disk Since the physical/logical damage of the storage medium occurs individually, intermittently, and continuously, it is very difficult to respond adaptively to it and it is difficult to fundamentally prevent it, so it can be said that it is constantly exposed to the above-mentioned problems.

In order to solve this problem, it is possible to consider a method of strengthening security solutions or policies for each individual PC, but since this is a human-dependent method, it can be said that there is an inherent limitation that its effectiveness is inevitably low.

Recently, a cloud-based document centralization system has been disclosed in order to more fundamentally solve these problems. Such a centralized system refers to an environment or system that stores and utilizes various work results from organizations such as companies in a central server rather than a personal PC.

This document centralization can effectively solve the above-mentioned problems in that information leakage, virus inflow, and hacking prevention can be realized in a unified way and a unified and reinforced solution, etc., and in distributed processing for information / data backup It can also be said to be more effective.

In addition, such a document centralization system can provide a significant advantage in that related tasks and tasks can be performed in a ubiquitous environment without time and space constraints.

In this way, while the document centralization system has fundamental advantages such as being able to centralize or centralize document resources effectively, since a vast amount of data or information is collected, it is difficult for administrators to efficiently manage documents, etc. for general users. There is a problem in that it is difficult to accurately and quickly search/check.

Recently, in order to solve these problems, a method of classifying a large amount of centralized documents using artificial intelligence-based supervised learning or unsupervised learning or based on keywords as performed by SNS companies As a method of grouping multiple object objects (documents, etc.), etc. are disclosed.

However, since this method categorizes documents based on a technique for mechanically processing data, it is difficult to reflect the actual content of the document, and it focuses on grouping itself, so the characteristics of the group can be viewed from the user's point of view. It is difficult to intuitively and easily recognize in the document, so it can be said that the process of searching and checking documents is still difficult.

On the other hand, in the case of a system that implements document centralization, since a vast number and capacity of document objects are gathered, users who use them must be able to easily and accurately access documents to be created, documents to be confirmed, or documents to be circulated.

However, in the document centralization system, a huge number of documents are only classified into directories such as a tree structure, so users have to find and access the documents individually with their own memory and their own know-how, which takes time to access documents. Problems such as lengthening of documents, overlooking or omission of documents that need to be checked or prepared periodically can occur at any time.

In this regard, a method of recommending a plurality of documents with high preference based on the total number of documents read (including read, write, etc.) out of a vast number of documents, and a plurality of documents matching the preferences of users identified through surveys, etc. A method of recommending one document or a method of recommending a document based on the degree of matching between the content of a document (content) and the user's personal information has been partially applied to a conventional document centralization system.

However, although this method can be helpful to some extent for statistical significance or reference utilization of documents, it is basically not based on the pattern analysis of individual users, so it cannot be optimized for the use and utilization of documents based on each user. There are limits.

In addition, in the case of a system implementing document centralization, since a vast amount of documents or document objects are stored and used, there is a problem in that it is difficult to quickly search and check a desired document or a document to be used as a reference.

Although a keyword-based search tool is implemented, this is a search method that is uniformly applied to the OS, etc. The search environment is implemented based only on the file name or some words in the document content, so the document the user wants is quickly and accurately provided. It can still be difficult to do.

In addition, in order to increase the efficiency of search or search, when using a document (creation, revision, editing, etc.), a method of assigning various tag information to the document is induced by policy, but this is basically dependent on the document creator. Since it is a human-dependent method, human error can occur at any time, and since it is additional information based on subjective and arbitrary judgment, accuracy and reliability are quite low.

In the document centralization system, since a large number of document objects by multiple users are made into a DB, it is more important to implement a system that accurately recommends or provides documents that have substantial similarities to documents used by users in order to further increase the synergy effect on document creation, reference, and utilization. can be said to be necessary.

The present invention was conceived to solve the above-mentioned problems in the background as described above, and analyzes and tracks various patterns of users that change dynamically in a time series in various ways, and based on this, recommends a list of documents necessary for the user, thereby providing a vast Its purpose is to provide a customized document recommendation system using dynamic changes in time-series pattern information that can more extensively apply convenience functions for personalized use while maintaining the essential advantages of a document centralization system in which a database of capacity is built. there is

Other objects and advantages of the present invention can be understood by the following description, and will be more clearly understood by the embodiments of the present invention. In addition, the objects and advantages of the present invention can be realized by the configuration shown in the claims and the combination of the configuration.

To achieve the above object, a customized document recommendation system using dynamic change of time-series pattern information of the present invention includes a document object DB unit in which document objects, which are documents collected by document centralization, are stored; When the document object is used by a user, a history information generator for storing history information including user information using the document object and usage time information when the document object is used in association with the document object; a request information input unit into which a request signal including requester information and at least one request period information among a plurality of different set periods is input; a first sorting unit which selects a used document object, which is a document object used by a user corresponding to the requestor information, from among the document objects by using the history information; a recommendation document sorting unit which selects one or more recommended documents, which are document objects having use time information relatively close to the reference point in time of the request period information, among the used document objects; and a list providing unit generating recommendation list information corresponding to the recommendation document and providing the generated recommendation list information to the requestor.

Here, the recommended document selection unit of the present invention assigns first rank information to each used document object, and the smaller the distance between the use time information of the used document object and the reference time point of the request period information, the higher the rank information. a periodic weight processing unit that gives first-rank information; and a sorting processor which selects the recommended document based on the size of the first rank information among the used document objects.

Preferably, the recommended document sorting unit of the present invention assigns second rank information to each used document object, and further includes a continuous weighting processing unit that assigns relatively high second rank information as the number of consecutively used periods is longer or the number of times of use is greater. In this case, the selection processing unit of the present invention may be configured to select the recommended document based on the computed sizes of the first and second rank information among the used document objects.

According to an embodiment, the recommended document selection unit of the present invention extracts the last use time information from among the use time information included in the history information of the used document object, and the closer the last use time information is to the current time point, the higher the priority. It may further include a proximity weight processing unit that grants 3 rank information. In this case, the selection processing unit of the present invention selects the recommended document based on the calculated sizes of the first to third rank information among the document objects in use. can be configured to

In addition, the recommended document selection unit of the present invention performs morpheme parsing processing on the file name information of the used document object to select a plurality of texts corresponding to nouns, and selects a plurality of texts corresponding to nouns, and among the selected plurality of texts, based on the frequency of use, the top m a text selection unit for selecting main texts (m is a natural number equal to or greater than 2); and a frequency weighting processor that assigns fourth rank information to each used document object, and assigns a relatively higher fourth rank information as the number of main texts included in the file name of the corresponding used document object among the top m main texts increases. can include more.

When the above embodiment is implemented, the screening processing unit of the present invention may be configured to select the recommended document based on the calculated size of the first to fourth rank information among the used document objects.

Furthermore, the recommended document sorting unit of the present invention tracks tracking information that generates tracking information including at least one of modification count information in which the content of the used document object has been modified or information on the revision rate of the modified portion of the entire content of the used document object. In this case, the selection processing unit of the present invention may be configured to select the recommended document based on the size of the first rank information and the calculated size of the tracking information among the used document objects.

In order to implement a preferred embodiment, the recommendation document selection unit of the present invention assigns an attention index to each document object, and the number of users other than the user corresponding to the requestor information who has modified the contents of the document object It may further include an attention level processing unit that gives a relatively high attention level index as the number increases.

When the above embodiment is implemented, the selection processing unit of the present invention may be configured to select the recommended document based on the size of the first rank information and the calculated size of the attention index among the used document objects.

In order to achieve the object according to another aspect, a user-oriented document classification system using a dual application of the classification system of the present invention is a plurality of representing the corresponding category for each of N (N is a natural number equal to or greater than 2) variably set. a first input unit into which sample documents, which are individual documents, are input; a vector generator for generating a multi-dimensional number-based embedding vector representing the contents of a corresponding sample document for each of the sample documents by using an artificial intelligence-based embedding model; a representative vector setting unit that performs statistical calculations on embedding vectors of sample documents belonging to the same category and sets the resulting value as a representative vector of the corresponding category; a target vector generator for generating target vectors, which are embedding vectors of each of the m target documents, targeting m (m is a natural number greater than or equal to 2) target documents to be centralized; and a data processing unit that selects at least one belonging category to which each of the m target documents belongs, using a similarity between the representative vector of each of the N categories and the object vector.

Here, the data processing unit of the present invention selects P (P is a natural number of 1 or more and N or less) categories having a degree of similarity greater than or equal to the first reference value through a contrast operation between the representative vector of each of the N categories and the target vector. It may be configured to perform classification processing in which each of the m target documents is selected as a category belonging to each of the target documents, and ranking information according to the degree of similarity is assigned to each of the m target documents for each of the P categories.

Preferably, the present invention applies a clustering algorithm to e (e is a natural number of 1 or more and m or less) target documents in which no category having a similarity higher than the first reference value among the m target documents exists, and k (k is a clustering unit for generating groups (at least 1 natural number); a second representative vector setting unit which performs a statistical operation on target vectors of target documents belonging to each of the k groups and sets the resultant value as a representative vector of each of the k groups; and S (S is a natural number between 1 and k) groups having a degree of similarity greater than or equal to the second reference value through a contrast operation between the representative vector of each of the k groups and the target vector of each of the e target documents. It may be configured to further include a second data processing unit that selects each document as a group to which it belongs, and assigns ranking information according to the degree of similarity to each of the e target documents for each of the S categories.

Furthermore, the present invention provides a text extraction unit for extracting text corresponding to a noun by applying a morpheme parsing algorithm to file name information of target documents belonging to each of the k groups; and a naming unit that selects a category naming for each of the k groups by using a central text, which is a text having a relatively high frequency among the texts.

Specifically, the naming unit of the present invention includes a selection unit for selecting a plurality of central texts based on frequency of use; a calculation processing unit for calculating information on the total number of times each of the plurality of central texts is used in a target document of a corresponding group and information on the number of repetitions in which each sentence of the same target document is repeatedly used; and a naming selector configured to select a category naming for each of the k groups by using the total number of times information and the overlapping number information.

Here, the naming selection unit of the present invention may be configured to select one of the plurality of central texts as the category naming of the corresponding group using the size of the resulting value of the formula below.

In the above formula, A is total number information, B is repetition count information, a is a real number greater than or equal to 1, and b is a real number greater than or equal to 1 and less than or equal to a.

More preferably, the present invention includes a monitoring unit for monitoring whether the content of the target document is modified by a user; When the content of the target document is modified by the user, the target vector generator is controlled to generate a second target vector for the modified target document, and the classification process is performed using the second target vector, and the modification is performed. a processing control unit controlling the data processing unit to generate time-series ranking information of the modified target document for each time period; and a feedback processor configured to transmit related information to a corresponding user when a change in the time series ranking information is greater than or equal to a reference value.

In order to achieve the object according to another aspect, the user profile-based similar document recommendation system of the present invention stores document objects, which are documents collected by document centralization, having a hierarchical tree structure of upper and lower S (S is a natural number of 2 or more). Document DB unit stored in the space; an embedding vector generator for generating a multi-dimensional number-based embedding vector representing the content of each document object for each of the document objects by using an artificial intelligence-based embedding model; a request input unit including information about a requester and information about a target document and inputting a request signal for recommending a similar document; a similar document selection unit which selects k (k is a natural number of 2 or more) similar documents having similarities to the target document through first similarity processing between the embedding vector of the target document and the embedding vector of the document object; and a main processing unit providing similarity list information including information on the k similar documents to the requester.

The present invention may further include a first ranking information generating unit generating similarity ranking information for each of the k similar documents. In this case, the main processing unit of the present invention provides the similarity list information according to the similarity ranking information. can be configured to

Preferably, the present invention provides profile information including storage path information of the document object stored in the storage space, use time information of the corresponding document object, user information using the corresponding document object, and belonging information of the corresponding user to the corresponding document. A profile creation unit that stores in association with an object; a learning object creation unit that creates k learning objects consisting of storage path information of the k similar documents, first time information at which the request signal was input, information about the requestor, and elements about membership information of the requestor; and a learning processing unit that applies a learning algorithm using the profile information to the k learning objects to generate weight information for each of the k similar documents representing a probability that profile information corresponding to the corresponding learning object exists among the profile information. may further include.

In this case, the main processing unit of the present invention may be configured to generate final ranking information by reflecting the weight information to the similarity ranking information and provide the similarity list information according to the final ranking information.

Furthermore, the learning object creation unit of the present invention processes encoding for each element constituting the storage path information of the k similar documents, the first time information at which the request signal was input, the requestor information, and the requester's belonging information. an individual vector generator for generating an individual vector of a number system representing a corresponding element by performing; and an object generator for generating a multidimensional final vector obtained by combining the individual vectors and generating the k learning objects using the generated multidimensional final vector.

In this case, the individual vector generating unit of the present invention generates independent individual vectors for each of the storage path information of the k similar documents, the first time information at which the request signal was input, the information about the requester, and the belonging information of the requester; , It is preferable to generate an independent individual vector for each element of S hierarchies forming a hierarchical tree structure as an element constituting the storage path information of the k similar documents.

Furthermore, the running processing unit of the present invention includes the storage path information of the k similar documents, the first time information at which the request signal was input, the number of corresponding elements among elements related to information about the requestor and information about the belonging of the requester, and It may be configured to generate the weight information by applying similarities of corresponding elements in combination.

According to an embodiment of the present invention, the time required for document search and the like can be drastically reduced by effectively combining analysis of the time-series dynamic change of individual pattern information of each user using document centralization and statistical utilization.

In addition, according to an embodiment of the present invention, it is possible to induce a user to effectively check an important document at each predetermined cycle set variably, thereby effectively suppressing problems such as overlooking or missing an important document. Document access and user convenience can be further optimized.

According to another embodiment of the present invention, by complementarily grafting category classification using an identification system such as an embedding vector and category classification using a clustering technique, etc., it is possible to avoid compulsory classification based on similarity proximity, thereby substantially improving document objects. It is possible to implement category classification that reflects the content more precisely.

According to another embodiment of the present invention, information retrieval and information accessibility can be drastically improved by allowing the naming of categories to be selected based on the actual contents of individual documents and the results of statistical calculations.

In addition, in the case of the present invention, a general rule of thumb (UX) for creating a document or naming the title or file of the document can be reflected, so that the category or category/group to which the document belongs can be more intuitively recognized. and confirmation, the use environment of centralized documents can be implemented as a more user-oriented environment.

According to another embodiment of the present invention, an embedding vector capable of representing the actual content of a document object and processing for determining similarity are applied to the embedding vector, so similar documents having similarity based on the target document and its content It can be selected and provided accurately and objectively.

In addition, according to another embodiment of the present invention, the storage path or access path of the document object, as well as the time information (month, day, time, etc.) when the document object was used, and user information and affiliation representing human relevance Since it is configured to create a learning object in the form of a multi-dimensional vector that comprehensively reflects information, etc., and provide a quantitative index for similarity based on this, it is possible to recommend and provide similar documents that can further maximize the reference usability between documents. there is.

The following drawings attached to this specification illustrate preferred embodiments of the present invention, and together with the detailed description of the present invention serve to more effectively understand the technical idea of the present invention, the present invention is described in these drawings should not be construed as limited to
1 is a diagram showing a document classification system and overall configuration related thereto according to a preferred embodiment of the present invention;
2 and 3 are block diagrams showing the detailed configuration of a document classification system according to a preferred embodiment of the present invention;
4 is a flowchart showing a category classification process using an embedding vector;
5 is a flowchart showing clustering processing and naming selection processing;
6 is a flowchart showing a category classification process performed for new documents;
7 is a block diagram showing the detailed configuration of a document recommendation system according to a second embodiment of the present invention;
8 is a block diagram showing a detailed configuration of the recommended document selection unit of FIG. 7;
9 is a flowchart showing a document recommendation processing process according to a second embodiment of the present invention;
10 is a flowchart showing a specific processing process for generating a recommendation document;
11 is a block diagram showing the detailed configuration of a similar document recommendation system according to a third embodiment of the present invention;
12 is a flowchart showing similar document recommendation processing according to a third embodiment of the present invention;
13 is a flowchart showing a processing procedure for recommending similar documents by comprehensively considering weight information, etc.;
14 is a diagram showing a document storage structure according to an embodiment of the present invention having a hierarchical tree structure up and down;
15 is a diagram showing an example of individual vectors generated for each element of a learning object.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Prior to this, the terms or words used in this specification and claims should not be construed as being limited to the usual or dictionary meaning, and the inventor appropriately uses the concept of the term in order to explain his/her invention in the best way. It should be interpreted as a meaning and concept consistent with the technical idea of the present invention based on the principle that it can be defined.

Therefore, since the embodiments described in this specification and the configurations shown in the drawings are only one of the most preferred embodiments of the present invention and do not represent all of the technical ideas of the present invention, various equivalents that can replace them at the time of the present application It should be understood that there may be waters and variations.

1 is a diagram showing a document centralization system (100, 500, 1000) according to an embodiment of the present invention and its overall configuration.

As described above, the document centralization system 100 of the present invention is a system implementing cloud-based document centralization, and is capable of communicating with one or more user terminals 200-1 and 200-n through a communication network 50. Connected.

When a user (employee in the company, etc.) performs tasks such as creating, modifying, and editing through the user terminals 200-1 and 200-n, the result (hereinafter referred to as 'document' or 'document object') is It is stored and managed in the document centralization system (100, 500, 1000) of the invention.

It can be set in various ways according to company policies or supplement-related protocols, but basically, the user terminal 200 may not be loaded with software for business processing, and in addition to temporary storage (volatile storage), the resulting data by the software, Since it may not be stored, the user terminal 200 may be configured to perform the function of an interfacing means in relation to the document centralization system 100, 500, 1000 of the present invention.

The document centralization system (100, 500, 1000) of the present invention can be communicatively connected to the user terminal (200) through a limited non-open network such as an in-house intranet, and according to the embodiment, authentication and security are reinforced web It may be communicatively connected to the user terminal 200 through a method such as via the server 300 .

If a method of authenticating a user's qualification with an enhanced technique such as exchange and matching of a unique key, authentication by OTP (One Time Password), etc. is mediated, the document centralization system (100, 500, 1000) of the present invention is a user's mobile terminal (400) or PC-based terminals can also be directly connected to enable communication.

The document centralization system (100, 500, 1000) of the present invention provides a space in which documents written or edited by individual users are stored, and among various data stored in the server space, data, information, documents, etc. that meet the user's request are stored. It may be configured to provide users with various additional functions such as searching and providing, as well as automatically recommending similar documents.

The document centralization system (100, 500, 1000) of the present invention performs parsing processing on a newly created document and stores the document in a logical space such as a directory that can be formed in a hierarchical or tree structure. is configured to

In particular, the document centralization system (100, 500, 1000) of the present invention classifies and processes a plurality of documents (hereinafter referred to as 'target documents') that are subject to document centralization to enable various utilization of documents stored in the server. It is configured to categorize into a plurality of categories by performing.

The document centralization system (100, 500, 1000) of the present invention uses log information for each user (user) to identify information about documents or document objects used (read, write, read, modify, edit, etc.), corresponding It is configured to continuously update and create history information including user information using a document object or time (date, time, etc.) information using the document.

In addition, the document centralization system 100, 500, 1000 of the present invention is configured to generate and provide customized recommendation list information specialized for each user by processing and calculating the generated history information based on various parameters and setting environments.

In addition, the document centralization system (100, 500, 1000) of the present invention selects one or more similar documents that are similar to the target document based on the target document, which is a document currently used by the user or a specific document by the user, and provides a GUI environment It may be configured to provide selected similar documents or list information about them through such interfacing.

Hereinafter, each embodiment of unique processing and configuration of the present invention will be described in detail with reference to the accompanying drawings.

On the other hand, in the following description, the document centralization system (100, 500, 1000) of the present invention is referred to by different terms or names based on the embodiment, but this effectively reflects the main function and technical idea of each embodiment and emphasizes or It is just a different rhetorical expression to highlight, but it means substantially the same document centralization system (100, 500, 1000).

As shown in FIG. 2, a user-oriented document classification system (hereinafter referred to as 'document classification system') 100 using a dual application of a classification system according to a first embodiment implementing the document centralization system of the present invention. may include a first input unit 110, a vector generator 120, a representative vector setting unit 130, a target vector generator 140, and a data processing unit 145.

Prior to the detailed description of the present invention, the document centralization system (100, 500, 1000) of the present invention can be implemented in the form of a server that can be accessed online. Each component constituting the document centralization system (100, 500, 1000) or detailed configuration according to the present invention should be understood as a logically separated component rather than a physically separated component.

That is, since each configuration corresponds to a logical component for realizing the technical idea of the present invention, even if each component is integrated or separated, if the function performed by the logical configuration of the present invention can be realized, the present invention It should be interpreted that it is within the scope, and any component performing the same or similar function should be construed as within the scope of the present invention regardless of whether or not the name is identical.

In the category setting unit 105 of the present invention, information on N categories into which documents are to be classified is set (S400, see FIG. 4), and information of the manager who additionally inputs the corresponding information, information on the time when the corresponding information was input, etc. are stored. It can be.

As will be described later, since N categories correspond to category information (e.g., planning, finance, research, personnel, etc.) that is the criterion for classifying target documents that are subject to document centralization, it is a criterion for understanding the overall business of the company. It is preferable to be configured to be set by a person in charge of a level or higher or a plurality of administrators, etc., and configured to be configured to be set variably by reflecting various situations, but it is preferable to be configured so that information on the N categories is not changed without a procedure such as authentication. . In terms of setting a plurality of categories, N is a natural number of 2 or more.

In the first input unit 110 of the present invention, sample documents, which are a plurality of documents representing the corresponding category, are input for each of the N categories (S410).

The sample document is a document input by a system administrator, etc., and means a document clearly belonging to a corresponding category, and corresponds to a document used as raw data that is a standard for artificial intelligence-based learning or modeling.

In order to improve the precision of modeling, it is preferable to have a plurality of sample documents for each category, and for efficiency of calculation processing, the same number of sample documents for each category may be input, but is not limited thereto. Of course.

In this way, when a plurality of sample documents are input for each category, the vector generator 120 of the present invention uses an artificial intelligence-based embedding model to represent the content of the corresponding sample document for each of the sample documents, a multidimensional number-based embedding vector. (Embedding vector) is created (S420).

The system, expression method, and rules constituting the vector are slightly different for each model that generates the vector, but basically, an embedding vector (vector embedding) is a variable vector used in machine learning and deep learning. It is expressed as an N-dimensional number, and since the numbers of each dimension are configured to have the characteristics of the corresponding word, words with similar meanings have characteristics expressed by similar numerical values.

The integer encoding is one of the techniques for converting natural language into a number form that can be understood by a machine, and means processing that maps each word to a unique integer. Starspace, fastText, word2vec, GloVe, ELMo, BERT, etc. This corresponds to a representative example of an embedding model.

If there are 8 categories and 7 sample documents are input for each category, the vector generator 120 of the present invention generates 56 embedding vectors (S420).

In this way, when an embedding vector is generated for each sample document, the representative vector setting unit 130 of the present invention performs statistical calculations on the embedding vectors of the sample documents (7 in the example above) belonging to the same category, and calculates the resulting value. It is set as a representative vector of the corresponding category (S430).

Statistical calculation may be average value calculation, and according to the embodiment, of course, it may be weight calculation reflecting distance matrix of embedding vector generated from each sample document, distribution map related to variance/concentration, and the like.

The target vector generator 140 of the present invention targets m (m is a natural number of 2 or more) target documents stored in the document DB unit 103, that is, m documents or document objects to be centralized. An embedding vector for each of the target documents is generated (S440). In the following description, in order to distinguish the sample document from the embedding vector generated for the target document, the embedding vector generated for the target document (document subject to classification processing) is referred to as 'object vector'.

When the object vectors of each of the m target documents are generated, the data processing unit 145 of the present invention performs similarity processing with each of the N categories based on the object vectors of the m target documents (S450). Using the result, one or more belonging categories to which each of the m target documents belongs are selected (S480).

Specifically, the data processing unit 145 of the present invention selects P categories having a degree of similarity equal to or higher than the first reference value (S460) through a contrast operation (S450) between the representative vector of each of the N categories and a specific target vector as a specific target document. You can select by category.

Since a category related to a specific target document is set, the P categories can logically be 1 or more and N or less, and the first reference value, which is a criterion for determining whether or not belonging to a specific category, reflects various elements and parameters. Of course, it can be variably adjusted by doing so.

In this case, the data processing unit 145 of the present invention converts ranking information according to which category a specific target document more closely matches, that is, the result value of similarity processing (size of similarity) to attribute information or tag information of a specific target document. It is more preferable to be configured to give (S480).

In this way, one or more categories of the target document are selected/set, but the ranking information is given to the attribute information of the target document, so that the process of user recommendation, search, etc. can be performed more efficiently in the future, and furthermore, the user's direct confirmation process And through the reflection of opinions, etc., it can be the basis for more precise category setting.

As described above, through a comparison operation between the representative vector and the target vector of each of the N categories, P categories having similarities greater than or equal to the first reference value are selected as belonging categories of each of the m target documents, and the ranking information is converted into m target documents The processing given to is generically referred to as 'classification processing' below.

Hereinafter, the clustering processing (S470) of the present invention, which is performed when there is no category having a similarity higher than the first reference value (S460), that is, when there is a target document that has not been classified into a category through classification processing, is shown in FIG. 5 and the like. Please refer to it for a detailed explanation.

Even if the above-described classification processing is performed, some of the m target documents may not belong to any one of the N categories. In this case, if the type and number of categories set in step S400 are increased, useless processing may be repeated, so the present invention performs clustering processing only for the remaining target documents excluded from classification processing.

Specifically, the clustering unit 150 of the present invention selects e target documents in which no category having a similarity higher than the first reference value exists among m target documents (S500).

For example, if one or more categories corresponding to 700 target documents are set through classification processing among a total of 1,000 (m) target documents, 300 (e = 1,000-700) target documents are subject to clustering processing. In this respect, logically, e corresponds to a natural number of 1 or more and m or less.

The clustering unit 150 of the present invention applies a clustering algorithm to the e target documents to generate k (k is a natural number greater than or equal to 1) groups (S510).

An unsupervised learning technique such as K-means is a representative example of the clustering algorithm, and hierarchical clustering, maximum distance algorithm, Isodata algorithm, Robin-Monroe method, and the like can also be applied as the clustering algorithm.

When k groups are generated, the second representative vector setting unit 160 of the present invention performs a statistical operation on target vectors (embedding vectors) of target documents belonging to each of the k groups, and calculates the resultant value in the k groups. It is set as a representative vector for each group (S520). As described above in the representative vector setting unit 130, an average value calculation may be applied to the statistical calculation.

In this way, when the representative vectors of each of the k groups are set, the second data processing unit 165 of the present invention performs similarity comparison processing between the k representative vectors and the object vectors of each of the e target documents (S530).

Using the result of the similarity comparison processing, the second data processing unit 165 of the present invention assigns S (S is a natural number between 1 and k) groups having a similarity higher than the second reference value to each of the e target documents. , but ranking information according to the degree of similarity for each of the S categories is given to each of the e target documents (S550).

Since a specific target document belonging to a specific group among k groups belongs to a specific group by a clustering algorithm, a representative vector representing its group in the similarity comparison processing (representative vector of the corresponding group vs. similarity between its target vector). may have the highest degree of similarity.

In this regard, since the present invention performs similarity comparison processing between the representative vectors of each of the k groups and the subject vector of a specific target document, it is possible to additionally identify groups highly related to the present invention, thereby further enhancing the precision of category classification. .

In addition, since the relationship with a plurality of groups (categories) is confirmed through numerical information through ranking information, it is possible to provide a basis for various utilization of subsequent application processing as described above.

On the other hand, in the case of the clustering technique, it is possible to classify a plurality of objects into a plurality of groups in which properties or characteristics correspond to each other, but unless objects belonging to the group are directly identified, the characteristics of the group It is impossible for users to recognize intuitively and with high visibility.

The present invention may further include a text extraction unit 107 and a naming unit 170 to solve these conventional problems.

The text extraction unit 107 of the present invention applies a morpheme parsing algorithm to file name information of target documents belonging to each of the k groups to extract text corresponding to nouns (S550).

When the text corresponding to the noun is extracted, the naming unit 170 of the present invention selects the category naming of each of the k groups using the center text, which is a relatively high-frequency text among the extracted text (S580).

Specifically, when the selection unit 171, which is one component of the naming unit 170, selects a plurality of center texts based on the frequency of use (S560), the calculation processing unit 173 selects each of the plurality of center texts as a corresponding group. Duplicate count information, which is the number of times each of the plurality of central texts is used repeatedly, is calculated based on the total count information used in the target document and each sentence constituting the same target document (S570).

The naming unit 170 of the present invention selects one of the central texts as a category naming that can represent the characteristics of the corresponding group by applying the total number of times information and the overlapping number information of each central text in combination (S580). Depending on the embodiment, a central text such as a secondary ranking may be selected as preliminary category naming.

In this way, the present invention primarily reflects the subjective purpose or intention intended by the user by selecting the central text based on the file name set by the user himself, and furthermore, the frequency with which these central texts are actually used in the target document In addition, since the category naming can be selected by comprehensively considering status information repeatedly used in the same sentence of the target document through morpheme analysis and data analysis, the substance of the document content can also be reflected.

In this case, the naming selector 175 of the present invention may be configured to select one of the plurality of central texts as the category naming of the corresponding group by using the size of the resulting value of Equation 1 below.

In the above formula, A is information on the total number of times each of the plurality of central texts is used in the target document of the corresponding group, and B is information on the total number of times each of the plurality of central texts is used repeatedly based on each sentence of the same target document belonging to the corresponding group. It is information on the number of duplicates, constant a is a real number of 1 or more, and constant b is a real number of 1 or more and less than or equal to a.

In this way, by reflecting the total count information as a numerical value for algebraic processing and reflecting the duplicate count information as a numerical value for exponential processing, the total count information and the duplicate count information are comprehensively considered. Since a more weighted result can be derived, the importance of the corresponding central text in the actual target document can be more practically reflected.

In addition, by setting a constant a weightedly reflecting the total number of times information (A) and a constant b weightedly reflecting the overlapping number information (B) to be correlated, it is possible to suppress excessive reflection of one of the two parameters compared to the other. can be reflected more harmoniously.

According to the embodiment, the document classification system 100 of the present invention may further include a monitoring unit 180, a processing control unit 190, and a feedback processing unit 195 as shown in FIG.

The monitoring unit 180 of the present invention may be configured to monitor whether the contents of the target document are modified based on nouns using a morpheme analysis engine or the like.

In this way, when the content of a specific target document is modified, the processing control unit 190 of the present invention controls the target vector generator 140 to generate a second target vector for the modified target document. The second target vector refers to an embedding vector generated for a modified target document.

When the second object vector is generated in this way, the processing control unit 190 of the present invention controls the data processing unit 145 to perform the above-described 'classification processing' using the second object vector, and furthermore, the target document is modified. Control to generate time-series ranking information of the modified target document by time.

In this way, if time series ranking information is given as attribute information to a specific target document, the user or manager can directly or indirectly check the change in the content of the specific target document, and based on this, the effectiveness of other subsequent response processing can be increased.

In addition, the feedback processing unit 195 of the present invention controls the sharing of related information to the first document creator, document modifier, or manager when the change in the time series ranking information occurs above the reference value, so that misuse or incorrect modification of the document is prevented. Centralized documents can be managed more effectively by continuously monitoring whether or not they occur.

6 is a flowchart illustrating a category classification process performed for new documents.

When a new document is input through the first input unit 110, the object vector generator 140 of the present invention generates an embedding vector for the input new document (S600). Of course, depending on the embodiment, the embedding vector corresponding to the new document may be input together with the new document through a preprocessing process or the like.

When the embedding vector of the new document is confirmed or generated, the data processing unit 145 of the present invention accesses and reads the representative vector of each of the N categories used for document centralization (S610), and as described above, Based on the embedding vector of the document, similarity comparison processing is performed between the embedding vector of the new document and the representative vector of each of the N categories (S620).

If there is one or more categories having a similarity higher than the third reference value, the data processing unit 145 of the present invention assigns ranking information according to the degree of similarity to set the category to which the new document belongs (S640).

The third reference value or the fourth reference value described later may be set to the same value as the first or second reference value described above, but may be set to different values depending on the embodiment in order to reflect the relationship with the category in various ways. is of course

In the similarity comparison processing (S630), when there is no category having a similarity higher than the third reference value, the data processing unit 145 of the present invention loads representative vectors of k groups used for clustering processing (S650) Processing is performed to compare the degree of similarity between the representative vector of each of the k groups and the target vector (embedding vector) of the new document (S660).

As a result of similarity comparison processing, if there is a category (group) having a similarity higher than the fourth reference value (S670), as described above, ranking information according to the degree of similarity is assigned to each category (group) to obtain the new document. An affiliated category (group) is set (S640).

If, as a result of the similarity comparison processing (S670), if there is no category to which the new document can belong, this means that the new document is a document with content that cannot be classified into the existing N+k categories, so that the new document In order to process as , preliminary processing such as storing in a separate storage space is performed (S680).

When the number of new documents stored in the separate storage space corresponds to the predetermined reference number, the clustering processing described above may be automatically performed so that grouping and naming of only new documents may be performed.

In addition, according to the embodiment, when a new document is stored in a separate storage space so that direct processing by an administrator can be induced, a corresponding fact may be informed to an administrator terminal or the like.

Of course, the above-described processing of the present invention may be configured to be applied cyclically if a preset termination condition is not satisfied (S690).

Of course, a method of loading all N+k representative vectors in the classification process of a new document and performing similarity comparison processing with a target vector (embedding vector) corresponding to the new document based on this is also applicable.

However, since it is common for N categories to be set based on the overall content of in-house documents, etc., classification processing is performed first based on N categories, and in this process, if a target document (new document) that is not classified is generated, For one year, it may be preferable in terms of efficiency of calculation processing to be designed to perform processing using representative vectors of k groups.

Hereinafter, a customized document recommendation system (hereinafter referred to as 'document recommendation system') using dynamic change of time-series pattern information according to the second embodiment implementing the document centralization system of the present invention with reference to the accompanying drawings (500 ) will be described in detail with respect to the specific configuration and processing.

7 is a block diagram showing the detailed configuration of a document recommendation system 500 according to the second embodiment of the present invention, and FIG. 9 is a flowchart showing a document recommendation processing process according to the second embodiment of the present invention.

As shown in FIG. 7, the document recommendation system 500 according to the second embodiment of the present invention includes a document object DB unit 510, a history information generator 520, a request information input unit 530, and a first selection unit. 540, a recommendation document selection unit 550, and a list providing unit 560 may be included.

The document object DB unit 510 of the second embodiment has a configuration corresponding to the document DB unit 103 described in the above-described embodiment, and stores a plurality of document objects, which are documents collected by document centralization (S3000, FIG. 9). reference).

When the document object is used by the user (S3100), the history information generation unit 520 of the present invention is information including user information using the document object and usage time information (time, date, etc.) when the document object was used. The history information is stored in association with the corresponding document object (S3200).

Although it may vary according to various factors such as security level, handling authorization standard, or operation policy, if the number of document objects stored in the document object DB unit 510 is 100,000, the history information is logically based on each of the 100,000 document objects. can be created

Since the user accesses the server where the document recommendation system 500 is implemented through log-in using an ID or security or authentication method, in this process, the user's identification information (ID information, etc.) or a unique number (IP) of a specific user terminal , MAC address, etc.) and the like can be confirmed. The history information generation unit 520 of the present invention generates the history information using such information.

Of course, the history information can be continuously generated, updated, and stored through cyclical processing of an algorithm implemented inside the system, time stamp, and the like.

Therefore, the history information is time-series information for each user using a document object, and can function as raw data for analyzing a pattern of each user that dynamically changes over time.

The request information input unit 530 of the present invention is a component that receives a request signal from a user (S3300), and the request signal is one or more of a plurality of different set periods such as information of a requester requesting information and each day, week, month, etc. It includes request period information, which is information.

The requestor information refers to information of a requestor who wants to receive periodic reports as a member (employee) who may be a member of an organization, company, etc. in which the document recommendation system 500 is operated. It goes without saying that this requestor information can be confirmed through various methods including procedures such as log-in or authentication as described above.

In this way, when a request signal is input, the first sorting unit 540 of the present invention uses the request signal and the history information stored in the history information generation unit 520 in three dimensions to use the user (requester) corresponding to the requestor information. ) Selects a used document object, which is a used document object (S3400).

Here, 'use' of a document object refers to various work activities related to the document object, such as read, write, make, amend, and edit.

In this way, when a used document object, which is a document object classified by the requester's personal criteria, is selected among document objects, the recommendation document selection unit 550 of the present invention selects a recommended document that meets a specific requirement among the used document objects (S3500 ).

As an example of this, the recommended document selection unit 550 of the present invention selects the recommended document using one or more document objects having use time information relatively close to the reference point of the request period information among the used document objects. can

When a recommended document is selected in this way, the list providing unit 560 of the present invention generates recommendation list information, which is list information corresponding to the recommended document itself or the recommended document, and generates it to the user terminal corresponding to the requester information or to the requester. It is provided by designated e-mail, mobile terminal, etc. (S3600).

Of course, the above-described processing according to the second embodiment of the present invention can be applied repeatedly or cyclically (S3700) unless a specific end condition is satisfied.

FIG. 8 is a block diagram showing the detailed configuration of the recommended document selection unit 550 of FIG. 7 , and FIG. 10 is a flowchart illustrating a specific processing process for generating a recommended document in the recommendation document selection unit 550 .

As shown in FIG. 8, the recommended document selection unit 550 of the present invention includes a period weight processing unit 551, a continuous weight processing unit 552, a proximity weight processing unit 553, a text selection unit 554, a frequency weight processing unit ( 555), a tracking unit 556, a attention level processing unit 557, and a selection processing unit 558.

As described above, when a request signal including request period information, which may consist of information such as day, week, month, quarter, year, etc., is input, the first screening unit 540 of the present invention checks the requested period information. (S4000) and inquires and checks the history information during the inquiry period set differently according to the requested period information (S4100).

As an example related to the implementation of the embodiment, when a requester applies to be provided with recommendation list information based on a day of the week (Monday), the corresponding usage time information ( Monday) search history information. If the requested period information is weekly, 30 days based on 4 weeks or 90 days, which is an integer multiple thereof, may be the inquiry period.

Depending on the embodiment, of course, the request period information can be set to a plurality, and in this case, the document recommendation system 500 of the present invention provides one of day-of-week recommendation list information, weekly recommendation list information, and monthly recommendation list information to a specific user. provides more than

In this way, when the history information for a specific inquiry period is inquired and confirmed, the period weight processing unit 551 assigns the first rank information to each document object used, and uses the history information to determine the document object used (the document object used by the requester) ) and the reference time of the requested period information (day, week, month, quarter, etc.), the relatively higher first rank information is assigned (S4200).

If the request period information is set for each day of the week and the base date is Thursday, the usage time information corresponding to Thursday among the used document objects during the inquiry period (for example, 21 days, etc.) before the base date is included as history information in the used document object. The highest first rank information is given, and the second rank first rank information is given to the used document objects whose usage time information is Wednesday and Friday.

From a corresponding point of view, the first rank information, which is relatively the lowest, is assigned to a used document object in which Monday and Sunday are used time information.

As an example of the implementation of the algorithm, each day of the week is given a number system (e.g., Monday: 1, Tuesday: 2, Wednesday: 3, Thursday: 4, Friday: 5, Saturday: 6, Sunday: 7) and based on An operation is performed on the difference between the number system corresponding to the day of the week (Thursday) and the absolute value of the day of the usage time information, and if the result value exceeds 3, a method of calculating the result value by subtracting the result value from 7 can be applied. there is.

In this way, when the numerical result value for each day of the week is derived, the lower the result value (separation degree (difference between the reference day and the corresponding day)), the higher score should be given, and the greater the difference (separation degree), the lower score should be given, so the hyperbolic tangent The first rank information may be calculated using a first function that reflects an inverse proportional relationship such as a hyperbolic tangent function.

It goes without saying that, depending on the embodiment, processing such as standardizing the first function so that the range (first rank information) of the first function is in a specific range (eg, 0 to 1.5) may be applied.

As such, if the first rank information is calculated for each used document object, the selection processing unit 558 of the present invention selects one or more recommended documents based on the size of the first rank information among the used document objects (in descending order) (S4800). When the recommended document is selected, as described above, the list providing unit 560 provides the requester with recommendation list information for the selected recommended document or list of recommended documents (S3600).

If the request period information is set by week, similarly, 1 to 4 (1st week: 1, 2nd week: 2, 3rd week: 3, 4th week (5th week): The number system of 4) is assigned, and the recommendation document is selected using information on parking at the current time and parking time information and a function using it as an independent variable.

As a document object with periodicity, documents are recommended based on the document object used by the requester on the corresponding day, week, month, etc., so the requester (user) can significantly reduce the search time for documents with high relevance based on the corresponding day/week/month. It can be shortened to , and it can be induced to accurately check all the document objects (documents) to be checked/created in the corresponding cycle.

The continuous weight processing unit 552 of the present invention assigns second rank information to each used document object, but gives relatively higher second rank information as the period (time) of continuous use is longer or the number of times of use is greater (S4300). .

In the case of the second rank information, the higher the independent variable (period of continuous use, number of times of use) is, the higher the score should be. Therefore, the second function that implements the second rank information as an algorithm is a log-based function (function based on It is preferable to apply a function having a characteristic of representing a proportional relationship and converging at the same time, such as log).

In this case, the selection processing unit 558 of the present invention selects the recommended document based on at least one of the first rank information and the second rank information, and preferably, to reflect the user pattern more three-dimensionally, the first rank information and the second rank information The recommendation document may be selected based on the sum of rank information.

The proximity weight processing unit 553 of the present invention extracts the last use time information among the use time information included in the history information of the used document object, and gives a relatively high third rank information as the last use time information approaches the current time. It may be configured to (S4400).

In the case of the third rank information, the higher the independent variable (interval from the last use time information) is, the lower the score should be, and it is desirable to assign a relatively high weight to the independent variable in the vicinity area. function can be applied.

In the above formula, y is the third rank information, and x (a real number greater than or equal to 0) is the difference between the last use time information and the current time.

In this case, the selection processing unit 558 of the present invention selects the recommended document based on one or more of the first rank information to the third rank information, and preferably, the first rank information to the third rank information in order to more comprehensively reflect the user pattern. The recommendation document may be selected based on a summed result of rank information or an operation value corresponding thereto.

The text selection unit 554 of the present invention selects a plurality of texts corresponding to nouns by performing morpheme parsing processing on the file name information of the used document object, and selects a plurality of texts corresponding to nouns, and selects the top m based on the frequency of use of the selected plurality of texts. (where m is a natural number of 2 or more) main texts are selected.

In this way, when a plurality of main texts are selected, the frequency weighting processing unit 555 of the present invention assigns fourth rank information to each of the used document objects, and the main text included in the file name of the corresponding used document object among the top m main texts. As the number of is increased, relatively high fourth rank information is assigned (S4500).

In this case, the selection processing unit 558 of the present invention selects the recommended document based on at least one of the first rank information to the fourth rank information, and preferably, the first rank information to the fourth rank information to reflect the user pattern in more diverse ways. The recommendation document may be selected based on a result of summing up a plurality of pieces of rank information, preferably all of them.

To implement a more preferred embodiment, the recommendation document selection unit 550 of the present invention may further include a tracking unit 556. The tracking unit 556 is configured to generate tracking information including one or more of information on the number of times the content of the document object has been modified or information on the modification rate of the entire content of the document object that has been modified (S4600). do.

When the tracking information is generated in this way, the screening processing unit 558 of the present invention can select the recommended document based on the size of the tracking information, and according to the embodiment, one or more of the above-described first rank information to fourth rank information And it may be configured to select the recommended document based on the calculated size of the tracking information.

In the case of a document object that has been modified a lot or has a high modification rate, it is highly likely that it corresponds to a document in progress. Therefore, if configured as above, the relevance of the requester's task or the current progress of the task can be effectively reflected. be able to

In addition, the attention processing unit 557 of the present invention assigns an attention index to each document object, but the higher the number of users other than the requester information who modified the contents of the document object, the higher the degree of attention. It is configured to give indicators.

When the attention index is generated in this way, the selection processing unit 558 of the present invention may select the recommended document based on the size of the attention index, and at least one of the above-described first to fourth rank information or tracking information. And it may be configured to select the recommended document based on the computed size of the attention index.

In this way, in the case of a document object that is frequently used or modified by a person other than himself based on the information requester, it is highly likely that the document is jointly created by multiple people or team members, so that periodic confirmation is induced. The need can be said to be great.

As shown in FIG. 11, the user profile-based similar document recommendation system (hereinafter referred to as 'similar document recommendation system') 1000 according to the third embodiment implementing the document centralization system of the present invention is a document DB unit. 1100, embedding vector generator 1200, request input unit 1300, similar document sorter 1400, main processing unit 1500, first ranking information generator 1600, profile generator 1700, learning It may include an object creation unit 1800 and a running processing unit 1900.

The similar document recommendation system 1000 according to the third embodiment of the present invention is a document currently in use by a user or a document designated by the user (hereinafter referred to as 'target document') and a plurality of documents having similarities and/or It corresponds to a system that improves document utilization more effectively by providing list information to the user through a GUI environment.

In the document DB unit 1100 of the present invention, like the document DB unit 103 of the first embodiment and the document object DB unit 510 of the second embodiment described above, document objects, which are documents collected by document centralization, are stored in a storage space. It is stored (S700, see FIG. 12).

As illustrated in FIG. 14, the document objects are linked to each other and stored in each storage space having a hierarchical tree structure of upper and lower S (S is a natural number of 2 or more). It is stored in association with a specific file in the form of information linking the hierarchical interrelationships of a corresponding series of tree structures.

14 shows a DB structure having a hierarchical tree structure of four (S=4), but this is an example and may have a different number of upper and lower hierarchies.

In addition, as will be described later, in order to more effectively implement the vector function processing of storage path information, the top layer space (directory, archive, etc.) of the tree structure is set to length 1, etc. It can be set to have a length that increases sequentially, and the naming of the storage space for each layer can be stored as digital information that independently classifies and represents it.

As described above, the embedding vector generator 1200 of the present invention applies an artificial intelligence-based embedding model to each document object stored in the document DB unit 1100 to represent the contents of the corresponding document object. A multi-dimensional number-based embedding vector is generated (S710).

Preferably, the embedding vector generation unit 1200 is configured to update and generate the embedding vector of each document object every specific period that can be set variably or when an event in which the content of the stored document object is modified or edited occurs. Of course, it can be configured to automatically generate an embedding vector for a new document object when a document object is created.

The request input unit 1300 of the present invention receives a request signal for similar document recommendation from the user (S720).

The request signal includes personal information of the user (requester) who requested the similar document and ID information for identifying the target document. When a user browses a specific document, user access information and currently viewed document information are used. According to the embodiment, it can be input in various ways, such as displaying an environment for receiving requested information in the form of a pop-up window and receiving a selection for it.

In this way, when a request signal is input, the similar document sorting unit 1400 of the present invention obtains a first value equal to the cosine similarity between the embedding vector of the target document and the embedding vector of the document object stored in the document DB unit 1100. Similarity processing is performed (S730) to select k (k is a natural number equal to or greater than 2) similar documents having content similarity with the target document (S740).

In this case, the first ranking information generator 1600 of the present invention may generate similarity ranking information expressing the degree of similarity between each of the k similar documents and the target document as a quantitative value.

In this way, when k similar documents are selected and similarity ranking information for each similar document is generated, the main processing unit 1500 of the present invention provides similarity list information including information on the k similar documents to the requestor ( S750), depending on the embodiment, list information sorted according to the similarity ranking information may be provided, or a rank or a similarity index according to the similarity ranking information may be provided together.

It goes without saying that the above-described similar document provision processing may be repeatedly and continuously performed if a specific end condition is not met (S760).

Hereinafter, an embodiment of the present invention that more substantially reflects the similarity or degree of similarity of similar documents will be described in detail with reference to FIG. 13 and the like.

The profile creation unit 1700 of the present invention includes storage path information (access information) in which a document object is stored in the storage space (a space having an upper and lower hierarchical tree structure), usage time information when the corresponding document object is used, and corresponding document Profile information including user information using an object and belonging information of the corresponding user is created and stored (S800).

The profile generator 1700 corresponds to some functions of the history information generator 520 of the second embodiment. In addition, as described above, the profile information can be continuously generated, updated, and stored through cyclical processing of an algorithm implemented inside the system and a time stamp.

Therefore, the profile information is time-series information for each user using a document object, and can function as raw data for analyzing patterns of each user that dynamically change over time.

The learning object creation unit 1800 of the present invention creates k learning objects corresponding to each of the k similar documents determined to have similarities with the target document based on the content by using the embedding vector (S810).

The learning object is composed of each element for the storage path information of the similar document, the first time information at which the request signal was input, the personal information of the requestor, and the affiliation information (department, team, etc.) of the requestor, and corresponds to time. It is preferable that the first time information is configured to include month, day, and time information in order to precisely calculate the last name.

Specifically, the learning object generator 1800 may include an individual vector generator 1810 and an object generator 1820. The individual vector generator 1810 performs encoding processing for each element of the storage path information of the k similar documents, the first time information at which the request signal was input, information about the requester, and the belonging information of the requestor. An individual vector of a number system representing the corresponding element is created (S811).

More preferably, as illustrated in FIG. 15, the learning object creation unit 1800 provides independent individual information for each similar document's storage path information, the first time information at which the request signal was input, requester information, and requester's affiliation information. A vector may be generated, but in the case of the storage path information of the k similar documents, an independent individual vector may be generated for each element of the S hierarchical tree structure as an element constituting the corresponding storage path information.

Among the elements of the learning object above, time information and human-related information are not time information, user information, and the user's affiliation information contained in the profile information of similar documents, but time information when the request signal was entered, requester information, and the requester's affiliation. becomes information.

When a learning object is created in this way, the probability that profile information identical/similar to these 9 pieces of information exists based on all 9 (examples) of information (i.e., information converted to a vector system) included in the learning object, or 0 Through a learning model, a probability of existence of an element corresponding to each of the nine elements or a numerical result for the degree of similarity may be generated and used comprehensively.

In this way, when an individual vector is generated for each element, the object generator 1820 of the present invention generates a multidimensional final vector by combining the individual vectors and creates the k learning objects using the generated multidimensional final vector (S813 ).

In the case of the learning object illustrated in FIG. 15, the multidimensional final vector becomes a 703-dimensional vector system (10+50+100+200+12+7+24+100+200) in which 9 individual vectors are combined.

When k learning objects are generated for k similar documents (S810), the learning processing unit 1900 of the present invention performs a learning algorithm using the profile information generated for each document object for the k learning objects (S520 ) Among the profile information, weight information representing the probability that the profile information corresponding to the learning object exists is generated for each k similar document (S830).

As described above, this weight information is the probability of existence of profile information including one or more elements corresponding to the learning object among the profile information of document objects collected by document centralization, or numerical size information similar to the learning object. Since it is represented, it means the probability that the requester reads similar documents (access, etc.).

For example, in the case of the processing of the present invention, whether or not the information requester's affiliation information (department, team, etc.) is related to each other and the degree of the correlation can be fed back as a result of calculation based on quantitative values, so that the relevance of the affiliation information is information It can function as a basis for recommending similar documents to the requester.

That is, the learning processing unit 1900 of the present invention provides the corresponding element among the storage path information of the k similar documents, the first time information at which the request signal was input, the information about the requestor, and the elements about the belonging information of the requestor. The weight information is generated by applying the number of and the degree of similarity of corresponding elements in combination.

In this way, when weight information is generated in the learning processing unit 1900, the main processing unit 1500 of the present invention generates final ranking information that comprehensively reflects the weight information and similarity ranking information (quantitative numerical value for similarity in content) (S840 ) It is configured to provide the similarity list information according to the final ranking information to the requester (S850).

Although the present invention has been described above with limited examples and drawings, the present invention is not limited thereto and will be described below and the technical spirit of the present invention by those skilled in the art to which the present invention belongs. Of course, various modifications and variations are possible within the scope of the claims.

In the description of the present invention described above, modifiers such as first and second are only terms of instrumental concepts used to relatively distinguish components from each other, so they are used to indicate a specific order, priority, etc. It should be interpreted that it is not a term that

Although the accompanying drawings and the like for illustration of the description of the present invention and its embodiments may be shown in a slightly exaggerated form in order to emphasize or highlight the technical contents according to the present invention, the above-described contents and matters shown in the drawings Taking into account, it should be interpreted that it is obvious that various types of modifications can be applied at the level of those skilled in the art.

The above-described category classification method, document recommendation method, and similar document recommendation method of the present invention may be implemented as computer readable codes on a computer readable recording medium. Computer-readable recording media include all types of recording devices (CD-ROM, RAM, ROM, floppy disk, magnetic disk, hard disk, magneto-optical disk, etc.) in which data that can be read by a computer is stored. It also includes a server for transmission.

100: document classification system 200: user terminal
300: web server 400: mobile terminal
103: document DB unit 105: category setting unit
107: text extraction unit 110: first input unit
120: vector generator 130: representative vector setting unit
140: object vector generation unit 145: data processing unit
150: clustering unit 160: second representative vector setting unit
165: second data processing unit 170: naming unit
171: selection unit 173: calculation processing unit
175: naming selection unit 180: monitoring unit
190: processing control unit 195: feedback processing unit
500: document recommendation system 510: document object DB unit
520: history information generation unit 530: request information input unit
540: first selection unit 550: recommendation document selection unit
551: period weight processing unit 552: continuous weight processing unit
553: proximity weight processing unit 554: text selection unit
555: frequency weighting unit 556: tracking unit
557: Attention level processing unit 558: Selective processing unit
560: list provision unit
1000: Similar document recommendation system 1100: Document DB unit
1200: Embedding vector generator 1300: Request input unit
1400: Similar document selection unit 1500: Main processing unit
1600: first ranking information generating unit 1700: profile generating unit
1800: learning object generation unit 1810: individual vector generation unit
1820: object creation unit 1900: running processing unit

Claims (7)

a document object DB unit in which document objects, which are documents collected by document centralization, are stored;
When the document object is used by a user, a history information generator for storing history information including user information using the document object and usage time information when the document object is used in association with the document object;
a request information input unit into which a request signal including requester information and at least one request period information among a plurality of different set periods is input;
a first sorting unit which selects a used document object, which is a document object used by a user corresponding to the requestor information, from among the document objects by using the history information;
a recommendation document sorting unit which selects one or more recommended documents, which are document objects having use time information relatively close to the reference point in time of the request period information, among the used document objects; and
and a list providing unit generating recommendation list information corresponding to the recommended document and providing the generated recommendation list information to the requester. The method of claim 1, wherein the recommendation document selection unit,
a periodic weight processing unit that assigns first rank information to each used document object, and assigns relatively higher first rank information as the distance between the use time information of the used document object and the reference time point of the request period information is smaller; and
and a selection processing unit for selecting the recommended document based on the size of the first rank information among the used document objects. The method of claim 2, wherein the recommendation document selection unit,
Further comprising a continuous weight processing unit that assigns second rank information to each used document object, and assigns relatively high second rank information as the continuously used period is longer or the number of times of use is greater,
The selection processing unit,
A customized document recommendation system using dynamic change of time-series pattern information, characterized in that the recommended document is selected based on the calculated size of the first and second rank information among the used document objects. The method of claim 3, wherein the recommendation document selection unit,
Further comprising a proximity weight processing unit for extracting the last use time information from among the use time information included in the history information of the used document object, and giving a relatively high third rank information as the last use time information is closer to the current time point; ,
The selection processing unit,
A customized document recommendation system using dynamic change of time-series pattern information, characterized in that the recommended document is selected based on the calculated size of the first to third rank information among the used document objects. The method of claim 4, wherein the recommendation document selection unit,
A plurality of texts corresponding to nouns are selected by performing morpheme parsing processing on the file name information of the used document object, and among the selected plurality of texts, the top m (m is a natural number of 2 or more) main texts based on the frequency of use. a text selection unit that selects; and
A frequency weighting processing unit that assigns fourth rank information to each used document object, and assigns a relatively high fourth rank information as the number of main texts included in the file name of the corresponding used document object among the top m main texts increases. include,
The selection processing unit,
A customized document recommendation system using dynamic change of time-series pattern information, characterized in that the recommended document is selected based on the calculated size of the first to fourth rank information among the used document objects. The method of claim 2, wherein the recommendation document selection unit,
Further comprising a tracking unit for generating tracking information including at least one of information on the number of times the content of the document object has been modified or information on a modification rate of a modified portion of the entire content of the document object,
The selection processing unit,
A customized document recommendation system using dynamic change of time series pattern information, characterized in that the recommended document is selected based on the size of the first rank information and the calculated size of the tracking information among the used document objects. The method of claim 2, wherein the recommendation document selection unit,
Further comprising an attention processing unit that assigns an attention index to each document object, and assigns a relatively high attention index as the number of users who have modified the content of the document object increases, as a user other than the user corresponding to the requester information. ,
The selection processing unit,
A customized document recommendation system using dynamic change of time-series pattern information, characterized in that the recommended document is selected based on the size of the first rank information and the calculated size of the attention index among the used document objects.

Images