KR102132142B1 - Method and apparatus for recommending vocabulary from data dictionary based on natural language processing technique - Google Patents

Abstract

Provided is a method for recommending a term from a data dictionary which performs pre-processing for a data dictionary including a plurality of terms as data items and determines at least one term related to a target word from the pre-processed data dictionary based on a natural language processing technique to provide the at least one term as a recommended term.

Description

METHOD AND APPARATUS FOR RECOMMENDING VOCABULARY FROM DATA DICTIONARY BASED ON NATURAL LANGUAGE PROCESSING TECHNIQUE}

The following description relates to a method and apparatus for recommending terms from a data dictionary based on a natural language processing technique, in particular, ranking and recommending terms similar to target words from a data dictionary according to a statistical method or a machine learning based method It is about technology.

As the 4th industry is activated and interest in it is growing, various kinds of data are being generated in large scale in various industries and sectors such as IT, finance, economy, and healthcare, and a lot of resources are consumed to manage this data. Is becoming.

To manage data, business/data dictionary is used. These business/data dictionaries are tools for maintaining and managing business terminology used by companies in the field.

However, in the business/data dictionary, there is a problem in that duplicate words are generated and maintained, and additional costs are often required to organize, remove, and maintain these terms. Also, for example, in the case of a merger or acquisition of a company, or when overlapping terms (words) among terms used by affiliates are to be deleted, and shortened abbreviations are to be integrated, an integrated dictionary that incorporates multiple data dictionaries is redesigned. Need to build.

As such, in order to perform data integration and data dictionary standardization incorporating data dictionaries, it is important to determine similar terms (words) from terms (words) included in the data dictionaries.

On the other hand, Korean Patent Publication No. 10-2014-0133383 (published on November 19, 2014) is a data management device, data management method, and data management system that encrypts and stores data and keywords in an external storage space under a cloud environment. And, it discloses a technique for generating a searchable ciphertext for keywords and enabling retrieval of data including the corresponding keyword from keywords encrypted using a token for the keyword to be searched.

The information described above is for illustrative purposes only, may include content that does not form part of the prior art, and may not include what the prior art can present to those skilled in the art.

A data dictionary that performs pre-processing on a data dictionary that includes a plurality of terms as data items and determines at least one term related to a target word from a pre-processed data dictionary based on a natural language processing technique and provides it as a recommended term A method of recommending terms from can be provided.

For multiple data dictionaries composed of a plurality of data dictionaries, a method of ranking a plurality of terms with high similarity to a target word and outputting them as recommended terms using a statistical method or a machine learning based method may be provided. have.

In one aspect, a method for recommending a term from a data dictionary, performed by a computer system, comprising: preprocessing at least one data dictionary comprising a plurality of terms as data items, a natural language processing technique. Based on the target word, determining at least one term related to the target word from the pre-processed data dictionary and outputting the determined term as a recommended term for the target word How to recommend is provided.

The data dictionary is a multiple data dictionary composed of a plurality of data dictionaries, each of a plurality of terms that the data dictionaries include as data items is an abbreviation, and the step of performing the preprocessing is for each data dictionary of the data dictionaries , Extracting terms corresponding to a predetermined language and removing duplicate terms among the extracted terms to obtain pre-processed terms in which the duplicate terms are removed from the extracted terms. In the determining, among the plurality of terms, the terms having a similarity with respect to the target word are determined, and the outputting step may rank and output the determined terms.

The natural language processing technique is a statistical method, and determining the term includes: determining a word vector of the target word, calculating cosine similarity between each word vector of the pre-processed terms and the word vector of the target word. And determining at least two terms having a higher cosine similarity as the plurality of terms having a higher similarity.

The natural language processing technique is a machine learning-based method, and the step of determining the term uses a machine learning-based model trained to determine a plurality of terms having high similarity to the target word among the pre-processed terms. Thus, a plurality of terms having high similarity to the target word are determined, and the machine learning-based model is configured such that the learning rate, the number of epochs, and the concealment are performed so that the loss converges. At least one hyper parameter among hidden layers, hidden units, and activation functions may be adjusted and trained.

The machine learning based model may be a Word2Vec based model.

In determining the term, a window size used to determine the context of the target word is set to 5 or less, and the predetermined language may be the same language as the target word.

The method of recommending the term may further include outputting a notification indicating that registration of the target word to the data dictionary is necessary when the target word is not included in the plurality of terms.

Each data dictionary is a standard dictionary constructed based on data related to one company or industry, and standard dictionaries corresponding to the plurality of data dictionaries are associated with data related to companies or industries associated with the data dictionaries. Built by an integrated data storage management system that manages data storage systems for, as an integrated dictionary, the integrated dictionary, each of the plurality of terms, individual standard terms associated only with any one of the standard dictionaries, the It can be classified as a divisional standard term associated with at least two of the standard dictionaries, but not all, or a transcription standard term associated with all of the standard dictionaries.

When a specific term is searched as the target word through the integrated dictionary, whether the specific term corresponds to the individual term, the division term, or the transcription term, along with a plurality of terms having similarity to the specific term end Can be queried together.

The machine learning-based model is a trained model based on reinforcement learning, and the machine learning-based model can be updated based on a determination of a plurality of terms with high similarity to the target word. have.

Through the embodiments, not only a single data dictionary, but also multiple data dictionaries composed of a plurality of data dictionaries may be recommended a plurality of terms having high similarity to a target word, through which multiple data dictionaries Can be standardized.

By using natural language processing techniques to provide recommended terms for the target word, similar terms for the target word can be effectively determined even for a data dictionary containing abbreviations of English words.

It is possible to maximize the usability of the integrated dictionary that integrates multiple data dictionaries and the convenience of data management through the integrated dictionary.

1 illustrates a method for recommending terms from a data dictionary, according to one embodiment.
2 is a block diagram illustrating an example of an internal configuration of a computer system that recommends terms from a data dictionary, according to one embodiment.
3 is a flow diagram illustrating a method for recommending terms from a data dictionary, according to one embodiment.
4 illustrates a method of determining recommended terms for a target word using a statistical method as a natural language processing technique according to an example.
5 illustrates a size of a window used to determine recommended terms for a target word, according to an example.
6 shows a change in loss according to hyperparameter tuning in the case of using a machine learning-based method as a natural language processing technique according to an example.
7 shows a result of providing recommended terms for a target word according to an example.
8 shows the result of providing recommended terms for a target word from a single data dictionary, according to an example.
9 shows the result of providing recommended terms for a target word from a multiple data dictionary that includes a plurality of data dictionaries, according to an example.
10 illustrates a method of constructing an integrated data storage management system and querying data through the integrated integrated data storage management system, according to an example.
11 illustrates a method of integrating and linking data for each industry using an integrated data storage management system constructed according to an example.
12 illustrates a classification system of terms/domains/standards defined in an integrated integrated data storage management system according to an example.
13 and 14 illustrate a method for managing individual/sector/enterprise standards using an integrated integrated data storage management system, according to an example.
15 illustrates an integrated dictionary implemented using an integrated integrated data storage management system, according to an example.

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

1 illustrates a method for recommending terms from a data dictionary, according to one embodiment.

As shown, in an embodiment, based on the natural language processing technique, from the data dictionary 10 (eg, at least one of data dictionary 1 to data dictionary N), recommended terms 100 for a target word (eg, PRDT) ) (Eg, PDNO, FNNC, PPUL, NMCO and FLCU). The natural language processing technique may be, for example, a statistical-based method (statistical method) or a machine learning-based method.

The statistical-based method may be a technique for determining recommended terms 100 for a target word from the data dictionary 10 through statistical analysis.

The machine learning based method may be a technique of determining recommended terms 100 for a target word from the data dictionary 10 through a machine learning model, such as a deep learning model. The machine learning model is a deep learning model trained to determine (infer) recommendation terms 100 related to (eg, highly similar to) a given target word, eg, from data items in the data dictionary 10 Can. By applying this machine learning model to the data dictionary 10 and target words, terms related to the data dictionary 10 may be provided and output as recommended terms 50.

Here, the data dictionary 10 may include a term or word used in a specific industry, a company (institution), or an affiliate of a company as a data item. Meanwhile, one term may be composed of a plurality of words. The data item of the data dictionary 10 may be metadata.

The terms or words included in the data dictionary 10 are standardized (ie, standardized and registered) and may be standard terms or standard words. At this time, the data dictionary 10 may correspond to a standard dictionary (that is, a standard term dictionary).

The data dictionary 10 in the present disclosure may be a business (term) dictionary, or may be a standard term dictionary. Accordingly, terms included in the data dictionary 10 may be business terms or standard terms.

'Standardization' can mean a series of activities that establish standards such as codes, terms, and data domains and apply them consistently to a public database (DB), and standard terms (words) clearly communicate between users who use data. It can refer to the definition of a term consisting of a word and a combination of words that are used by each institution in order to support. The standard term dictionary can refer to a set of these standard terms and common standard terms. Each company (institution) and its affiliates may have separate standard term dictionaries.

In an embodiment, using natural language processing techniques, recommendations 100 related to a given target word, not only for a single data dictionary 10, but also from multiple data dictionaries 10 composed of multiple data dictionaries Can be provided.

As described above, a method of receiving the recommended terms 100 from the data dictionary(s) 10 using the natural language processing technique will be described in more detail with reference to FIGS. 2 to 15 to be described later.

2 is a block diagram illustrating an example of an internal configuration of a computer system that recommends terms from a data dictionary, according to one embodiment.

As shown in FIG. 2, the computer system 200 is a processor 210 as a component for executing a method for providing recommended terms 100 for a target word from the data dictionary(s) 10, It may include a memory 220, a permanent storage device 230, a bus 240, an input/output interface 250 and a network interface 260. The computer system 200 may be composed of a plurality of computer systems, unlike the one shown. The computer system 200 may be, for example, a server or other computer used by an enterprise or organization that manages and utilizes the data dictionary 10 (or data related thereto), its affiliates, or the head office.

Processor 210 may include or be part of any device capable of processing a sequence of instructions to implement a method for providing recommended terms 100 for a target word from data dictionary(s) 10. have. The processor 210 may include, for example, a computer processor, a processor in a mobile device or other electronic device, and/or a digital processor. The processor 210 may be included in, for example, a server computing device, a server computer, a series of server computers, a server farm, a cloud computer, a content platform, and the like. The processor 210 may be connected to the memory 220 through the bus 240.

The memory 220 may include volatile memory, permanent, virtual, or other memory for storing information used or output by the computer system 200. The memory 220 may include, for example, random access memory (RAM) and/or dynamic RAM (DRAM). The memory 220 can be used to store arbitrary information, such as status information of the computer system 200. The memory 220 stores instructions of the computer system 200, including, for example, instructions for performing a method to provide recommended terms 100 for a target word from the data dictionary(s) 10. Can also be used to Computer system 200 may include one or more processors 210 as needed or appropriate.

The bus 240 may include a communication infrastructure that enables interaction between various components of the computer system 200. The bus 240 may carry data between components of the computer system 200, for example, between the processor 210 and the memory 220. The bus 240 may include wireless and/or wired communication media between components of the computer system 200, and may include parallel, serial or other topology arrangements.

Persistent storage 230 is a component, such as memory or other permanent storage, as used by computer system 200 to store data for a predetermined extended period (eg, compared to memory 220). It may include. The permanent storage device 230 may include non-volatile main memory as used by the processor 210 in the computer system 200. The permanent storage device 230 may include, for example, a flash memory, hard disk, optical disk, or other computer readable medium.

The machine learning model used to provide the recommended terms 100 according to the aforementioned machine learning based method may be implemented in the memory 220 or the permanent storage device 230. Alternatively, such a machine learning model may be implemented on another computer system existing outside the computer system 200.

The input/output interface 250 may include interfaces to a keyboard, mouse, voice command input, display, or other input or output device.

The network interface 260 can include one or more interfaces to networks such as a local area network or the Internet. The network interface 260 can include interfaces for wired or wireless connections.

Also, the computer system 200 in other embodiments may include more components than those in FIG. 2. However, it is not necessary to clearly show most prior art components. For example, the computer system 200 is implemented to include at least some of the input/output devices connected to the input/output interface 250 described above, or a transceiver, a global positioning system (GPS) module, a camera, various sensors, Other components, such as a database, may also be included.

Through embodiments implemented through such a computer system 200, recommended terms 100 for a target word may be provided from a single data dictionary 10 or multiple data dictionaries 10 including a plurality of data dictionaries. Can.

The description of the above-described technical features with reference to FIG. 1 can be applied to FIG. 2 as it is, and duplicate description is omitted.

In the detailed description to be described later, an operation performed by a configuration (eg, processor 210) of the computer system 200 may be described as an operation performed by the computer system 200 for convenience of description.

Further, in the present disclosure, the term "dictionary" may be used interchangeably with "dictionary", and unless otherwise specified, both may be used interchangeably.

3 is a flow diagram illustrating a method for recommending terms from a data dictionary, according to one embodiment.

In step 310, the computer system 200 may perform preprocessing on at least one data dictionary 10 that includes a plurality of terms as data items. This pre-processing operation may be to process the data dictionary 10 to be suitable for application to natural language processing techniques (eg, statistical methods or machine learning-based methods) to be described later.

The data dictionary 10 may include a term or word used in a specific industry, company (institution), or affiliate of the company as a data item. The data item of the data dictionary 10 may be metadata. The terms or words included in the data dictionary 10 are standardized (ie, standardized and registered) and may be standard terms or standard words. At this time, the data dictionary 10 may correspond to a standard dictionary (that is, a standard term dictionary).

The data dictionary 10 may be a multiple data dictionary 10 composed of a plurality of data dictionaries.

Also, each of the plurality of terms included in each data dictionary 10 as data items may be an abbreviation. The abbreviation may be, for example, an abbreviation for English terms. In one example, the word "code" (term) can be expressed as an abbreviation for CD, CDE, CE, or CODE. The target word for determining the recommended terms 100 may also be an abbreviation. Each of the provided recommended terms 100 may also be such an abbreviation.

In the case where the data dictionary 10 is a multi-data dictionary 10 composed of a plurality of data dictionaries, in the pre-processing step of step 310, the computer system 200 determines a predetermined language for each data dictionary 10. Terms corresponding to can be extracted (step 312). At this time, the predetermined language may be Korean or English. Alternatively, the predetermined language may be other languages. The predetermined language may be the same language as the target word.

In step 312, when the term is Hangul, each of the words separated through a space or "_(underscore)" may be extracted as a term. For the pasted term, each of the terms is based on a morpheme analysis. Can be extracted. When the term is English or an English abbreviation, each of the words separated by a space or "_" may be extracted as a term.

In step 314, the computer system 200 may remove duplicate terms from the terms extracted from each data dictionary 10 in step 312. According to the removal, the computer system 200 may obtain pre-processed terms in which duplicate terms are removed among the extracted terms. For example, the computer system 200 may remove redundant terms (ie, redundant abbreviations) from among data dictionaries. The computer system 200 may remove the same abbreviations from the terms extracted from the data dictionaries and the original terms corresponding to the abbreviations. The computer system 200 may not remove any of the terms extracted from the data dictionaries, and only one of the original terms corresponding to the abbreviation or the abbreviation is the same.

Duplication of the terms extracted in step 314 by step 314 may be prevented.

In step 320, the computer system 200 may determine, for the target word, at least one term associated with the target word from the pre-processed data dictionary 10 according to step 310, based on the natural language processing technique. have. The target word can be input or selected from the user. The computer system 200 may determine a plurality of terms having a similarity to a target word among a plurality of terms included in the data dictionary 10.

When the data dictionary 10 is a multiple data dictionary 10 composed of a plurality of data dictionaries, the computer system 200 includes a plurality of terms (ie, multiple data dictionaries) included in the multiple data dictionary 10. Among the plurality of terms included in the field) may determine a plurality of terms with similarity to the target word.

The determined plurality of terms may correspond to the recommended terms 100 described above.

In step 330, the computer system 200 may output the determined term as the recommended term(s) 100 for the target word.

For example, the computer system 200 may rank and output a plurality of terms determined as having similarity to the target word. The outputted terms may correspond to the recommended terms 100 described above. On the screen of the computer system 200, the recommended term with the highest similarity to the target word may be exposed at the top. In addition, each recommended term may be exposed on the screen of the computer system 200 along with the similarity with the target word of the recommended term. Accordingly, the user can identify the recommended term having a high relevance to the target word and assumptions, and check the degree (similarity) of each recommended term related to the target word.

The target word described above may be included in a plurality of terms (words) included in the data dictionary 10. When the target word is not included in a plurality of terms included in the data dictionary 10, the computer system 200 may output a notification indicating that registration of the target word with the data dictionary 10 is required. . Accordingly, the user can first register the target word in the data dictionary 10.

In other words, in order to determine recommended terms 100 for a target word, a standardization operation may need to be performed in relation to the target word. If the target word is not registered in the data dictionary 10, similar terms may not be searched. After registration of the target word is completed, a similar term can be searched.

Meanwhile, a method of determining the recommended terms 100 related to the target word using the natural language processing technique will be described in more detail with reference to FIG. 4 to be described later.

The description of the technical features described above with reference to FIGS. 1 and 2 can be applied to FIG. 3 as it is, and a duplicate description is omitted.

4 illustrates a method of determining recommended terms for a target word using a statistical method as a natural language processing technique according to an example.

In step 410, the computer system 200 may determine a word vector of the target word. As described above, the target word may be included in a plurality of terms (words) included in the data dictionary 10. The computer system 200 may parse the target word and determine a word vector corresponding thereto.

In step 420, the computer system 200 may calculate the cosine similarity between each word vector of the pre-processed terms (with the redundant terms removed according to step 314) and the word vector of the target word. The computer system 200 may parse each of the preprocessing terms to determine a word vector corresponding thereto, and calculate cosine similarity between the word vector of the preprocessing term and the unit vector of the target word.

In operation 430, the computer system 200 may determine at least two terms having a higher cosine similarity as the plurality of terms having a higher similarity with respect to the target word. The determined terms may correspond to the recommended terms 100 described above.

As another example, the computer system 200 may calculate the Euclidean distance between each word vector of the preprocessing terms and the word vector of the target word, and based on the calculated Euclidean distance, as a plurality of terms having higher similarity to the target word You can also decide.

A method of calculating the cosine similarity between vectors may be expressed as Equation 1 below. x and y may represent each word vector.

[Equation 1]

According to the above-described steps 410 to 430, using a statistical method, a plurality of suggested terms similar to a target word from a single data dictionary 10 or multiple data dictionaries 10 including a plurality of data dictionaries ( 100) can be determined and provided.

The statistical method described above may be based on a distributional hypothesis that "the meaning of a word is formed by surrounding words." For example, a co-occurrence matrix may be constructed by calculating how many times a word appears around the target word, and a vector of words may be constructed from each row of the co-occurrence matrix. The size of the surroundings of the target word may be limited by the window size to be described later. Accordingly, the word vector(s) associated with the target word and the meaning of the target word can be grasped.

The following shows a method of determining recommended terms for a target word using a machine learning-based method as a natural language processing technique.

The computer system 200 uses a machine learning-based model trained to determine a plurality of terms with high similarity to a target word among pre-processed terms, in which duplicate terms are removed according to step 314). It is possible to determine a plurality of terms with similarity to words.

The machine learning based model may be a deep learning model, for example, a CNN based model, an RNN based model, or a reinforcement learning based model.

The CNN-based model, which uses a convolutional neural network (CNN), may be a kind of multilayer perceptrons designed to use minimal preprocessing. The CNN-based model may consist of one or more convolutional layers and general artificial neural network layers on top of it, and further utilize weights and pooling layers.

The RNN-based model uses a cyclic neural network, and a connection between units constituting an artificial neural network may be a model using a neural network constituting a directed cycle. The RNN-based model can utilize memory inside the neural network to process arbitrary input.

In the reinforcement learning-based model, the agent takes an action from the environment, and accordingly, the agent is given a reward and a state, and the agent is set to have the greatest reward. It can be a training model. Models based on reinforcement learning may be classified into models using a model-based algorithm or models using a model-free algorithm, depending on whether the agent has prior knowledge.

For example, in an embodiment, a model based on reinforcement learning (ie, a model trained based on reinforcement learning) may be used as a model based on machine learning. The model based on reinforcement learning may be continuously (or periodically) updated according to learning without user (ie, administrator) intervention.

Accordingly, the computer system 200 may update the machine learning based model based on the result after determining the recommended terms 100 by applying the machine learning based model. For example, the computer system 200 may be updated by learning the recommended terms 100 determined for the target word, and may be updated to further increase the accuracy of the reasoning of the later recommended terms 100. In other words, the model based on machine learning may be updated based on a determination of a plurality of terms (ie, recommended terms 100) having similarity to a target word.

The machine learning based model used for the determination (inference) of the recommended terms 100 of the embodiment is such that the learning rate, the size of the batch, the number of epochs, so that the loss converges , Hyper layers of at least one of hidden layers, hidden units, and activation functions may be adjusted to train/learn. Accordingly, the machine learning based model may determine recommended terms for the target word from the data dictionary 10 after the hyper parameters are adjusted and trained.

Depending on how you adjust the hyperparameters, the results of the inference can vary. Adjustment of these hyperparameters can be done through test or empirical methods. For example, if the learning rate is too large, overfitting occurs too quickly, and the result may span the local minimum, and it may also be undesirable when the learning rate is too small. The number of epochs may indicate how many times to repeat the training data set, and may be set to 50-300, for example.

In relation to this, FIG. 6 shows a change in loss due to hyperparameter tuning in the case of using a machine learning-based method as a natural language processing technique according to an example.

FIG. 6(a) shows the loss in a state in which the hyper parameters are not tuned, and FIG. 6(b) may represent the loss in a state after tuning the hyper parameters. It can be seen from FIG. 6(a) that the loss does not converge. The loss in FIG. 6 may indicate a cross entropy error, and iteration may be an epoch divided by a batch size. In an embodiment, as a hyper parameter, the size of the batch and the size of the concealment layer may be adjusted to obtain a result as shown in FIG. 6B.

Hyperparameters may include at least one of a learning rate, number of epochs, hidden layers, hidden units, and activation functions, as described above. . Loss can be determined by adjusting these hyper parameters. Depending on the loss, the results of training/learning and reasoning of the model may be different. Tuning of the hyperparameters can be performed by empirical or experimental methods. The tuning of the hyperparameters can be adjusted so that the loss reaches the absolute minimum rather than the local minimum.

The machine learning based model may be, for example, a Word2Vec based model. As an example, the machine learning based model may be a CBOW model as a Word2Vec based model.

Using a machine learning based method as described above, a plurality of recommended terms 100 similar to a target word are determined from a single data dictionary 10 or multiple data dictionaries 10 including a plurality of data dictionaries Can be provided.

The description of the technical features described above with reference to FIGS. 1 to 3 may be applied to FIGS. 4 and 6 as it is, and a duplicate description is omitted.

5 illustrates a size of a window used to determine recommended terms for a target word, according to an example.

In an embodiment, the window size used to determine the context of the target word can be set to 5 or less, or this window size can be set to 1.

In FIG. 5, the case where the window size is 1 is illustrated. As shown, in an embodiment, only words adjacent to the target word may be used to determine the context of the target word. The computer system 20 may change the size of the window according to the user's setting.

As such, in consideration of the determined context (the context represented by words adjacent to the target word) of the target word, terms similar to the target word may be determined as recommended terms 100 using a natural language processing technique.

The window size can be variously set according to the needs of the user. When the window size is increased, it may include more meaning around the target word, but because the surrounding words are included too much, the difference in context between words may be less. Therefore, it may be difficult to grasp the exact meaning of the target word. On the other hand, if the window size is too small, it can be determined that the target word contains only the meanings that are too local, so that the actual meanings of the target words are not grasped, but only the local meanings can be grasped.

In the embodiment, the window size is set to 1 or 5, but it can be set variously as needed.

The description of the technical features described above with reference to FIGS. 1 to 4 and 6 may be applied to FIG. 5 as it is, and a duplicate description is omitted.

Hereinafter, a method of determining and providing the recommended term 100 from the data dictionary 10 using the above-described natural language processing technique (machine learning-based method or statistical method) will be described in more detail.

As mentioned above, institutions and companies are using business/data dictionaries to improve data management in each industry, and are consuming significant resources to manage them. Meanwhile, the database in which the data dictionary is implemented has a limited number of bytes in each column, so when stored in the database, the term (English) can be abbreviated. These terms may be abbreviated differently depending on the company or institution, and may be abbreviated differently depending on the industry or affiliate within the company.

Therefore, it takes considerable resources to maintain and standardize such data. In addition, when data/business dictionaries are combined between companies/institutions, the integrated dictionaries need to be redesigned.

Through embodiments, it is possible to reduce the cost of companies/organizations in standardizing data and increase the efficiency of data management. In addition, in an embodiment, by applying a natural language processing technique, it is possible to increase the usability of a business/data dictionary including abbreviations of terms.

[Standardized Data Dictionary]

Standard terms may be business terms that can communicate with database users. to be. These terms can be coined by standardized words defined by public/governmental organizations (or corporations). The standardized term dictionary may represent a set of standardized terms and general standardized terms. The concept of a standardized term dictionary can be a standard, so it can be followed by different companies and industries.

[Natural Language Processing (NLP)]

Natural language processing (NLP) can be a discipline that trains computers to understand human language. For example, the natural language processing technique may include a thesaurus-based technique, a statistical technique, and an inference-based technique, and the computer system may be trained in different ways. The statistical method may be a method of processing natural language using a corpus. In order to extract important information from the corpus, the method may first divide the term from the corpus. Thereafter, the term is kept in a list of IDs, and these terms can be vectorized to present it in a distributed representation.

In the field of NLP, the distribution hypothesis can be defined as In the sense that the context of a term comes from the surrounding term, the meaning of the term may come from the surrounding term.

To track surrounding terms from the corpus, a co-occurrence matrix can be generated, and a window size can be defined to limit the number of terms surrounding the target term. For example, when the window size is 1, it may mean that the context should be limited to words immediately adjacent to the target word. The size of the window can vary, but in the case of ordinary business terms, it can be determined in that the word consists of 5 words or less.

As a method for evaluating vector similarity, there may be a method such as using Euclidean distance similarity and cosine similarity, for example.

The method of calculating the cosine similarity between vectors may be expressed as Equation 2 below. x and y may represent each word vector.

[Equation 2]

Statistical approaches can be improved using PMI (Pointwise Mutual Information). When using a co-occurrence matrix, the frequency of words can be independent of the window size. However, if the sentence contains a term such as'a', these terms may not be as important as other terms in finding the context of the word. PMI indicators can help you find these words when you find context, and eliminate these components.

Meanwhile, a method based on machine learning may be using models such as Word2Vec and GloVe. Models based on Word2Vec may include CBOW (Continuous Bag of Words) model and Skip-gram model. The CBOW model predicts the context of the target word using the surrounding words, and the Skip-gram model predicts the surrounding window of the context words using the target words. The GloVe model can use statistical and machine learning models.

[Multiple data standard integration]

For example, when a large multinational company wants to expand its business, individual subsidiaries can operate on their own, manage information and data separately, and create local standards. However, it is important to keep track of global standards to exercise control over subsidiaries. In the case of mergers and acquisitions (M&A), the parent company, the parent company, may want to take over subsidiaries by converting all regional standards into global standards. There are generally methods to create global standards between enterprises, centralized management and decentralized management.

By applying natural language processing to these different standards in the examples, it relates to integrating the standards.

In embodiments, statistical methods and machine learning-based methods may be used to standardize multiple data dictionaries. By using these two approaches, we propose a top five recommendation system that can list similar terms in different dictionaries.

For data, prior to applying the above method, pre-processing as described above with reference to FIG. 3 should be performed for data (ie, data dictionary 10).

First, (English) terms can be extracted from different data dictionaries. An English term may consist of several terms/words. Because there is a limit to the number of bytes stored for each column in the database, these dictionaries can consist of abbreviations (abbreviations). For example, the word "code" can be abbreviated as CD, CDE, CE or CODE. For Korean government agencies and companies, the English term is supplementary, so the term can be divided into "_". For example, the business term "Market English Name" can be abbreviated as "MRKT_ENG_NM" and stored in the dictionary. Can be. These abbreviations can generally be used because the column has limited storage capacity. At this time, it may be very important to extract these similar terms from business terms in that different businesses tend to abbreviate terms according to their rules, and there may be multiple abbreviated versions of the same word. . In addition, since the same abbreviation can be found in other databases, it may be necessary to delete duplicate terms.

[Apply natural language processing to data dictionary]

Since the data/business dictionary is composed of English (abbreviation), similarity between terms can be determined by applying a natural language processing technique to such a dictionary. Therefore, as described above with reference to FIGS. 3 and 4, recommended terms 100 may be determined and provided as similar terms for the target word.

Natural language processing techniques include, for example, statistical methods and machine learning-based methods, through which the top five similar terms can be determined as recommended terms.

As the data set, for example, privately owned data sets of three companies (X, Y, Z) were used. To exclude heterogeneity of the data set, a data set limited to financial sectors such as banks, credit card companies, and insurance companies was used. Since these three companies are in a similar field, the datasets owned by the three companies (X,Y,Z) may have similar business words/terms.

The company's dictionary contains 28987 standard words, 3461 unique words are included, the company's dictionary contains 2218 unique words and 14835 standard words, and the company's dictionary contains 4684 standard words and 2303 unique words. Is included.

In an embodiment, the window size is limited to size 1 as described above with reference to FIG. 5. Therefore, only one word adjacent to the target word can be used to determine the context. The window size can be changed according to the user's settings. The suggestion of the top five similar terms can also be changed according to the adjustment of parameters by the user.

[Statistics method]

Provision of recommended terms 100 according to a statistical method may be performed according to the following. First, the computer system 200 may search for a target word vector in a word list (eg, the pre-processing terms described above) searched for in the pre-processing device. Next, the computer system 200 may calculate all cosine similarities of all word vectors and target word vectors. Finally, the computer system 200 may output the top five items as recommended terms 100 based on the cosine similarity calculation result.

Statistical methods using cosine similarity may include a mutual information amount method for reducing unrelated word duplication and singular value decomposition, and a method for reducing a vector dimension when calculating cosine similarity for all words.

[Machine learning-based method]

As a machine learning based method, a model based on Word2Vec can be used.

In this example, three different models were trained: Word2Vec using only corporate X data, Word2Vec using corporate X and Y data, and Word2Vec using corporate X, Y, and Z data. Meanwhile, the hyperparameter tuning of the model may be performed so that the loss converges. Loss can mean the score by the softmax layer for the input and output layers of the model, and this probability can be applied to the cross entropy error with the correct label. The model's embedding layer can store a dense vector representation, reducing memory usage and unnecessary computation. In addition, negative sampling can reduce the computation of the Softmax layer, which reduces training time. In FIG. 6 described above, the loss of before (a) and after (b) tuning for the hyperparameter is shown.

[result]

As an evaluation according to the above example, the statistical method using the data of the company X (STAT1), the Word2Vec (INFER1) model trained with the data of the company X, the statistical method using the data of the company X and Y (STAT2), the company X and The trained Word2Vec model (INFER2) with Y's data was analyzed.

The results are shown in FIG. 7. 7 shows a result of providing recommended terms for a target word according to an example.

In the illustrated example, it can be confirmed that YMD was used as a target word, and that five terms with similarity are provided as recommended terms 100. As shown, each recommended term and the similarity (probability) associated therewith may be output together.

[Single data dictionary analysis]

8 shows the result of providing recommended terms for a target word from a single data dictionary, according to an example.

The corporate X dataset described above contains 28987 standardized business terms and 3461 unique words/terms. The result of determining the recommended terms 100 for the target word by selecting 10% of the terms is shown in FIG. 8.

As illustrated, in the case of applying a machine learning-based method, terms having different abbreviations may be provided as recommended terms although the same words such as SELL/SNS and NEW/NW are represented in a single data dictionary.

In addition, in the case of applying a statistical method, YMD may be provided as a recommended term as a word similar to the target word YM.

[Multiple data dictionary analysis]

9 shows the result of providing recommended terms for a target word from a multiple data dictionary that includes a plurality of data dictionaries, according to an example.

In order to provide recommended terms 100 for a target word from a multi-data dictionary, the pre-processing described above with reference to FIG. 3 may need to be performed on the multi-data dictionary.

In the example shown in FIG. 9, data of company X has 28987 business terms, 3461 unique words, data of company Y has 14835 business terms, and may have 2218 unique words. A total of 5679 unique words can be calculated, and among them, 518 words were used equally. Among the standardized words from which duplicate words were removed, 10% of personal terms were selected to determine results.

9 may indicate that different scenarios exist for the integrated data dictionary. As can be seen in Figure 9, SELL/SNS can be identified in the same dictionary. In addition, it can be seen that through different data dictionaries, the same word "transfer" is abbreviated in two different ways of TRSF/TRSFL.

Meanwhile, in the data dictionary of company X data, PRDT is defined as a product, and in the data dictionary of company Y data, PRNO is defined as a product number. This may mean that the abbreviations may be abbreviated differently or subdivided into smaller abbreviations, so that there is a high probability that other abbreviations of product numbers such as PRDTNO may exist instead of PRNO. The computer system 200 of the embodiment can determine that this possibility is high.

In addition, for KOSDAQ, which stands for KOSDAQ (Korean Securities Dealer Automated Quotation), it can be seen that STCK for abbreviating stock and DMST for domestic abbreviation are recommended in similar terms.

The description of the technical features described above with reference to FIGS. 1 to 4 may be applied to FIGS. 5 to 9 as it is, and redundant description will be omitted.

10 illustrates a method of constructing an integrated data storage management system and querying data through the integrated integrated data storage management system, according to an example.

Each of the plurality of data dictionaries described above can be managed in each system shown in FIG. 10 (ie, each system of A to C systems).

Each system may correspond to a data storage system for data in each data dictionary. The data in the data dictionary may be metadata, and the data storage system may be an integrated metadata repository (MDR).

The integrated data storage management system may be a system that integrally manages data for each of a plurality of systems. The integrated data storage management system may enable multi-repository management (MRM) for data of multiple systems (ie, multiple data dictionaries).

That is, each of the illustrated A system, B system, and C system is a data storage system, and may include a data dictionary.

The (each) data dictionary may be a standard dictionary built on the basis of data related to one company or industry, and such data dictionary is a system for storing and managing data related to the company or industry (A system, B shown) System or C system).

Standard dictionaries corresponding to a plurality of data dictionaries can be built as an integrated dictionary by an integrated data storage management system that manages data storage systems for data related to companies or industries associated with the data dictionaries. The integrated dictionary may correspond to the integrated standard dictionary shown in FIG. 10.

An integrated dictionary, each of the terms (included in the standard dictionaries), individual standard terms associated with only one of the standard dictionaries, sector standard terms associated with at least two of the standard dictionaries, but not all Or it can be classified as a transcription standard term that relates to all of the standard dictionaries.

Referring to the illustrated example, data items of respective data of the A system, the B system, and the C system may be mapped to respective standard dictionaries. For example, at least one of terms, words, and domains included in data of each system may be mapped to a standard dictionary.

In addition, the standard dictionary associated with each system can be mapped to the integrated standard dictionary. The integrated standard dictionary may be a standard dictionary of a group that manages companies/affiliates that manage the A system, the B system, and the C system. Standard dictionaries associated with each system may be mapped to an integrated standard dictionary by determining whether terms, words, or domains included in the standard dictionary correspond to individual/department/company standards (or terms).

Users of the integrated standard dictionary can query the data included in the standard dictionary associated with each system by querying the integrated standard dictionary.

For example, when a specific term is searched as a target word through an integrated standard dictionary (integrated dictionary), a specific term is an individual term, along with terms with high similarity to a specific term (ie, recommended terms 100). Whether it corresponds to a divisional term or a company term Can be queried together. Also, for each of the recommended terms 100, does each recommended term correspond to an individual term, a section term, or a transcription term? Can be queried together. The result of the inquiry may be displayed on the screen of the computer system 200.

The description of the technical features described above with reference to FIGS. 1 to 9 may be applied to FIG. 10 as it is, and a duplicate description is omitted.

Hereinafter, with reference to FIGS. 11 to 15, an integrated data storage management system and a method of performing MRM on data of a plurality of data storage systems (MDR) using the integrated data storage management system will be described in more detail. do.

Firms (having multiple industries) are premised on independent business operations in the case of expansion of business divisions or the establishment of individual corporations, so it is common that informatization is also promoted individually.

In addition, integrated management of each distributed metadata sharing target for data sharing systems that have various industries in the organization or exist in a physically separated state or for standard sharing and data sharing between group companies and affiliates Is needed. In addition, it is necessary to secure data compatibility between the distributed metadata management system based on the standards of each organization (or by industry) in order to share standards between multiple standards that occur due to the nature of the business and groups and affiliates.

In relation to this, MRM (Multi Repository Management) defines the standards of group companies for data sharing between group companies and affiliates based on multiple standards between individual affiliates (or industries) under the group, and based on the group standards (Global Standard). It may mean managing a mapping or inclusion relationship between each local standard. In other words, MRM manages enterprise-specific metadata managed according to the characteristics of different industries as one unified storage, creates individual standards for individual metadata, and continuously manages the structural relationship between group standards and individual standards. Can mean

To solve this, based on the integrated metadata repository (MDR: Meta Data Repository), mutual relations between standards existing in each metadata are defined, and through MRM, interoperability in data sharing between systems with different standards May need to be secured.

The main functions of the integrated data storage management system that provides MRM include the ability to collect multiple metadata in one integrated storage; Standard predefined and individual mapping functions for each data storage system; The ability to define and load standard dictionaries (words, terms, domains) for each metadata; It can include structure and relationship management between multiple standards (definition and relationship establishment of company/sector/individual standards).

[Concept of MRM]

Industry can have characteristics by industry. For example, the manufacturing industry may exist for each production line, the finance centers on the holding company, and in the case of an administrative agency, the regional offices and affiliated organizations may be centered around the main office. A separate IT system can be applied to each of these, and data related to it (meta data) is often managed separately.

MRM may be integrated management of metadata for each industry.

Further, the MRM may be to enable synchronization of standards, reference information, and other metadata between MDRs with respect to the MDR.

11 illustrates a method of integrating and linking data for each industry using an integrated data storage management system constructed according to an example.

In FIG. 11, an integrated data storage management system 1100 is shown, and each of MDR1 to 3 may correspond to the data storage system described above.

The words, terms, and/or domains (eg, domains recommended by computer system 200 described above) included in the data storage systems (ie, MDRs 1 to 3) as shown, are of the integrated data storage management system 1100. In the integrated MDR, it can be assigned to a transcription standard, a partial standard, or an individual standard. Accordingly, through the integrated data storage management system 1100, integrated inquiry and access to data of MDR1 to 3 and standard management may be possible.

12 illustrates a classification system of terms/domains/standards defined in an integrated integrated data storage management system according to an example. The illustrated companies A to C metadata systems may correspond to the data storage systems described above.

13 and 14 illustrate a method for managing individual/sector/enterprise standards using an integrated integrated data storage management system, according to an example.

As illustrated, standard terms and standard domains defined in each data storage system may be assigned to an enterprise standard (term), a sector standard (term), or an individual standard (term) by an integrated data storage management system.

As illustrated in FIG. 14, each of the standard (Korean) words of MDR 1 to 4 may be identified as corresponding to a transcription standard (term), a division standard (term), or an individual standard (term). In addition, domains (type, number of digits) corresponding to words and terms divided based on terminology among the collected metadata may also be identified as corresponding to a transcription standard, a division standard, or an individual standard.

15 illustrates an integrated dictionary implemented using an integrated integrated data storage management system, according to an example.

As illustrated, the integrated dictionary 1500 may be built in the integrated data storage management system. In the integrated dictionary, data dictionaries or standard dictionaries implemented in individual data storage systems may be mapped.

Terms, words, and domains may be defined in the integrated dictionary 1500. In addition, for each term, it may be defined whether the term corresponds to a transcription term, a section term, or an individual term. In addition, a structured structure between standards (associated with individual data storage systems) may be defined in the integrated dictionary 1500. Also, a system for terms may be defined in the integrated dictionary 1500.

The integrated data storage management system can collect data dictionary based on data collected from each institution (each data storage system) and standard dictionary information of each institution. In addition, the integrated data storage management system may collect mapping information linking the standard dictionary and meta information. The integrated data storage management system can perform mapping by creating a new standard dictionary without a standard dictionary. Accordingly, the integrated data storage management system can build the integrated dictionary 1500.

In constructing the integrated dictionary 1500, it is possible to determine whether or not a transcription/sector/individual standard (term) is applied to each term and timely indicate the term.

The description of the technical features described above with reference to FIGS. 1 to 10 may be applied to FIGS. 11 to 15 as it is, and overlapping descriptions are omitted.

The device described above may be implemented with hardware components, software components, and/or combinations of hardware components and software components. For example, the devices and components described in the embodiments may include a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor (micro signal processor), a microcomputer, a field programmable gate array (FPGA), and a programmable programmable logic array (PLU). It may be implemented using one or more general purpose computers or special purpose computers, such as a logic unit, microprocessor, or any other device capable of executing and responding to instructions. The processing device may perform an operating system (OS) and one or more software applications running on the operating system. In addition, the processing device may access, store, manipulate, process, and generate data in response to the execution of the software. For convenience of understanding, a processing device may be described as one being used, but a person having ordinary skill in the art, the processing device may include a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that may include. For example, the processing device may include a plurality of processors or a processor and a controller. In addition, other processing configurations, such as parallel processors, are possible.

The software may include a computer program, code, instruction, or a combination of one or more of these, and configure the processing device to operate as desired, or process independently or collectively You can command the device. Software and/or data may be embodied in any type of machine, component, physical device, computer storage medium, or device in order to be interpreted by the processing device or to provide instructions or data to the processing device. have. The software may be distributed over networked computer systems, and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. In this case, the medium may continuously store a program executable on a computer or may be temporarily stored for execution or download. In addition, the medium may be various recording means or storage means in a form of a single or several hardware combinations, and is not limited to a medium directly connected to a computer system, but may be distributed on a network. Examples of the medium include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks, And program instructions including ROM, RAM, flash memory, and the like. In addition, examples of other media may include an application store for distributing applications or a recording medium or storage medium managed by a site, server, or the like that supplies or distributes various software.

As described above, although the embodiments have been described by a limited embodiment and drawings, those skilled in the art can make various modifications and variations from the above description. For example, the described techniques are performed in a different order than the described method, and/or the components of the described system, structure, device, circuit, etc. are combined or combined in a different form from the described method, or other components Alternatively, even if replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (10)

A method for recommending terms from a data dictionary, performed by a computer system, comprising:
Pre-processing at least one data dictionary including a plurality of terms as data items;
Based on a natural language processing technique, for a target word, determining at least one term associated with the target word from the preprocessed data dictionary; And
Outputting the determined term as a recommended term for the target word
Including,
A data dictionary is a multiple data dictionary composed of a plurality of data dictionaries,
Each of the plurality of terms that the data dictionaries include as data items is an abbreviation,
The step of performing the pre-treatment,
For each data dictionary of the data dictionaries, extracting terms corresponding to a predetermined language; And
Obtaining pre-processed terms in which duplicate terms are removed from the extracted terms by removing duplicate terms from the extracted terms
Including,
Determining the term,
Among the plurality of terms, the terms having similarity to the target word are determined,
The step of outputting,
Rank and output the determined terms,
Each data dictionary is a standard dictionary built based on data related to one company or industry,
Standard dictionaries corresponding to the plurality of data dictionaries are built as an integrated dictionary by an integrated data storage management system that manages data storage systems for data related to companies or industries associated with the data dictionaries,
The integrated dictionary,
Each of the plurality of terms, with an individual standard term associated with only one of the standard dictionaries, a sector standard term associated with at least two but not all of the standard dictionaries, or all of the standard dictionaries Classified by the relevant transcription standard terminology,
When a specific term is searched as the target word through the integrated dictionary, the specific term is the individual standard term, the sector standard term, or the transcription standard term, along with a plurality of terms having similarity to the specific term. Does it correspond to How to recommend terms that are queried together. delete According to claim 1,
The natural language processing technique is a statistical method,
Determining the term,
Determining a word vector of the target word;
Calculating a cosine similarity between each word vector of the preprocessing terms and the word vector of the target word; And
Determining at least two terms having a higher cosine similarity as the plurality of terms having a higher similarity.
How to recommend a term, including. According to claim 1,
The natural language processing technique is a machine learning-based method,
Determining the term,
Among the pre-processed terms, a plurality of terms having high similarity to the target word are determined using a machine learning based model trained to determine a plurality of terms having high similarity to the target word,
The machine-learning-based model includes learning rate, number of epochs, hidden layers, hidden units, and activation functions so that the loss converges. ), wherein at least one hyper parameter is adjusted and trained. According to claim 4,
The machine learning-based model is a Word2Vec-based model, a method for recommending terms. According to claim 1,
Determining the term,
The window size used to determine the context of the target word is set to 5 or less,
A method of recommending a term, wherein the predetermined language is the same language as the target word. According to claim 1,
If the target word is not included in the plurality of terms, outputting a notification indicating that the target word needs to be registered in the data dictionary
A method of recommending a term, further comprising a. delete delete According to claim 4,
The machine learning based model is a model trained based on reinforcement learning,
The machine learning based model is a method for recommending a term, which is updated based on a determination of a plurality of terms with similarity to the target word.

Images