KR20210001681A - A method for operating a service providing of corporate insolvencies information based on automatic sentiment information labelings - Google Patents

Abstract

According to an embodiment of the present invention, an operation method of a service providing device comprises the steps of: setting terminology dictionary information of a news article labeling module to predict a possibility of the insolvency of an enterprise; collecting article information labeled by the news article labeling module according to the terminology dictionary information setting according to a preset period in response to the prediction of the possibility of the insolvency of the enterprise; predicting insolvent enterprise probability information based on a prediction model based on the collected article information; and providing an insolvent enterprise information service using the insolvent enterprise probability information to a user terminal.

Description

A METHOD FOR OPERATING A SERVICE PROVIDING OF CORPORATE INSOLVENCIES INFORMATION BASED ON AUTOMATIC SENTIMENT INFORMATION LABELINGS}

The present invention relates to a method and apparatus for providing a corporate insolvency information service. More specifically, the present invention relates to a method and apparatus for providing a corporate insolvency information service according to prediction of corporate insolvency based on emotional information labeling of a news article.

In line with the development of the financial industry, as awareness of risk management spreads, financial institutions are developing techniques to evaluate corporate credit in various ways and applying them to actual work. In general, a company's credit evaluation is performed through the company's financial information and sales information.

In order to find out the signs of corporate insolvency, a financial model based on a statistical numerical linear model was created using the explicit sentence or numerical information described in the structured data based on financial information or public information, and the possibility of default was predicted using this.

In the traditional default prediction, a model that evaluates the possibility of a corporate default by reflecting the current financial status and status of a company using financial information is used as the most accurate and objective method.

However, since financial information is prepared only on a regular basis after each company's settlement of accounts, there is a limitation in immediately reflecting rapid changes in business management status. To compensate for this, a default prediction model is used that utilizes the company's stock price information, but the company's stock price is the result of real-time evaluation by investors and reflects information such as the company status and financial status most quickly, but the stock price information is financial. Depending on the market and the situation of the relevant industry, it is highly likely to fluctuate with other variables besides the value of the company, and there is a limitation that stock price information cannot be used except for listed companies.

Therefore, there is a need to develop a technology that can complement the method of analyzing the cause of bankruptcy based on limited financial information for companies that have caused bankruptcy.

On the other hand, with the rapid spread of high-speed internet in recent years, the internet has become an indispensable necessity in modern life. Many users use various services such as news article search and e-commerce through the Internet. Accordingly, news articles related to the company (company sales promotion article, company management change article, company share sale article, company disclosure article, etc.) have been generated from numerous media companies and distributed to users through the Internet.

Recently, by collecting such news information and constructing big data to classify and analyze labels, attempts have been made to develop various applications and provide services to grasp the awareness or propensity of a specific person or company, or to identify recent trends. This is being accelerated by the introduction of artificial intelligence technology based on the 4th industrial revolution.

In particular, companies can refer to marketing strategies by grasping how their products or services are being evaluated on social media, and policy agencies can determine policy revision directions and promotion methods through public opinion analysis on policies. As such a need arises, brand monitoring services specialized in emotion analysis through data mining are also becoming active.

However, the method of analyzing news information based on artificial intelligence technology to date mainly uses semi-supervised learning, and a small number of experts on news information directly read articles and assign labels to create a small set of reference data. It is a method of processing emotional labeling for all news information using this, and in the end, excessive time and cost are spent in the process of handling all news articles directly by experts.

In addition, the performance and efficiency of this classmate learning method are still low, so only high-cost and low-efficiency services can be provided, and commercialization and popularization are practically difficult.

The present invention was conceived to solve the above problems, constructing dictionary data based on paragraph-based sentence structure and terminological analysis of documents, and processing automated emotional label tagging corresponding to a large amount of online news articles that are unstructured texts. Thus, by performing the predictive processing of the possibility of corporate insolvency using the sentiment information labeling data of the refined news article based on learning, it is possible to predict and automate the possibility of an efficient, logical and accurate corporate insolvency. An object thereof is to provide a method and apparatus for providing an information service.

A method according to an embodiment of the present invention for solving the above-described problems includes: setting term dictionary information of a news article labeling module in order to predict a possibility of corporate insolvency; Collecting article information labeled by the news article labeling module according to the term dictionary information setting according to a preset period in response to prediction of the possibility of corporate insolvency; Predicting probability information about a probability of corporate insolvency based on a prediction model based on the collected article information; And providing a corporate insolvency information service using the corporate insolvency probability information to a user terminal.

In addition, an apparatus according to an embodiment of the present invention for solving the above-described problems includes: a term presetting unit for setting term dictionary information of a news article labeling module in order to predict a possibility of corporate insolvency; An article information collection unit that collects article information labeled by the news article labeling module according to the term dictionary information setting according to a preset period in response to the prediction of the possibility of corporate insolvency; An insolvency probability predictor for predicting information on probability of insolvency probability of a company based on a predictive model based on the collected article information; And a service provider providing a corporate insolvency information service using the corporate insolvency probability information to a user terminal.

Meanwhile, the method according to an embodiment of the present invention for solving the above-described problems may be implemented with a program for executing the method on a computer and a recording medium in which the program is recorded.

According to an embodiment of the present invention, dictionary data based on paragraph-based sentence structure and terminology analysis of a document is constructed, and automated emotional label tagging corresponding to a large amount of online news articles is processed, and refined news articles By performing a learning-based process of predicting the possibility of corporate insolvency using emotional information labeling data, it enables more accurate news information-based sentiment analysis while minimizing the time and cost of manpower input, and complex analysis of conventional financial data. It is possible to provide a method and apparatus for predicting corporate insolvency based on emotional information labeling of news articles that can predict the possibility of corporate insolvency quickly, accurately and logically in an automated manner, compared to the unpredictable and uncertain corporate analysis tool.

In addition, according to an embodiment of the present invention, by predicting the possibility of a company's insolvency based on analysis data of a news article, which is an unstructured text, based on a modularized process, its causal relationship and logic are clear and reasonable prediction is possible.

In addition, since the prediction of the possibility of insolvency is automated and processed with an accurate probability-based numerical value, it is possible to prevent in advance an inaccurate prediction due to human association as in the past, and the data collection period can be shortened.

In addition, according to an embodiment of the present invention, the online news article data is composed of larger and larger big data over time, and can be expanded to national and regional units, so the analysis accuracy and efficiency can be determined by time and data processing performance. It gets higher according to.

1 is a conceptual diagram schematically showing an entire system according to an embodiment of the present invention.
2 is a more detailed block diagram of a labeling apparatus according to an embodiment of the present invention.
3 is a block diagram showing in more detail an apparatus for providing corporate insolvency information service according to an embodiment of the present invention.
4 is a flowchart illustrating a labeling method according to an embodiment of the present invention in more detail.
5 is a diagram illustrating a labeling process and a document verification and filtering process according to an embodiment of the present invention.
6 is a more detailed flowchart illustrating a method of providing a corporate insolvency information service according to an embodiment of the present invention.
7 and 8 are diagrams for explaining terminology preset information for providing a corporate insolvency information service according to an embodiment of the present invention.
9 and 10 are diagrams for explaining a collection data period setting of labeled article information for providing a corporate insolvency information service according to an embodiment of the present invention.
11 is a diagram illustrating a table of information on probability of insolvent probability calculated according to a corporate insolvency information service according to an embodiment of the present invention.

The following content merely illustrates the principles of the present invention. Therefore, those skilled in the art can implement the principles of the present invention and invent various devices included in the concept and scope of the present invention, although not clearly described or illustrated herein. In addition, it is understood that all conditional terms and examples listed in this specification are, in principle, expressly intended only for the purpose of making the concept of the present invention understood, and are not limited to the embodiments and states specifically listed as such. Should be.

In addition, it is to be understood that all detailed descriptions listing specific embodiments as well as principles, aspects and embodiments of the present invention are intended to include structural and functional equivalents of these matters. It should also be understood that these equivalents include not only currently known equivalents, but also equivalents to be developed in the future, that is, all devices invented to perform the same function regardless of structure.

Thus, for example, the block diagrams herein are to be understood as representing a conceptual perspective of exemplary circuits embodying the principles of the invention. Similarly, all flowcharts, state transition diagrams, pseudocodes, etc. are understood to represent various processes performed by a computer or processor, whether or not the computer or processor is clearly depicted and that can be represented substantially in a computer-readable medium. Should be.

In addition, the explicit use of terms presented as processor, control, or similar concepts should not be interpreted exclusively by referring to hardware capable of executing software, and without limitation, digital signal processor (DSP) hardware, ROM for storing software. It should be understood to implicitly include (ROM), RAM, and non-volatile memory. Other commonly used hardware may also be included.

The above-described objects, features, and advantages will become more apparent through the following detailed description in connection with the accompanying drawings, whereby those of ordinary skill in the technical field to which the present invention pertains can easily implement the technical idea of the present invention. There will be. In addition, in describing the present invention, when it is determined that a detailed description of a known technology related to the present invention may unnecessarily obscure the subject matter of the present invention, a detailed description thereof will be omitted.

Hereinafter, a preferred embodiment according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a diagram schematically showing an entire system according to an embodiment of the present invention.

Referring to FIG. 1, a system according to an embodiment of the present invention may include an emotional information labeling apparatus 100, a corporate insolvent information service providing apparatus 200, and a user terminal 300, and the emotional information labeling apparatus ( 100) may be connected to the news article database 400.

More specifically, the emotional information labeling apparatus 100 obtains news document data from the news article database 400, extracts paragraph information from the obtained document data, and is obtained from the document data according to the paragraph information. Articles in which the emotional label is tagged by comparing the paragraphs with the pre-built learning-based dictionary unit, determining an analysis subject emotional information label corresponding to the relevant paragraph to be analyzed, and tagging the analysis subject emotional information label to the one or more paragraphs Process it into information.

Accordingly, the processed article information may be provided to the corporate insolvent information service providing device 200, and the corporate insolvent information service providing device 200 is sentiment information based on news data for predicting corporate insolvency through the emotional information label. It is possible to collect and perform prediction processing of corresponding insolvency probability information, and perform a service provided to the user terminal 300 by processing corporate insolvency information including prediction processing information.

In particular, the enterprise insolvency information service providing device 200 provides the term dictionary and rule setting information used for labeling in the emotional information labeling device 100 to enable tagging of emotional information labels specialized for predicting corporate insolvency. , Based on this, related article information tagged with a label may be collected according to preset range and period information, and the possibility of insolvency for each company based on the emotional information label of the collected article information may be predicted based on probability. Here, the probability-based prediction may be processed according to a prediction model for insolvency based on learning data, and detailed article information collection and prediction processes will be described in more detail later.

On the other hand, the emotional information labeling device 100, the service providing device 200, and the user terminal 300 may be connected by wire or wirelessly through a network, and each emotional information labeling device 100, a service for communication between networks The providing device 200 and the user terminal 300 are provided through an Internet network, LAN, WAN, PSTN (Public Switched Telephone Network), PSDN (Public Switched Data Network), cable TV network, WIFI, mobile communication network, and other wireless communication networks. You can send and receive data. In addition, each of the emotional information labeling apparatus 100, the service providing apparatus 200, and the user terminal 300 may include respective communication modules for communicating with a protocol corresponding to each communication network.

In addition, the user terminal 100 described in the present specification includes a mobile phone, a smart phone, a laptop computer, a terminal for digital broadcasting, a personal digital assistant (PDA), a portable multimedia player (PMP), and a navigation system. Although it may be included, the present invention is not limited thereto, and may be various devices capable of user input and information display.

In such a system, the user terminal 300 may be connected to the company insolvency information service providing apparatus 200 to receive the news data-based analysis target company's insolvency information providing service.

2 is a block diagram illustrating in more detail an apparatus 100 for labeling emotional information according to an exemplary embodiment of the present invention.

2, the emotional information labeling apparatus 100 according to an embodiment of the present invention includes an input unit 110, a paragraph information extraction unit 120, a label determination unit 130, a dictionary unit 140, and a log generation unit. It includes a unit 160, a verification unit 170, and an article information processing unit 180.

First, the input unit 110 receives a news document for processing article information from the news article database 400 and transmits it to the paragraph information extraction unit 120.

The input unit 110 may include one or more input interfaces for receiving the above-described news article data from the news article database 400. For example, the news document may be determined according to the news service, may be a comprehensive article document of each portal site, an article document of a news provider, and further include an article evaluation comment or an article-related novel media (SNS) document. .

In addition, the input unit 110 may receive information to be analyzed. The information to be analyzed may include, for example, a company name, a company name, a person name, and the like.For example, article information to which an emotional information label corresponding to a specific company is assigned is corporate insolvency information that provides prediction information of the company. It may be provided to the service providing device 200.

The paragraph information extracting unit 120 may extract paragraph information for classifying the news document into one or more paragraph data that can be assigned a label.

In particular, in the case of online news articles, especially economic-related articles, there are not many articles that are filled with only one company in one article, and in many cases, they are filled with content about two or more companies. In this case, it is difficult to automatically classify the positive or negative evaluation of a specific company, and the positive or negative sentiment evaluation corresponding to the entire document and the positive or negative evaluation of each individual company often do not coincide. Evaluation can be difficult.

Therefore, the paragraph information extraction unit 120 according to an embodiment of the present invention extracts only the paragraph containing a specific analysis object (company, etc.), rather than using the entire original text of the news article as it is, and processes it as article information tagged with a related emotional label. Output is possible, a process of extracting paragraph information for this is performed, and the extracted paragraph information may be transmitted to the label determination unit 130.

In addition, the label determination unit 130 indexes each paragraph of the document by referring to the paragraph information and the keyword and sentiment data set of the dictionary unit 140, and scores according to a specific keyword or sentiment word of the analysis target information for each paragraph. According to the matching comparison operation, the emotion information label to be tagged may be primarily determined, and the determined emotion information label and paragraph information may be transmitted to the verification unit 170.

Here, the dictionary unit 140 used for extracting paragraph information by the paragraph information extracting unit 120 and comparing score matching by the label determining unit 130 is, according to a keyword or sentiment frequency of a paragraph related to the analysis object. It may include a keyword set and a sentiment word set that enable determination of the sentiment label score of the analysis target.

More specifically, the dictionary unit 140 according to an exemplary embodiment of the present invention may be pre-built according to the provision of setting information from the apparatus 200 for providing a corporate insolvency information service and be provided in the emotional information labeling apparatus 100. The dictionary unit 140 includes a plurality of language data sets built on a learning basis to determine an analysis target emotional information label of a paragraph obtained from the document data, wherein the language data set is an analysis target from the paragraph. Includes a keyword set capable of identifying a noun-type term having a positive or negative meaning corresponding to the economy or credit of, or the language data set is a verb having a positive or negative meaning corresponding to the reputation of the analysis object from the paragraph or It may include a set of sentiment words that can identify adjective terms.

Here, the positive keyword set and the negative keyword set enable positive and negative identification of individual news articles, and may include document feature words (or feature words) used for machine learning based on this.

For example, the positive keyword set and the negative keyword set may include keywords having a positive or negative meaning related to economic information or credit information of an analysis target company expressed in a news article, and may generally be a noun-type word. , Can be used as a marker in evaluating whether a news article is positive or negative.

The apparatus 200 for providing corporate insolvency information service includes setting information to the dictionary unit 140 to construct keywords constituting a positive keyword set and a negative keyword set for more accurately predicting a corporate insolvency prediction in the dictionary unit 140. Send to set a certain keyword in advance, collect and set keywords from expert documents or various technical documents in related fields, or related keywords related to existing collected keywords by a related word search method (eg Word2vec). By setting a list or performing a process of registering and setting only keywords determined according to filtering among the derived related word keywords, a positive keyword set and a negative keyword set may be generated and stored in the dictionary unit 140.

For example, the dictionary unit 140 may include a set of positive and negative sentiment words. A sentiment word set can have a verb form such as'increase' or'decrease' in a news article, or an adjective form such as'good' and'bad', and it is one that expresses the emotion of the analysis target through combination with keywords. It may include the above sentiment words.

The sentiment word can be used as a value variable to determine the emotional affirmation or negativity of a news article, and a document feature (or feature word, features) for machine learning in the apparatus 200 for providing corporate insolvent information, such as a keyword set. ) Can also be used.

Accordingly, the dictionary unit 140 constructs a set of sentiment words only with terms that are not limited to a specific domain area, but collects sentiment words from expert documents or various technical documents in a related field, or a related word search method (for example, Word2vec) generates a list of related sentiment words related to the previously collected sentiment words, or registers only sentiment words determined by filtering among the derived list of sentiment words, positive sentiment set and negative. You can create and store sentiment sets.

Accordingly, for terms registered in the dictionary unit 140, the label determining unit 130 may give different scores according to the type of the term (positive, negative, keyword, sentiment word), and the total is It can be used when determining the final positive or negative emotional label.

The label determination unit 130 identifies a paragraph related to an actual analysis object from the news document using the dictionary unit 140 generated as described above and the paragraph information of the paragraph information extraction unit 120, and responds to the identified paragraph. You can decide which emotional label to do. Here, the emotional label may be an automated label classified as positive or negative. In addition, news documents are news of various fields produced by various media outlets, and there may be no significant differences in terms of publication format, although there is a difference in description method for each medium.

More specifically, when a single news article is input, the label determination unit 130 may perform indexing for each paragraph based on the paragraph information extracted by the paragraph information extracting unit 120 and check the indication variable. Here, a paragraph may mean a set of sentences above and below the document divided by one or more blank lines. Therefore, a single sentence can be a paragraph.

To this end, the paragraph information extracting unit 120 first performs paragraph indexing of news document data, identifies a paragraph containing information to be analyzed among the indexed paragraphs, and sets the center of the index range corresponding to the identified paragraph. The indicator variable for can be assigned to the identified paragraph.

In addition, the label determination unit 130 may determine at least one of an index range and an index order of a target paragraph for determining an emotional label based on the indicator variable. In addition, the label determination unit 130 may determine a positive or negative emotional score of the relevant paragraph to be analyzed according to a matching comparison operation with the dictionary unit 140.

More specifically, the label determination unit 130 uses the analysis target information (for example, a list of company names) input through the input unit 110 to determine in which paragraph the analysis target information exists in the news document. You can index within the index range centered on.

If the analysis target information does not exist in any paragraph in the document, the corresponding news article index is terminated, and the input unit 110 receives the next news article.

For example, if the number of searched paragraphs in the news document is 5 and a specific company name is found in the third paragraph, the paragraph information extracting unit 120 may assign 2 to index_p (an indicator variable for paragraph indication). Then, the label determination unit 130 may repeatedly perform a label score operation corresponding to the specific company name within the upper and lower two paragraphs centering on the 2, and processing of article information or news according to the result of performing the score operation You can decide to skip the document.

In addition, the label determination unit 130 may determine an analysis target emotional information label to be tagged in the paragraph according to the score determination. It can be determined whether the paragraph tagged with the emotional information label is processed into article information by the verification unit 170, and the processing unit 180 determines article information tagged with the emotional label based on the verified paragraph and the emotional information label. Can be processed.

For example, the label determination unit 130 calculates +2 in response to a positive keyword generated in a paragraph based on the paragraph information, calculates -2 in response to a negative keyword, and calculates +2 in response to a positive sentiment word. By calculating 1.5 and calculating -1.5 in response to negative sentiment words, the sentiment label comparison matching score with the dictionary 140 may be determined. Here, the score is an example and may be changed as necessary.

In addition, after the emotional label for each positive or negative is determined in the label determination unit 130, various evidence records used to determine the label are configured as log data through the log generation unit 160, and the external file (or database) Can be stored in the form. Log data can be used as a basis for labeling emotional information.

In addition, when a plurality of article documents each including paragraphs to which the label of the emotion information to be analyzed is first tagged are output according to the score comparison in the label determination unit 130, the verification unit 170 performs the emotion of the plurality of article documents. A verification value is calculated according to a similarity comparison operation for each classification, and a document to be processed into article information and an analysis target emotion information label to be tagged to the article information are finally determined based on the verification value.

Accordingly, the verification unit 170 may perform a verification process to ensure stability of the emotional information label tagging result. However, the verification process at this time can also perform a mechanical process to minimize human intervention and perform automatic processing.

More specifically, the verification unit 170 collects the labeling result of the label determination unit 130, calculates the similarity between news article paragraphs tagged with the same emotional label, and calculates the similarity level for each news article paragraph according to the document similarity measurement method. After accumulating, ranking may be determined according to the accumulated size, and a process of removing subordinate news article paragraphs may be performed. Such a process may include a process of normalizing the result for convenience of the customer and dividing the section into [1, 0~1, 0], and the verification unit 170 sets a threshold corresponding thereto, and A positive or negative emotional label can be determined for only paragraphs that fall within the range of or not. This will be described later in more detail with reference to FIG. 6.

On the other hand, the article information processing unit 180 may output the paragraph tagged with the finally determined emotional label as article information, which is provided to the corporate insolvent information service providing device 200, and predicts corporate insolvency information to be analyzed. It may be used for a service, and a prediction service may be provided to the user terminal 300.

In addition, the emotional information labeling device 100 according to an embodiment of the present invention can be applied not only to news articles or report data composed of long texts, but also to various newsletters composed of relatively short texts, online comments, and SNS data. have.

In addition, the emotional information labeling device 100 according to the embodiment of the present invention may be provided on the system as a separate device, but each functional element is modularized to be provided as an element of the corporate insolvent information service providing device 200. I can. In this case, the apparatus 200 for providing a corporate insolvency information service according to an embodiment of the present invention may include all or some functions of the emotional information labeling apparatus 100. Conversely, the emotional information labeling apparatus 100 may include all or part of the functions of the apparatus 200 for providing a corporate insolvency information service.

Accordingly, the embodiment of the present invention may be implemented as a single service providing system including the corporate insolvency information service providing device 200 and the emotional information labeling device 100, and is limited by separation or combination between the modules and functions. It does not become.

3 is a block diagram for explaining in more detail an apparatus for providing corporate insolvency information service according to an embodiment of the present invention.

Referring to FIG. 3, the apparatus 200 for providing a corporate insolvency information service according to an embodiment of the present invention includes a term presetting unit 210, an article information collecting unit 220, an insolvent probability predicting unit 230, and a service provider. Includes study 240.

The term presetting unit 210 configures preset information for predicting corporate insolvency information and insolvent probability, and transmits the configured preset information to the dictionary 140 of the emotional information labeling apparatus 100 to transmit the preset information to the dictionary 140 ), a positive keyword set, a negative keyword set, a positive sentiment set, and a negative sentiment set are generated.

The term presetting unit 210 may set one or more keywords that can increase prediction performance and accuracy in the dictionary unit 140 according to an insolvent probability prediction model, and further, by setting score information for each keyword type, It can be processed to be used for information labeling.

To this end, the term dictionary setting unit 210 may construct term dictionary generation rule data in which variables can be set, and may transmit setting information based on the term dictionary generation rule data to the dictionary unit 140.

In addition, the insolvency probability prediction model is a processing model used in the insolvency probability prediction unit 230, and may be a learning-based data prediction model, which may be different according to each data attribute, and a more specific model and prediction process are It will be described later in FIG. 6.

In addition, the article information collection unit 220 sets a data collection period for each company classification according to the request of the insolvent probability prediction unit 230, and obtains the emotional information labeled article information for each set period from the emotional information labeling device 100 do.

For example, the article information collection unit 220 may request article information within a first period compared to the observation point to the sentiment information labeling device 100 in response to an observation point to be predicted by the insolvency probability predictor 230 In addition, the emotional information labeling apparatus 100 may collect article information within the first period processed by the article information processing unit 180 and transmit it to the article information collection unit 220.

Here, the first period may be changed according to the characteristics of the insolvent probability prediction model and training data, and may be determined differently according to the characteristics of individual companies. For example, the first period may be determined differently for each case, such as the case of a company with high probability of bankruptcy and a company with low probability of bankruptcy, small and medium-sized enterprises and large enterprises. In addition, the first period may be subdivided and determined according to a window sliding method according to learning of a default point based on accumulated training data.

On the other hand, the insolvency probability prediction unit 230 applies to the insolvency probability prediction model based on the collected data for the first period of emotionally labeled article information, and outputs the probability of insolvency of the company based on the applied result.

In this case, the insolvency prediction model may be a learning-based prediction model, and the insolvency prediction unit 230 performs a machine learning process by using article information labeled with emotional information as training data, and performs a machine learning process based on the performance result. The predicted probability information may be configured and output to the insolvency probability predictor 230.

In addition, the service provider 240 may provide a corporate insolvency information service to the user terminal 300 based on the insolvency probability prediction probability information output from the insolvent probability predictor 230. Here, the corporate insolvency information service includes a process of processing the possibility of insolvent information for each company obtained from the probability information of predicting the probability of insolvency, and a transmission process of converting the processed information about possible insolvency into video, audio, or image to the user terminal 300 can do. In addition, the user terminal 300 may perform user registration and payment processing for receiving a service, and may transmit specific company information for predicting the possibility of insolvency to the service provider 240 in advance.

For example, the service provider 240 may receive a list of companies requesting insolvency prediction from the user terminal 300, and obtain and process insolvency prediction probability information corresponding to the list from the insolvency prediction unit 230 In addition, the processed information on predicting probability of insolvency may be transmitted to the user terminal 300.

FIG. 4 is a flowchart for explaining in more detail an operating method of the emotional information labeling apparatus 100 according to an embodiment of the present invention, and FIG. 5 is a diagram illustrating a label tagging process and a document verification and filtering process accordingly to be.

4 illustrates a process of outputting article information tagged with a positive or negative emotional label associated with a company name when an analysis target is a company name according to an embodiment of the present invention. The company name may be composed of a list, and analysis target information is set. Depending on the company name, group name, corporation name, person name, etc. may be changed.

Referring to FIG. 4, first, the emotional information labeling apparatus 100 according to an embodiment of the present invention receives a new target article document through the input unit 110 (S101).

In addition, the emotional information labeling apparatus 100 performs paragraph indexing by extracting paragraph information corresponding to a new target document through the paragraph information extracting unit 120 (S102), and checks whether the company name information for each paragraph is indexed ( S103).

Here, a paragraph may mean a set of sentences divided into upper and lower blank lines while the document is divided into one or more blank lines. Therefore, a single sentence can be a paragraph.

In addition, the emotional information labeling apparatus 100 may index where the company name in the company name list exists in the news document through the paragraph information extracting unit 120. If the company name is not indexed in any paragraph in the news article document, the news article is ignored and the next news article document is entered.

When the company name is indexed, the emotional information labeling apparatus 100 allocates an indicator variable corresponding to the paragraph in which the company name information is indexed through the paragraph information extraction unit 120 (S105).

For example, if the number of searched paragraphs is 5 and the name of the company to be analyzed is found in the third paragraph, 2 may be assigned to index_p (indicating variable for paragraph indication).

Accordingly, the paragraph information extracting unit 120 may request the label determination unit 130 to perform a label determination process by moving to the paragraph designated by the index_p by using the index_p.

In addition, the emotion information labeling apparatus 100 designates a target paragraph range and order to determine the emotion label centered on the indicator variable index_p through the label determination unit 130.

Thereafter, the emotional information labeling apparatus 100 performs positive or negative keywords corresponding to each paragraph and positive or negative keywords according to the target paragraph range and order corresponding to the current news article document through repeated execution of steps S109 to S121. After indexing negative sentiment words, scoring is performed to output a paragraph tagged with a final sentiment information label.

More specifically, these repetitive processes can be processed only for paragraphs determined by the target paragraph range and order indicated by a specific indicator variable, not all paragraphs of the news article. One upper paragraph of a paragraph, two lower paragraphs, and the title paragraph of a news article document may be set as the target paragraph range.

In addition, the sequential processing order of the target document is preferably set in the order of the current paragraph, the lower first paragraph, the upper paragraph, the lower second paragraph, and the title paragraph, and the label determination unit 130 is registered in the dictionary unit 140. The scoring process may be performed according to whether terms are indexed, and the log generator 160 may record log information corresponding to the label determination.

First, the emotional information labeling apparatus 100 designates a current paragraph as a target paragraph through the label determination unit 130 (S109).

Then, the label determination unit 130 determines the type of the target paragraph based on the term information of the dictionary unit 140 (S111).

Here, the type is whether to process separate article information (go to step S112), move to the next paragraph (go to step S121), or skip the current news article document and receive the next article document (go to step S101). It may be an index to represent, and the label determination unit 130 may determine a paragraph type according to whether a keyword and a sentiment word registered in the dictionary unit 140 are indexed and a frequency number.

More specifically, if the paragraph type is 0 to 2, the label determination unit 130 may extract and generate the paragraph currently designated as index_p as a separate article to which relationship information with the originally input article document is allocated (S112). ).

For example, when the original article id is 12345678, the article id of the separately extracted and generated paragraph may be designated as 12345678_1, or separate common relationship metadata within the article may be included.

In addition, the label determination unit 130 may determine a positive or negative score according to a matching comparison operation with the dictionary unit 140 in response to the extracted and generated article (S113), and determine an emotional label based on the score to generate a log. A log is recorded through the unit 160 (S115), and the determined emotional label is tagged to the target paragraph and output to the verification unit 170 (S117).

More specifically, the label determination unit 130 may assign a score for each registered keyword and sentiment words of the dictionary unit 140 indexed within the extracted paragraph, and determine the sum score. And basically, the'positive' emotional label or the'negative' emotional label can be determined according to the positive or negative total score of the paragraph. In addition, when the total score is within a certain range (eg, -1.5 to 1.5), the emotional label may be determined as'neutral'. The range of these values can be adjusted according to the properties of data or news articles.

Thereafter, the emotional information labeling apparatus 100 checks whether the last paragraph through the label determination unit 130 (S119), and moves to the next paragraph within the target paragraph range in a specified order (S121), or the next news article document It is possible to return to the input step S101.

Meanwhile, the step S121 may be performed even when the type index is 3, and when the type index is repeatedly determined to be 3, the label determination unit 130 moves the target paragraph to the top, bottom, etc. The decision process can be repeated. However, if 3 is repeated even after processing the repetitive type determination within the target paragraph range, the label determiner 130 may determine the type index as 4. In this case, the current article is set to type 4 and may be performed again from step S101 in which the next news article document is input.

For example, referring to FIG. 5A, it can be seen that the inputted news article document is composed of seven paragraphs P1, P2, P3, P4, P5, and P7.

Accordingly, the emotional information labeling apparatus 100 obtains the paragraph composition information through the paragraph information extracting unit 120, and among them, the fourth paragraph (P4) in which the name of the company ('company B') as the analysis target information is found. ) Can be set as the reference paragraph (index_p) and the pointer can be moved to index_p.

In addition, the emotional information labeling apparatus 100 processes a keyword (kword) and an emotional word (eword) index through the label determining unit 130. In FIG. 4, the keyword and sentiment words do not exist in the current fourth index_pth paragraph P5, so the label determiner 130 may move to the index_p+1th paragraph P5. Since there are positive keyword 1, positive sentiment 1, negative keyword 2, and negative sentiment 2,3 in the P5 paragraph, if the sentiment score is calculated, sentiment_score = +2+1.5-2-1.5-1.5 = -1.5. The label determiner 130 may determine that it has a final negative sensibility and output a paragraph article labeled with negative sensibility information.

In addition, log data recorded as the label determination unit 130 performs an automated labeling process according to an embodiment of the present invention may be configured, and the structure thereof is a news information log, a paragraph information log, and term information as follows. It can be illustrated as a table including a log.

The news information log (News info.) is a log record of the news article itself, and may be composed of an article ID and original content.

Paragraph info. is a log record of extracted paragraphs, including the paragraph ID and the extracted paragraph content (filtered paragraph), the scored sum obtained through labeling corresponding to the terms that occurred, and It includes a decided label based on this.

The term information log (Lexicon info.) is a log record of terms constituting the extracted paragraph, and the term and type that occurred (terms per type), frequency, score, and counts per type ) Can be composed of.

In addition, FIG. 5(B) is a diagram for explaining a document verification and filtering process according to an embodiment of the present invention. Referring to FIG. 5(B), the verification unit 170 according to an embodiment of the present invention includes a document Based on the similarity measurement and the similarity accumulated value, a news article to be processed into article information labeled with emotional information may be verified, and finally articles to be output and provided to the service providing device 200 may be determined.

For example, assuming that news data with 10 positive emotion labels is generated as shown in FIG. 5(B), the verification unit 170 uses the terms registered in the dictionary unit 140 to verify the document Can be generated, and the similarity cumulative value is calculated after calculating the similarity between the articles containing the paragraphs labeled with the sentiment information and other news articles previously generated using the document verification model.

In this case, the verification unit 170 may perform normalization as necessary to calculate the cumulative value of the similarity for each news article, and then position each article in three sections as shown in [1, 0 to 1, 0]. An article whose value is 1 indicates that it is almost similar to other documents in terms of similarity and, for example, may be determined as an article sufficiently expressing positive emotions. However, when the value is 1, the frequency is generally low.

In the case of [1~0], articles that are partly similar and partly different from other documents may be determined as'positive' articles or articles that may be determined as'neutral' according to the user's judgment.

However, contrary to the numerical meaning, even if the actual similarity value is low (for example, 0.4), the positive sentiment value cannot be considered low, and differences may occur because the composition of terms is different from other articles belonging to the same category. .

Accordingly, in the case of [1 to 0], the verification unit 170 may set an upper or lower limit boundary value, check only samples of articles located within the boundary value, or set a separate reference value.

On the other hand, when the similarity is 0, the verification unit 170 may determine that the similarity value is too low or has characteristics that are different from other articles, and thus is'neutral' or'impossible to judge'. In this case, as in the case of 1, only a small number of articles may correspond to this.

Accordingly, when the similarity accumulated value exists among the three sections as shown in FIG. 6, the verification unit 170 may perform a process of removing specific news articles from the processing target by setting a threshold. .

5(B) shows a case where the threshold value is set to 0.32, and accordingly, the verification unit 170 removes documents 7 and 8 of news articles having 1.28 and 2.17 values lower than 0.32 accordingly.

In addition, only news article numbers 1 to 6 and 9 to 10 are finally left in the positive emotional news article set that has been verified by the verification unit 170, which is transmitted to the processing unit 180 to label the positive emotional information. Can be processed and printed as articles.

On the other hand, the verification unit 170 performs a similarity verification process for a set of news articles having a negative sensibility in the same manner, and collects only the news articles finally determined as having a negative sensibility by passing the verification as to whether or not the negative sensibility is , By passing it to the processing unit 180, negative emotion information may be processed and output as labeled articles.

6 is a flowchart illustrating the operation of the apparatus for providing corporate insolvent information service according to an embodiment of the present invention, and FIGS. 7 to 11 are diagrams illustrating the operation of the apparatus for providing corporate insolvent information service according to an embodiment of the present invention. These are diagrams illustrating a data structure processed by.

Referring to FIG. 6, the apparatus 200 for providing a corporate insolvent information service sets a term dictionary rule through the term preset unit 210 and sets a keyword for the dictionary unit 140 of the emotion information labeling apparatus 100. It performs (S201).

7 and 8 are views for explaining terminology preset information for providing a corporate insolvency information service according to an embodiment of the present invention, and FIG. 7 is a diagram illustrating keyword setting information according to an embodiment of the present invention. 8 is a diagram illustrating terminology dictionary rule setting information according to an embodiment of the present invention.

As described above, the dictionary unit 140 may process keyword setting for a positive or negative label, and the term presetting unit 210 may perform setting processing to be efficient in predicting corporate insolvency information.

In particular, as shown in FIG. 7, the terms of the keyword set in the term dictionary of the dictionary unit 140 generated for predicting the possibility of insolventity are positive nouns such as price limit width and rapid rise, and negative nouns such as audit, business difficulties, and restraint It may include. In addition, the terms of the sentiment set may consist mainly of verbs and adjective-type vocabularies, and may include positive terms such as stress, overcoming, high, etc., and may include negative terms such as diminishing and failing. .

In order to set such a keyword set, the term dictionary setting unit 210 may select domain-specific noun-type terms that become the basis thereof using related experts and document information. For example, the term dictionary information may be determined according to a vocabulary extraction process of reference documents collected from outside to predict the possibility of corporate insolvency.

More specifically, the term dictionary setting unit 210 may extract adjectives and verb vocabulary by performing a POS (Part of Speech) tagging process corresponding to various reference documents collected from outside to set a sentiment word set. In addition, the term presetting unit 210 may select and output a certain number of higher terms equal to or greater than a threshold value by calculating a term frequency inverse document frequency (tf-idf) value after deduplication processing corresponding to the extracted words. have. These vocabularies may be determined as sentiment word set terms through a positive or negative determination input.

In addition, the term presetting unit 210 may extract terms for each type, which is a basis, from reference documents collected from outside. To this end, the term presetting unit 210 extracts a predetermined number of terms with a high proximity ranking for each term using word2vec operation that calculates the proximity according to the front and rear relationship between each word in a vector form, Related terms may be obtained by performing redundancy removal, and terms having a tf-idf value higher than a preset threshold value may be extracted from the related terms and set by the dictionary 140.

Accordingly, the dictionary unit 140 may set terms for each type (positive/negative keywords, positive/negative emotional words) for predicting the possibility of corporate insolvency using news data. In addition, the dictionary unit 140 may set the distribution point information for label tagging of the emotional information labeling device 100, and this distribution point information is processed by the term presetting unit 210 and transmitted to the dictionary unit 140. The term dictionary generation rule setting information for determining the allocation information may be previously stored in the term preset unit 210.

FIG. 8 illustrates the configuration of term dictionary rule setting information according to an embodiment of the present invention. Point information according to the term dictionary rule setting information is determined for each keyword and sentiment type and transmitted to the dictionary unit 140. have.

For example, as shown in FIG. 8, the score of each positive keyword and positive emotional word according to Rule 1 may be greater than 0, and the value of the positive keyword according to Rule 2 is negative matching it. The keyword value may be the same as the negative calculation value, and according to Rule 3, the value of the positive sentiment word may be the same as the negative calculation value of the negative sentiment value matching it, and the absolute value of the keyword assigned value is the absolute value of the sentiment score. May be greater than the value.

Meanwhile, the article information collection unit 220 collects the article information according to the enterprise insolvency prediction model and the analysis section from the emotional information labeling device 100 (S203).

9 and 10 are diagrams for explaining a collection data period setting of labeled article information for providing a corporate insolvency information service according to an embodiment of the present invention, and FIG. 9 is a collection by company classification according to an embodiment of the present invention. Data period setting information is shown, and FIG. 10 shows period setting information using a window sliding method.

The article information collection unit 220 collects the news article information labeled from the emotional information labeling device 100 for predicting corporate insolvency, without limiting the field, and articles across all fields (eg, sports, art, etc.) News articles from all media (eg local media, specialized media by field, etc.) that are recognized for their quality can be collected.

Due to the nature of the news, there are many articles mainly on companies or large corporations that are well-known to people, but there are not many articles on small and medium-sized companies in the provinces. Although it is being used, all of the news of the company's new product advertisements, charity activities, support for various events, and participation can be helpful in generating emotionally positive signals for the company.

In addition, as described above, the collection unit 220 may preset the collection period of article information as a first period, or may be set as a complex period such as the first period and the second period, which is the type and classification of the company. It may be different according to.

For example, as shown in FIG. 9, in the case of bankrupt companies, the collection unit 220 collects articles for the past 14 months, including the year and month known to have caused bankruptcy for each individual company, and In this case, it is possible to collect articles for the past 12 months, including the month corresponding to the preset observation year. Accordingly, the collection unit 220 sets a first period corresponding to the insolvent company (data from 3 months to 14 months before the default date), and the second period corresponding to the normal company (12 months before the observation target month). Data) to collect article information from the emotional information labeling apparatus 100.

In addition, for example, as shown in FIG. 10, in the case of a bankrupt company, the collection unit 220 considers the rapid increase in articles related to insolvent in the last 3 months including the time when the bankruptcy occurred, News articles can be collected, and article information within a certain period set in a window-sliding method based on the time of bankruptcy can be collected in consideration of the different time periods of bankruptcy of individual companies.

More specifically, referring to FIG. 10, the collection unit 220 may collect article information of a bankrupt company as data according to a learning model for predicting corporate insolvency. The collection unit 220 determines a window section before, after, or before and after a predetermined period centering on the time when the bankruptcy occurs, and processes the sliding of the window section according to the time when the bankruptcy occurs for each company to determine the collection period. Can be set. Accordingly, the collection period set according to the window sliding may be different for each bankrupt company, and article information for each period may be individually collected and used for learning, training, and data analysis.

Here, in the case of a normal company, the collection unit 220 may use 12 months of data including the year and month at the time of observation as a target for collecting article information regardless of this, which means that the observation time is the current time (year and month). ), and trying to predict the possibility of the company's insolvency at the present time.

On the other hand, the enterprise insolvency information service providing apparatus 200 predicts the probability of insolvency based on the predicted model based on the collected article information through the insolvency probability predicting unit 230 (S205), and provides the service providing unit 240 A corporate insolvency information service based on the predicted probability information may be provided to the user terminal 300 (S207).

As described above, the insolvency probability prediction unit 230 according to an embodiment of the present invention constructs a learning model using positive and negative labeled keywords and sentiment words of article information, and uses the learning model to provide information on each period and company. It is possible to calculate and output default probability information according to an input.

FIG. 11 is a diagram illustrating a table of insolvency probability information calculated according to a corporate insolvency information service according to an embodiment of the present invention, and the insolvency probability may be calculated by being divided into calculated average probability information and final diagnostic probability information. .

More specifically, the insolvent probability prediction unit 230 from the training data consisting of a news article set having a positive emotion and a news set having a negative emotion 1) by company-by month-by emotional value-article frequency and 2) by company-by month -Each type of term frequency file (or DB) can be created. In addition, the insolvency probability predictor 230 may generate two or more predictive models for diagnosing insolvency using the data 1) and 2) described above. Here, the first model may be a monthly illegal article rate model and may have a continuous data attribute. In addition, the second model may be a monthly term score ratio model and may have discrete data properties.

For example, the insolvent probability prediction unit 230 includes a single model generation process (eg, SVM-Support Vector Machine, etc.) and a combined model generation process (eg, GB-gradient boosting) targeting continuous data. Etc.) can be used to create a first model. In addition, the insolvency prediction unit 230 uses a single model generation process (eg, DT-decision tree, etc.) and a combined model generation process (eg, RF-Random forest, etc.) targeting discrete data. Thus, a second model can be generated.

The calculation for predicting the probability of corporate insolvency according to each of these models may be described in more detail as shown in Table 1 below.

Model 1:
Irregular article
ratio ① Number of irregular incidents: The number of articles that were identified as'unholy' among articles collected monthly
② Total number of articles: Total number of articles collected per month

=> Calculate the percentage of irregular articles = ① Number of irregular articles / ② Total articles Model 2:
Positive negative
Keywords/sympathetic words
Score percentage ① Calculation of monthly positive and negative keywords and sentiment score:
Total scores of positive and negative keywords and sentiment words generated from articles collected monthly
Positive words for January: Sales (2), operating profit (2), surplus (2) >> 2+2+2=6
February negative words: Management category (-2), loss (-1.5) >> -2-1.5=-3.5
March Positive words: Orders (2), New (1.5), Negative words: Decrease (-1.5) >> 2+1.5-1.5=2
...

②Total: 12 months above from M-March to M-14M. ① Calculate the 12-month sum of the absolute value of each month's positive and negative keywords and scores
[6+3.5+2+ ... = 11.5]

=> Monthly positive and negative keywords and sentiment score rate calculation = Above ① / Above ②
January: 6 / 11.5 = 0.52
February: -3.5 / 11.5= -0.30
March: 2 / 11.5 = 0.17
...

Referring to Table 1, first, a monthly fraudulent article rate model based on the first model may be constructed, and the insolvent probability prediction unit 230 responds to the total number of articles per month for each target company regardless of the default/normal company. , After calculating the proportion of articles identified as negative emotions, a numerical value normalized in units of one year can be obtained as learning training data.

In addition, a keyword and sentiment score ratio model based on the second model can be constructed, and the insolvent probability prediction unit 230 sums the distribution points of terms for each type generated in monthly articles for each target company regardless of the default/normal company. Then, through the arithmetic processing as shown in Table 1, a numerical value normalized for each year may be obtained as learning training data.

At this time, insolvency probability prediction unit 230, in order to solve the problem of insufficient number of articles in small and medium-sized companies, training data for predicting insolvency only for companies with an average number of articles of 50 or more for two years. It can be processed so that it can be included as.

Meanwhile, the process of predicting the likelihood of insolvency according to an embodiment of the present invention may include a process of predicting the likelihood that insolvency will occur among normal companies that have not experienced bankruptcy within the last two years. Therefore, in order to apply machine learning according to each learning model, the insolvent probability prediction unit 230 uses a cross-validation method to set the total number of bankrupt companies and normal companies as the same number, and 90% of all companies are training group and 10 % Is set as a prediction group, and prediction probability calculation processing based on the above-described training data may be performed.

For example, the insolvent probability prediction unit 230, in analyzing the information of more than 100 articles over a four-year period, when the number of bankrupt companies is 100 and the number of normal companies is 100, in the case of bankrupt companies, consider the size and characteristics As a result, 90 companies are set as training sets and 10 companies are set as test sets, and in the case of normal companies, cross-validation is applied, and companies 1 to 90 are trained at the first attempt, and companies 91 to 100 are used. Can be set as test sets and applied to the learning model.

And, the insolvent probability prediction unit 230 fixed the data set of the bankrupt company in the second trial, but only for normal companies, the training set for companies (11 to 100) times, and the set for testing companies (1 to 10) It is set to, and the average value can be obtained by repeatedly learning 10 times in the same manner. Meanwhile, the insolvent probability prediction unit 230 may determine a test set through random sampling for 10% of the normal companies, and then set the remaining companies as a training set to obtain an average value after repeated learning.

On the other hand, in order to process performance improvement while considering the characteristics of the data constituting the first model and the second model, as shown in FIG. 11, the insolvency prediction unit 230 is a single type (SVM, DT) and After performing the combined (GB, RF) machine learning learning process, it is possible to calculate the final probability of insolvency by integrating the result values.

At this time, in the case of DT and RF that calculate discrete numerical values, there are often cases where the amount of articles per company is insufficient and the probability value calculated per month appears to be 0 or 50% or less due to the nature of the process (upper right square in Fig. 11). In consideration of this, the insolvency probability predictor 230 may calculate the final insolvency probability information as a diagnosis probability by obtaining an average of only the remaining calculated values excluding these values (lower right square in FIG. 11 ).

For example, at the time of observation in April 2018, if Company A has SVM and GB values of 80% and 70% (first model), respectively, and DT and RF values are 0% and 55% (second model), respectively. , When the predicted probability of failure is p_i∈{p_SVM,p_GB,p_DT,p_RF}, the failure probability prediction unit 230 performs a combined operation as shown in (Equation 1) below to diagnose a final predicted probability value of 68.3%. It can be calculated as probability information.

The apparatus 200 for providing corporate insolvency information service and the apparatus 100 for labeling emotional information according to an embodiment of the present invention may be used for various purposes of utilizing large-capacity news data of a big data scale. In particular, it can be applied to the development of credit management for various purposes, including forecasting the possibility of corporate insolvency by using recent online news articles, forecasting the stock market of listed companies, analyzing trends by industry, or predicting personal credit for specific purposes. .

In addition, the enterprise insolvency information service providing device 200 and the emotional information labeling device 100 according to an embodiment of the present invention are generally an early warning system for investigating whether borrowers are insolvent by using only financial data. It can also be used for the purpose of supplementing the limitations of the default or default prediction method.

In addition, the enterprise insolvency information service providing device 200 and the emotional information labeling device 100 according to an embodiment of the present invention separate and modularize the article prediction process and the insolvent probability prediction process using machine learning technology, Since it secures a large amount of online news article data that has been labeled negatively and uses it to predict the possibility of corporate insolvency, it can be transplanted into various application systems in module units.

In addition, since the enterprise insolvency information service providing device 200 and the emotional information labeling device 100 according to an embodiment of the present invention can effectively use the article frequency-based prediction and the keyword-based prediction together, logically understand the cause and effect. In doing so, it can provide a basis for reasonable reasoning.

Meanwhile, the above-described method according to various embodiments of the present invention may be implemented as a program and provided to each server or devices while being stored in various non-transitory computer readable media. Accordingly, the user terminal 100 can access the server or device and download the program.

The non-transitory readable medium refers to a medium that stores data semi-permanently and can be read by a device, not a medium that stores data for a short moment, such as a register, cache, or memory. Specifically, the above-described various applications or programs may be provided by being stored in a non-transitory readable medium such as a CD, DVD, hard disk, Blu-ray disk, USB, memory card, and ROM.

In addition, although the preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above, and the technical field to which the present invention belongs without departing from the gist of the present invention claimed in the claims. In addition, various modifications are possible by those of ordinary skill in the art, and these modifications should not be individually understood from the technical spirit or prospect of the present invention.

Claims (10)

The sentiment labeling device acquires news document data from the news article database and processes it into article information tagged with the sentiment information label to be analyzed based on the term dictionary information, and the company's insolvent information service providing device is the sentiment processed by the sentiment information labeling device. In the operating method of a company insolvency prediction information service providing device that collects labeled article information to predict corporate insolvency and provides the predicted information to a user terminal device,
Setting term dictionary information of the emotional labeling device to predict the possibility of corporate insolvency;
Collecting article information labeled by tagging positive and negative sentiment information in the sentiment labeling device according to the term dictionary information setting for a bankrupt company and a normal company, respectively, according to a preset period in response to the prediction of the company's insolvency;
Based on the collected article information, the first model normalized after calculating the company's (Monthly fraudulent article frequency)/(All article frequency per month) for bankrupt companies and normal companies, and the company's (sum of terms in all articles per month) Calculating a predicted probability value using a second model of )/(|sum of the term score for 12 months|), calculating an average of the predicted probability values, and predicting a probability of a company insolvency; And
Providing a corporate insolvency information service using the corporate insolvency probability information to a user terminal.
A method of operating a company insolvency prediction information service providing device. The method of claim 1,
The terminology dictionary information includes terminology information of a keyword type and terminology information of a sentiment type used to determine a sentiment information label for predicting the possibility of corporate insolvency.
A method of operating a company insolvency prediction information service providing device. The method of claim 2,
The glossary information is
Includes allocation information corresponding to the term information of the keyword type and the term information of the sentiment type,
The allocation information is determined for each keyword and a term of a sentiment type according to a preset term dictionary generation rule corresponding to the term information of the keyword type and the term information of the sentiment type.
A method of operating a company insolvency prediction information service providing device. The method of claim 2,
The terminology information of the keyword type and the terminology information of the sentiment type are determined according to the vocabulary extraction process of reference documents collected from outside to predict the possibility of corporate insolvency.
A method of operating a company insolvency prediction information service providing device. The method of claim 1,
The preset period is,
Including a first period corresponding to a bankrupt company and a second period corresponding to a normal company
A method of operating a company insolvency prediction information service providing device. The method of claim 5,
The first period includes a certain range of periods based on when the bankruptcy of the bankrupt company occurs.
A method of operating a company insolvency prediction information service providing device. The method of claim 6,
The collecting step,
In order to determine the predetermined range period, determining a window section centered on the time when the bankruptcy occurred, and determining the range period for each company by processing the sliding of the window section according to the time when the bankruptcy occurred for each company.
A method of operating a company insolvency prediction information service providing device. The method of claim 1,
The predicting step,
Construct learning data including a news article set having a positive emotion and a news article set having a negative emotion from the collected article information,
Generating a predictive model including a first model, which is a percentage of negative articles per period having continuous data characteristics, and a second model, which is a ratio of term scores for each period having discrete data characteristics, using the training data,
Applying the training data to a period-specific ratio model based on the first model to obtain a normalized numerical value, and applying the training data to a period-based term score ratio model based on the second model to obtain a normalized numerical value Repetitively performing the learning process to calculate the probability of insolvency probability information for each model; including
A method of operating a company insolvency prediction information service providing device. The method of claim 1,
The predicting step comprises the step of setting the training set and the test set ratio between the bankrupt company and the normal company to be the same or different for each iteration of the iteration for cross-validation corresponding to the iterative performance.
A method of operating a company insolvency prediction information service providing device. The method of claim 1,
The single model generation process used to generate the first model is a support vector machine (SVM)-based process, and the combined model generation process is a gradient boosting (GB)-based process,
The single model generation process used for generating the second model is a DT (Decision tree)-based process, and the combined model generation process is a random forest (RF)-based process.
A method of operating a company insolvency prediction information service providing device.

Images