Method for replacing personal information and electronic device for performing same

Abstract

According to one aspect of the present disclosure, a method performed by an electronic device may be provided. The method may comprise a step of acquiring text including a character string representing personal information. The method may comprise a step for converting the character string into a reference word embedding by using a word embedding model. The method may comprise a step for identifying the distances between candidate word embeddings converted from words used for training the word embedding model and the reference word embedding. The method may comprise a step for selecting replacement candidates of the character string from among the candidate word embeddings on the basis of the identified distances. The method may comprise a step for replacing the character string representing the personal information with a character string corresponding to one of the replacement candidates.

Description

Methods for replacing personal information and electronic devices for doing so

[0001] The present disclosure relates to an electronic device and method for replacing a string representing personal information with another string using embedding.

[0002] As interest in and importance of personal information protection grows, regulations to protect it are continuously being introduced, and various technologies for protecting personal information are also being developed. In particular, the need to protect personal information within text containing personal information is growing. However, protecting personal information embedded in text can be challenging because the text is not fixed in a specific format. Therefore, there is a growing need for methods that can protect personal information while preserving its meaning, even in texts not formatted in a specific way.

[0003] According to one aspect of the present disclosure, a method performed by an electronic device may be provided. The method may include obtaining a text including a character string representing personal information. The method may include converting the character string into a reference word embedding using a word embedding model. The method may include identifying a distance between candidate word embeddings converted from words used in training the word embedding model and the reference word embedding. The method may include selecting replacement candidates for the character string from among the candidate word embeddings based on the identified distance. The method may include replacing the character string representing the personal information with a character string corresponding to one of the replacement candidates.

[0004] According to one aspect of the present disclosure, an electronic device may be provided. The electronic device may include at least one processor including processing circuitry; and a memory storing instructions. By individually or collectively executing the instructions by the at least one processor, the electronic device may obtain a text including a character string representing personal information. By individually or collectively executing the instructions by the at least one processor, the electronic device may convert the character string into a reference word embedding using a word embedding model. By individually or collectively executing the instructions by the at least one processor, the electronic device may identify a distance between candidate word embeddings converted from words used for training the word embedding model and the reference word embedding. The electronic device can select replacement candidates for the string from among the candidate word embeddings based on the identified distance by individually or collectively executing the instructions by the at least one processor. The electronic device can replace the string representing the personal information with a string corresponding to one of the replacement candidates by individually or collectively executing the instructions by the at least one processor.

[0005] According to one aspect of the present disclosure, a computer-readable recording medium having recorded thereon a program for executing any one of the aforementioned and hereinafter described methods for operating an electronic device and / or an electronic device can be provided.

[0006] FIG. 1 is a drawing for exemplarily explaining an operation of an electronic device according to one embodiment of the present disclosure to replace a string expressing personal information with another string.

[0007] FIG. 2 is a flowchart illustrating an operation of an electronic device according to one embodiment of the present disclosure to replace a string representing personal information with another string.

[0008] FIG. 3 is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to obtain a text including a string representing personal information.

[0009] FIG. 4 is a diagram for explaining an operation of an electronic device according to one embodiment of the present disclosure to convert a word into a word embedding.

[0010] FIG. 5 is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to convert a string representing personal information into a reference word embedding using a word embedding model.

[0011] FIG. 6A is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to identify an angle between each candidate word embedding and a reference word embedding.

[0012] FIG. 6b is a diagram illustrating an operation of an electronic device according to an embodiment of the present disclosure to identify a length between each candidate word embedding and a reference word embedding.

[0013] FIG. 7 is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to identify a distance between each candidate word embedding and a reference word embedding.

[0014] FIG. 8A is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to select a set number of candidate word embeddings as replacement candidates for a string.

[0015] FIG. 8b is a diagram for explaining an operation of an electronic device according to one embodiment of the present disclosure to select a set number of candidate word embeddings as replacement candidates for each reference word embedding.

[0016] FIG. 9A is a diagram for explaining an operation of an electronic device according to one embodiment of the present disclosure to select replacement candidates based on a reference distance.

[0017] FIG. 9b is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to delete a candidate word embedding from a replacement candidate based on a distance from a reference word embedding vector.

[0018] FIG. 10 is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to replace a string representing personal information with another string.

[0019] FIG. 11 is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to replace a string representing personal information with another string.

[0020] FIG. 12 is a block diagram illustrating an electronic device according to an embodiment of the present disclosure.

[0021] FIG. 13 is a block diagram exemplarily illustrating the configuration of a server according to one embodiment of the present disclosure.

[0022] Hereinafter, terms used in this specification will be briefly described, and the present disclosure will be described in detail. In this disclosure, the expression “at least one of a, b, or c” can refer to “a,” “b,” “c,” “a and b,” “a and c,” “b and c,” “all of a, b, and c,” or variations thereof.

[0023] The terms used in this disclosure are selected from widely used, common terms, taking into account the functions of the disclosure. However, these terms may vary depending on the intentions of those skilled in the art, precedents, the emergence of new technologies, etc. Furthermore, in certain cases, terms may be arbitrarily selected by the applicant, in which case their meanings will be described in detail in the relevant description. Therefore, the terms used in this disclosure should not be defined simply as names, but rather based on the meanings of the terms and the overall content of the disclosure.

[0024] Singular expressions may include plural expressions unless the context clearly indicates otherwise. Terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the art described herein. Furthermore, terms containing ordinal numbers, such as "first" or "second," used herein may be used to describe various components, but such components should not be limited by such terms. Such terms are used solely to distinguish one component from another.

[0025] When a part of the specification is said to "include" a component, unless otherwise specifically stated, this does not exclude other components but rather implies the inclusion of other components. Furthermore, terms such as "part" and "module" used in the specification refer to a unit that processes at least one function or operation, which may be implemented in hardware, software, or a combination of hardware and software.

[0026] All functions or operations described in this document may be performed by a single processor or a combination of processors. A single processor or a combination of processors is a circuitry that performs processing, and may include circuitry such as an Application Processor (AP), a Communication Processor (CP), a Graphical Processing Unit (GPU), a Neural Processing Unit (NPU), a Microprocessor Unit (MPU), a System on Chip (SoC), or an Integrated Chip (IC).

[0027] The artificial intelligence-related functions according to the present disclosure are operated via a processor and memory. The processor may be comprised of one or more processors. In this case, one or more processors may be a general-purpose processor such as a CPU, an AP, a Digital Signal Processor (DSP), a graphics-only processor such as a GPU or a Vision Processing Unit (VPU), or an artificial intelligence-only processor such as an NPU. One or more processors control the processing of input data according to predefined operating rules or artificial intelligence models stored in memory. Alternatively, if one or more processors are artificial intelligence-only processors, the artificial intelligence-only processor may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

[0028] The predefined operation rules or artificial intelligence models are characterized by being created through learning. Here, being created through learning means that the basic artificial intelligence model is trained using a learning algorithm using a plurality of learning data, thereby creating a predefined operation rules or artificial intelligence model set to perform a desired characteristic (or purpose). This learning may be performed on the device itself on which the artificial intelligence according to the present disclosure is performed, or may be performed through a separate server and / or system. Examples of the learning algorithm include, but are not limited to, supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning.

[0029] An artificial intelligence model may be composed of multiple neural network layers. Each of the multiple neural network layers has multiple weight values, and performs neural network operations through operations between the operation results of the previous layer and the multiple weights. The multiple weights of the multiple neural network layers may be optimized based on the learning results of the artificial intelligence model. For example, the multiple weights may be updated so that the loss value or cost value obtained from the artificial intelligence model is reduced or minimized during the learning process. The artificial neural network may include a deep neural network (DNN), and examples thereof include, but are not limited to, a convolutional neural network (CNN), a deep neural network (DNN), a recurrent neural network (RNN), a restricted boltzmann machine (RBM), a deep belief network (DBN), a bidirectional recurrent deep neural network (BRDNN), or deep Q-networks.

[0030] Embodiments of the present disclosure relate to a method for replacing a string representing personal information with another string. Before describing specific embodiments, the meanings of terms frequently used in the present disclosure are defined.

[0031] In this disclosure, "personal information" may refer to any type of information that identifies an individual. For example, personal information may include direct identifiers that can directly identify an individual, such as name, social security number, date of birth, contact information, and address. Furthermore, personal information may include indirect identifiers that cannot directly identify an individual, but can be combined with other information to identify an individual, such as financial information, health information, employment information, ethnicity, gender, and religion.

[0032] "Embedding" can refer to a vector that represents the characteristics of unstructured data such as text. For example, certain unstructured data can be converted into a vector that includes the characteristics of the unstructured data, and the converted vector can correspond to an embedding. Therefore, an embedding corresponding to a text can include the characteristics of the text (e.g., semantic features, grammatical features, contextual features, etc.). Embeddings can be created through embedding transformation. Terms such as "embedding vector" or "feature vector" may also be used instead of "embedding."

[0033] "Embedding transformation" can refer to the process of converting unstructured data, such as text, into an embedding. For example, it can refer to the process of converting text into a numerical vector, making it accessible for electronic devices to process.

[0034] An "embedding model" can be a model that performs "embedding transformation." For example, an embedding model can extract features, such as text, and convert them into embeddings, enabling electronic devices to understand and process the text.

[0035] FIG. 1 is a drawing for exemplarily explaining an operation of an electronic device according to one embodiment of the present disclosure to replace a string expressing personal information with another string.

[0036] According to one embodiment of the present disclosure, an electronic device can use a personal information replacement model (100) to replace a string representing personal information with another string. The personal information replacement model (100) replaces personal information with another string having a similar format, category, or meaning, thereby protecting personal information while maintaining the format, category, and meaning of the existing personal information in the text as much as possible.

[0037] Text data containing personal information can be used for data analysis or to train artificial intelligence models. However, when data containing personal information is used for data analysis or artificial intelligence model training, it can lead to personal information leakage. A personal information replacement model (100) can protect personal information and prevent personal information leakage by replacing strings representing personal information with other strings.

[0038] Referring to FIG. 1, the personal information replacement model (100) can provide replacement text (20) such as 'Lee Young-hee uses Mirae Bank account number 583-742-9168 and credit card number 4321-8765-2109-6543' instead of the original text (10) such as 'Kim Cheol-su uses the best bank account number 728-164-3952 and credit card number 9876-5432-1098-7654'.

[0039] For example, the original text (10) may include strings expressing personal information such as the name 'Kim Cheol-su', the bank name 'Choi Best Bank' as financial information, the account number '728-164-3952' as financial information, and the credit card number '9876-5432-1098-7654' as financial information, but these are only examples and are not limited to the examples described above.

[0040] According to one embodiment of the present disclosure, a personal information replacement model (100) can provide a string representing personal information, such as a name, bank name, account number, or credit card number, by replacing it with another string while maintaining the format or meaning of the string as much as possible. For example, the personal information replacement model (100) can provide a string by replacing the string 'Kim Cheol-su' with the string 'Lee Young-hee'. The personal information replacement model (100) can provide a replacement for an original string, such as 'Kim Cheol-su', with a replacement string, such as 'Lee Young-hee', thereby maintaining the format or meaning of the name and providing a replacement with another name, such as 'Lee Young-hee'. The personal information replacement model (100) can provide 'Lee Young-hee' in the replacement text (20) to protect the personal information 'Kim Cheol-su' in the original text (10).

[0041] Additionally, for example, the personal information replacement model (100) can provide a replacement string such as '583-742-9168' for an original string such as '728-164-3952', thereby maintaining the format and meaning of the account number and providing a replacement string such as '583-742-9168'. The personal information replacement model (100) can provide '583-742-9168' in the replacement text (20) to protect the personal information '728-164-3952' in the original text (10).

[0042] According to one embodiment of the present disclosure, unlike general text consisting of a series of characters, numeric data does not have a unique linguistic meaning per word, and each number itself may not imply a separate meaning. The personal information replacement model (100) may use a sub-model that replaces numeric strings representing personal information, such as account numbers and credit card numbers, with other strings while preserving the format and meaning of the strings.

[0043] According to one embodiment of the present disclosure, the personal information replacement model (100) can replace part or all of a string expressing personal information with a replacement string. The personal information replacement model (100) can provide replacement text (20) in which a string expressing personal information is replaced with another string while maintaining the format or meaning of the personal information expressed in the original text (10).

[0044] According to one embodiment of the present disclosure, personal information may include personally identifiable information (PII) that can identify an individual, such as name, date of birth, contact information (e.g., phone number, email address, etc.), and financial information (e.g., account number, card number, etc.). Furthermore, for example, personal information may include health information (e.g., medical history, physical disability, mental disability, etc.), employment information (e.g., occupation, etc.), and sensitive information (e.g., personal information that may infringe upon an individual's privacy, such as thoughts, beliefs, political opinions, etc.).

[0045] According to one embodiment of the present disclosure, a string may be a data structure composed of a sequence of characters. A string may be a basic unit that constitutes text. Personal information may be expressed as a string composed of at least one of letters, numbers, or special characters. For example, personal information may be expressed as letters, such as a name (e.g., "Kim Cheol-su"). Alternatively, personal information may be expressed as numbers, such as a phone number (e.g., "01012345678"), or a combination of numbers and special characters (e.g., "+82)10-1234-5678"). However, these are merely examples and are not limited to the examples described above.

[0046] According to one embodiment of the present disclosure, an electronic device may be a device capable of performing a personal information replacement task and displaying and providing text and / or personal information replacement results. For example, the electronic device may be implemented as various types and forms of electronic devices that include a display. Electronic devices may include, but are not limited to, devices capable of displaying visual information through a display, such as a smart TV, a smartphone, a tablet PC, or a laptop PC.

[0047] According to one embodiment of the present disclosure, an electronic device may be a device that performs a personal information replacement task and provides text and / or the personal information replacement result to a user device. For example, the electronic device and user device of the present disclosure may be implemented in a server-client device configuration.

[0048] The specific operations of an electronic device using a personal information replacement model (100) to provide a string representing personal information by replacing it with another string will be described in more detail through the drawings and descriptions thereof described below.

[0049] FIG. 2 is a flowchart illustrating an operation of an electronic device according to one embodiment of the present disclosure to replace a string representing personal information with another string.

[0050] According to one embodiment of the present disclosure, an electronic device can perform a personal information replacement operation, which replaces a string expressing personal information with another string that can replace the string expressing personal information while maintaining the format or meaning of the string expressing personal information.

[0051] In operation S210, the electronic device may obtain text containing a character string representing personal information. The electronic device may obtain an original text. The original text may contain a character string representing the personal information. The original text may be used as input data for a personal information replacement operation performed by the electronic device.

[0052] According to one embodiment of the present disclosure, an electronic device can obtain original text based on user input. For example, the electronic device can obtain the original text by having the user directly input text via a keyboard or voice recognition.

[0053] Additionally, according to one embodiment of the present disclosure, an electronic device can obtain original text through communication with a server. The electronic device can obtain the original text based on text data collected from the server. For example, the electronic device can download a large dataset and use it as the original text.

[0054] The specific operation of the electronic device to obtain a text containing a character string representing personal information is described in the description of FIG. 3.

[0055] In operation S220, the electronic device may convert a string into a reference word embedding using a word embedding model. The electronic device may analyze the strings constituting the original text to extract structural meaning, such as words or sentences, from the original text. The electronic device may decompose the original text into semantic units to process the text. For example, the electronic device may decompose the original text into semantic units or structural units to extract words. In the present disclosure, a word may represent a semantic unit that has meaning in the text, or a structural unit, such as a stop word. In addition to letters, numbers and special characters may also have meaning depending on the context within the text. Therefore, in the present disclosure, a word may be a concept including letters, numbers, or special characters.

[0056] According to one embodiment of the present disclosure, an electronic device can convert words obtained from an original text, which represent personal information, into a reference word embedding. A word embedding may refer to an embedding that represents the characteristics of a text at the word level. The reference word embedding may represent a word embedding in which words obtained from the original text are converted into embeddings. The electronic device can use a word embedding model to map words to embeddings in a multidimensional space. For example, the electronic device can set the dimensionality of the embedding space to 50 to 300 dimensions, but the number of dimensions of the embedding space is not limited thereto. The electronic device can use the word embedding model to represent words representing personal information as numeric vectors in a vector space of preset dimensions.

[0057] According to one embodiment of the present disclosure, an electronic device can obtain a word embedding model. For example, the electronic device can load the word embedding model through communication with a server. According to one embodiment of the present disclosure, the electronic device can obtain the word embedding model externally and perform fine-tuning to adapt it to the task of converting personal information words.

[0058] According to one embodiment of the present disclosure, an electronic device can convert words into word embeddings using a word embedding model. The word embedding model may be a model that performs word embedding conversion. For example, the electronic device may use word embedding models such as Word2Vec or GloVe, but these are merely examples and are not limited to the examples described above.

[0059] According to one embodiment of the present disclosure, a word embedding model may be trained using a vocabulary containing a plurality of words as training data. The vocabulary may, for example, represent a pool of words used as training data for the word embedding model.

[0060] According to one embodiment of the present disclosure, a word embedding model may be a model specialized in converting personal information into an embedding, and is learned using a string representing personal information or a set of words representing personal information as learning data.

[0061] The specific operation of the electronic device converting a string into a reference word embedding using the word embedding model is described in the description of FIGS. 4 and 5.

[0062] In operation S230, the electronic device can identify the distance between candidate word embeddings converted from words used in training the word embedding model and the reference word embedding. The electronic device can identify the distance between each candidate word embedding and the reference word embedding. The candidate word embedding may represent the result of embedding conversion of each word used in training the word embedding model.

[0063] According to one embodiment of the present disclosure, an electronic device can calculate the distance between a candidate word embedding and a reference word embedding. In the embedding space, the distance between the candidate word embedding and the reference word embedding may indicate the similarity between the respective word embeddings. For example, the shorter the distance between the candidate word embedding and the reference word embedding, the more similar the meaning of the word corresponding to the candidate word embedding and the word corresponding to the reference word embedding may be.

[0064] According to one embodiment of the present disclosure, when an original text includes multiple words expressing different personal information, the electronic device can calculate the distance between candidate word embeddings based on each of the reference word embeddings. That is, the electronic device can calculate the distance from each of the multiple reference word embeddings to the candidate word embedding corresponding to the word used as training data for the word embedding model.

[0065] For example, if the original text is "Kim Cheol-su lives in Seoul," the original text contains two words representing personal information: the name "Kim Cheol-su" and the place of residence "Seoul." Accordingly, the electronic device can calculate the distance between the first reference word embedding vector corresponding to "Kim Cheol-su" and each candidate word embedding, and the second reference word embedding vector corresponding to "Seoul" and each candidate word embedding.

[0066] The specific operation by which the electronic device identifies the distance between each candidate word embedding and the reference word embedding is described in the description of FIGS. 6A to 7.

[0067] In operation S240, the electronic device may select replacement candidates for a string from among the candidate word embeddings based on the identified distance. The electronic device may compare the distance between each of the plurality of candidate word embeddings and the reference word embedding. By calculating the distance between each of the plurality of candidate word embeddings and the reference word embedding, the electronic device may obtain a candidate word embedding that is relatively close to the reference word embedding.

[0068] According to one embodiment of the present disclosure, an electronic device may list the distances between each of a plurality of candidate word embeddings and a reference word embedding in order of distance. Based on the identified distances, the electronic device may select a reference number of candidate word embeddings from among the candidate word embeddings in order of decreasing distances between them and the reference word embedding.

[0069] According to one embodiment of the present disclosure, an electronic device may preset a reference number. The reference number may be the number of replacement candidates for replacing a string representing personal information. For example, if the reference number is set to 3, the electronic device may select three candidate word embeddings in descending order of distance from the reference word embedding.

[0070] According to one embodiment of the present disclosure, an electronic device can select a reference number of candidate word embeddings and use the selected candidate word embeddings as replacement candidates for a string representing personal information.

[0071] According to one embodiment of the present disclosure, when an original text includes multiple words expressing different personal information, an electronic device may select replacement candidates based on each of the reference word embeddings. The electronic device may calculate the distance from each of the multiple reference word embeddings to the candidate word embedding, and select a reference number of candidate word embeddings from among the candidate word embeddings in descending order of the distance between each of the reference word embeddings.

[0072] For example, if the original text is 'Kim Cheol-su lives in Seoul,' the electronic device can calculate the distance between the first reference word embedding vector corresponding to 'Kim Cheol-su' and each candidate word embedding. The electronic device can select three preset candidate word embeddings in order of the smallest distance from the first reference word embedding vector among the candidate word embeddings. In addition, the electronic device can calculate the distance between the second reference word embedding vector corresponding to 'Seoul' and each candidate word embedding. The electronic device can select three preset candidate word embeddings in order of the smallest distance from the second reference word embedding vector among the candidate word embeddings.

[0073] According to one embodiment of the present disclosure, an electronic device may select a candidate word embedding of a reference order from among candidate word embeddings. The candidate word embedding of the reference order may be a candidate word embedding having the greatest distance from a reference word embedding vector among candidate word embeddings selected by a reference number. In other words, when candidate word embeddings are listed in descending order of distance from the reference word embedding vector, the candidate word embedding may be a candidate word embedding corresponding to the reference order. The reference order value may be the same as the reference number value. For example, if the reference number is 3, the reference order may also be the third.

[0074] According to one embodiment of the present disclosure, an electronic device may list a plurality of candidate word embedding vectors in descending order of distance from a first reference word embedding vector. The electronic device may select a candidate word embedding having a reference order among the candidate word embeddings. For example, a third candidate word embedding having a reference order may be a candidate word embedding having a longest distance (e.g., a first distance) from the first reference word embedding vector among the candidate word embeddings selected based on the first reference word embedding vector. In other words, when the candidate word embeddings are listed in descending order of distance from the first reference word embedding vector, the electronic device may be a candidate word embedding having a third shortest distance.

[0075] In addition, the electronic device may list the candidate word embeddings in descending order of distance from the second reference word embedding vector. The electronic device may select a candidate word embedding in the reference order from among the candidate word embeddings. For example, the third candidate word embedding in the reference order may be the candidate word embedding with the longest distance (e.g., the second distance) from the second reference word embedding vector among the candidate word embeddings selected based on the second reference word embedding vector. In other words, when the candidate word embeddings are listed in descending order of distance from the second reference word embedding vector, it may be the candidate word embedding with the third shortest distance.

[0076] However, when an electronic device selects alternative candidates based on each of the reference word embeddings, the privacy protection strength for different personal information may vary. For example, a first distance identified based on a first reference word embedding vector and a second distance identified based on a second reference word embedding vector may differ. Due to the difference between the first and second distances, the protection strength for the first personal information and the protection strength for the second personal information may vary. For example, as the distance between the reference word embedding vector and the candidate word embedding becomes shorter, the protection strength may decrease. As the distance between the reference word embedding vector and the candidate word embedding becomes longer, the protection strength may increase.

[0077] Accordingly, according to one embodiment of the present disclosure, an electronic device may compare a first distance identified based on a first reference word embedding vector with a second distance identified based on a second reference word embedding vector, and select a larger distance as the reference distance. The electronic device may select candidate word embeddings using the reference distance.

[0078] For example, an electronic device may select all candidate word embedding vectors within a reference distance from a first reference word embedding vector. Furthermore, the electronic device may select all candidate word embedding vectors within a reference distance from a second reference word embedding vector. The electronic device may use the selected candidate word embeddings as alternative candidates for strings representing personal information.

[0079] In operation S250, the electronic device may replace a string representing personal information with a string corresponding to one of the replacement candidates.

[0080] According to one embodiment of the present disclosure, an electronic device can select one of the alternative candidates. For example, the electronic device can randomly select one of the alternative candidates. The electronic device can obtain a word corresponding to the selected alternative candidate. The electronic device can replace a string representing personal information with the word corresponding to the selected alternative candidate. The electronic device can provide alternative text in which the string representing the personal information is replaced with another string.

[0081] FIG. 3 is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to obtain a text including a string representing personal information.

[0082] According to one embodiment of the present disclosure, an electronic device can obtain an original text including a character string representing personal information. Referring to FIG. 3 , the electronic device can obtain an original text (310) such as "Hello. This is Hong Gil-dong." The electronic device can identify a character string representing personal information in the original text (310). For example, the electronic device can identify "Hong Gil-dong" (311) in "Hello. This is Hong Gil-dong" as a character string representing personal information.

[0083] According to one embodiment of the present disclosure, an electronic device can identify a character string representing personal information based on an index (320) corresponding to a text (330). For example, the electronic device can obtain an index (320) of a character string representing personal information using metadata of the text (330). The electronic device can extract personal information using the index (320) of the character string.

[0084] According to one embodiment of the present disclosure, an electronic device can obtain an index of a string representing personal information from metadata of an original text. The metadata of the original text can include start index information and end index information of the string representing the personal information.

[0085] For example, an electronic device can obtain an index (320) of a string representing the name information 'Hong Gil-dong' from the metadata of the text 'Hello. This is Hong Gil-dong.' (330). The electronic device can obtain start index information '[7]' (321) corresponding to 'Hong' (331) of the string 'Hong Gil-dong'. The electronic device can obtain end index information '[9]' (323) corresponding to 'Dong' (333) of the string 'Hong Gil-dong'.

[0086] According to one embodiment of the present disclosure, an electronic device can extract a character string representing personal information from an original text based on start index information and end index information. For example, the electronic device can extract the character string "Hong Gil-dong" representing personal information from the text "Hello. This is Hong Gil-dong." (330) using start index information "[7]" (321) and end index information "[9]" (323). The character string index may have been described based on starting from 0, including spaces and special characters. However, spaces and special characters may be excluded, or the index start value may be set to a value other than 0. Meanwhile, the character string indexing criteria may vary depending on the language used or national policies.

[0087] Additionally, according to one embodiment of the present disclosure, an electronic device can extract a string representing personal information using a language model. For example, the electronic device can use a language model to detect personal information and obtain a string representing the personal information.

[0088] FIG. 4 is a diagram for explaining an operation of an electronic device according to one embodiment of the present disclosure to convert a word into a word embedding.

[0089] According to one embodiment of the present disclosure, an electronic device can convert words into word embeddings using a word embedding model. The word embedding model may be trained using a vocabulary containing multiple words as training data.

[0090] Referring to Fig. 4, the word embedding model may be learned using a string representing personal information or a set of words representing personal information as learning data.

[0091] For example, the word embedding model may be learned using addresses such as the word embedding vector Seoul (411) corresponding to the word 'Seoul', the word embedding vector Tokyo (412) corresponding to the word 'Tokyo', the word embedding vector Beijing (413) corresponding to the word 'Beijing', the word embedding vector London (414) corresponding to the word 'London', the word embedding vector Berlin (415) corresponding to the word 'Berlin', the word embedding vector Sydney (416) corresponding to the word 'Sydney', and the word embedding vector New York (417) corresponding to the word 'New York' as learning data.

[0092] In addition, for example, the word embedding model may be learned using disease names such as the word embedding vector obesity (421) corresponding to the word 'obesity', the word embedding vector diabetes (422) corresponding to the word 'diabetes', the word embedding vector hypertension (434) corresponding to the word 'hypertension', the word embedding vector hypotension (424) corresponding to the word 'hypotension', the word embedding vector stroke (425) corresponding to the word 'stroke', the word embedding vector asthma (426) corresponding to the word 'asthma', the word embedding vector pneumonia (427) corresponding to the word 'pneumonia', the word embedding vector depression (428) corresponding to the word 'depression', and the word embedding vector anxiety disorder (429) corresponding to the word 'anxiety disorder' as learning data, but this is only an example and is not limited to the examples described above.

[0093] According to one embodiment of the present disclosure, for a word (430) for which a word embedding model has completed learning, the word embedding model can convert the word (430) into an embedding vector (440). The word embedding model can generate a word embedding vector (440) representing the meaning of the word (430) by mapping the word to an embedding space.

[0094] According to one embodiment of the present disclosure, mapping information can be acquired as a result of training a word embedding model. The word embedding model can generate an embedding vector (440) using the mapping information. The mapping information may refer to information indicating how words (430) and embedding vectors (440) are connected. The mapping information may include a connection relationship between each word in a vocabulary and the embedding vector assigned to that word.

[0095] For example, a word embedding model can generate a word embedding vector Seoul (411) corresponding to the word 'Seoul' in the embedding space. The word embedding model can use mapping information to convert the word 'Seoul' into vector coordinates such as '[0.12, 0.56, ..., 0.78]'. Similarly, the word embedding model can use mapping information to generate word embedding vectors corresponding to each word.

[0096] According to one embodiment of the present disclosure, an electronic device can obtain mapping information between words and their corresponding embedding vectors. For example, the electronic device can obtain mapping information for a word embedding model while obtaining a word embedding model. Furthermore, for example, the electronic device can store mapping information for the word embedding model in memory.

[0097] According to one embodiment of the present disclosure, for words for which the word embedding model has not completed training, the word embedding model may not be able to convert the word into a word embedding. For example, the word embedding model may output "Out-Of-Vocabulary (OOV)" for words not included in the predefined vocabulary.

[0098] Furthermore, according to one embodiment of the present disclosure, for words for which the word embedding model has not completed training, the word embedding model can perform embedding transformation by dividing the word into smaller units (e.g., syllables, prefixes / suffixes, etc.). For example, the word embedding model can perform embedding transformation on a subword basis for compound words not included in a predefined vocabulary.

[0099] FIG. 5 is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to convert a string representing personal information into a reference word embedding using a word embedding model.

[0100] According to one embodiment of the present disclosure, an electronic device can convert a string representing personal information into a reference word embedding. The electronic device can convert words obtained from an original text representing the personal information into a reference word embedding. The reference word embedding may represent a word embedding in which words obtained from the original text are converted into an embedding. The reference word embedding may also be referred to as a reference embedding.

[0101] According to one embodiment of the present disclosure, an electronic device can convert a string representing personal information into a reference word embedding vector using a word embedding model. When an original text includes multiple words representing different pieces of personal information, the electronic device can generate reference word embeddings for each of the multiple pieces of personal information. The electronic device can convert a first string representing first personal information into a first reference word embedding vector. The electronic device can convert a second string representing second personal information into a second reference word embedding vector. That is, based on each piece of personal information, the electronic device can convert the string representing each piece of personal information into a reference word embedding corresponding to the string.

[0102] For example, an electronic device can obtain original text (510), such as "I live in Seoul. I regularly visit a hospital due to obesity." The electronic device can identify a first character string "Seoul" (511), which represents first personal information corresponding to a place of residence, from the original text (510). The electronic device can convert the first character string "Seoul" (511) into a first reference word embedding vector "Seoul" (521), which corresponds to the word "Seoul" in the embedding space.

[0103] Additionally, for example, the electronic device can identify a second string of characters, 'obesity' (512), representing second personal information corresponding to a disease name from the original text (510). The second string of characters, 'obesity' (512), can be converted into a second reference word embedding vector, 'obesity' (522), corresponding to the word 'obesity' in the embedding space.

[0104] That is, the electronic device can identify multiple words representing different personal information and generate a reference word embedding vector for each word.

[0105] FIG. 6A is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to identify an angle between each candidate word embedding and a reference word embedding.

[0106] According to one embodiment of the present disclosure, an electronic device can identify the distance between each candidate word embedding and a reference word embedding. When identifying the distance, the electronic device can also consider the direction. The electronic device can calculate the angular difference between the reference word embedding and each candidate word embedding.

[0107] According to one embodiment of the present disclosure, an electronic device can calculate a cosine similarity between each candidate word embedding and a reference word embedding. The electronic device can measure the directional similarity between the reference word embedding and each candidate word embedding.

[0108]

[0109] Here,

[0110] is the dot product of the two embedding vectors,

[0111] is the magnitude of each embedding vector,

[0112] is the cosine of the angle formed by the two embedding vectors.

[0113] According to one embodiment of the present disclosure, The closer the value is to 1, the more similar the meanings of the words corresponding to the reference word embedding and the words corresponding to the candidate word embedding may be. The closer the value is to 0, the less clear the correlation or semantic connection between the word corresponding to the reference word embedding and the word corresponding to the candidate word embedding may be. The closer the value is to -1, the more contrast there may be between the meanings of the words corresponding to the reference word embedding and the words corresponding to the candidate word embedding.

[0114] Referring to FIG. 6A, the electronic device can identify the distance between the first reference word embedding vector 'Seoul' (521) corresponding to the word 'Seoul' and all candidate word embeddings. For example, referring to 610a of FIG. 6A, the electronic device can identify the angle between the first reference word embedding vector 'Seoul' (521) and the word embedding vector Tokyo (412). Referring to 620a of FIG. 6, the electronic device calculates the angle between the first reference word embedding vector 'Seoul' (521) and the word embedding vector Beijing (413). can be calculated. Similarly, the electronic device can calculate the distance between the first reference word embedding vector 'Seoul' (521) and the word embedding vector London (414), the word embedding vector Berlin (415), the word embedding vector Sydney (416), and the word embedding vector New York (417) corresponding to the word set (vocabulary).

[0115] FIG. 6b is a diagram illustrating an operation of an electronic device according to an embodiment of the present disclosure to identify a length between each candidate word embedding and a reference word embedding.

[0116] According to one embodiment of the present disclosure, an electronic device can identify the distance between each candidate word embedding and a reference word embedding. When identifying the distance, the electronic device can consider both the size and direction between the embeddings. The electronic device can calculate the distance between the reference word embedding and each candidate word embedding.

[0117] According to one embodiment of the present disclosure, an electronic device can calculate a Euclidean distance between a reference word embedding and each candidate word embedding. The electronic device can quantify the similarity or difference between a word corresponding to the reference word embedding and a word corresponding to each candidate word embedding.

[0118]

[0119]

[0120] Here,

[0121] is each word embedding vector,

[0122] n is the dimension of the vector,

[0123] It is the component in the i-th dimension of each embedding vector.

[0124] According to one embodiment of the present disclosure, the Euclidean distance between a reference word embedding and a candidate word embedding can be used as a relative comparison index. For example, the shorter the Euclidean distance, the more similar the meanings of words corresponding to the reference word embedding and words corresponding to the candidate word embedding may be. The longer the Euclidean distance, the more dissimilar the meanings of words corresponding to the reference word embedding and words corresponding to the candidate word embedding may be. For example, as the dimension of the embedding space and the dimension of the embedding vector increase, the distance between the embedding vectors may increase. As the dimension of the embedding space and the dimension of the embedding vector decrease, the distance between the embedding vectors may decrease.

[0125] Referring to FIG. 6B, the electronic device can identify the length between the first reference word embedding vector 'Seoul' (521) corresponding to the word 'Seoul' and all candidate word embeddings. For example, referring to 610b of FIG. 6B, the electronic device can identify the length between the first reference word embedding vector 'Seoul' (521) and the word embedding vector Tokyo (412). Referring to 620b of FIG. 6, the electronic device calculates the length between the first reference word embedding vector 'Seoul' (521) and the word embedding vector Beijing (413). can be calculated. Similarly, the electronic device can calculate the length between the first reference word embedding vector 'Seoul' (521) and the word embedding vector London (414), the word embedding vector Berlin (415), the word embedding vector Sydney (416), and the word embedding vector New York (417) corresponding to the word set (vocabulary).

[0126] According to one embodiment of the present disclosure, an electronic device may use a method for calculating cosine similarity or Euclidean distance to identify the distance between each candidate word embedding and a reference word embedding. However, the method for identifying the distance between each candidate word embedding and the reference word embedding is not limited to the examples described above, and other similarity calculation methods or distance calculation methods may also be applied.

[0127] FIG. 7 is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to identify a distance between each candidate word embedding and a reference word embedding.

[0128] According to one embodiment of the present disclosure, when there are a plurality of reference word embeddings, the electronic device can identify the distance to all candidate word embeddings based on each reference word embedding.

[0129] According to one embodiment of the present disclosure, an electronic device may calculate a distance between a first reference word embedding vector and each candidate word embedding. Additionally, the electronic device may calculate a distance between a second reference word embedding vector and each candidate word embedding.

[0130] According to one embodiment of the present disclosure, an electronic device can calculate a distance between all candidate word embeddings and a reference word embedding, regardless of the type of word corresponding to the word embedding vector.

[0131] For example, based on the first reference word embedding vector 'Seoul' (521), the electronic device can identify the distance to the word embedding corresponding to the address, such as the word embedding vector Tokyo (412), the word embedding vector Beijing (413), etc. In addition, the electronic device can identify the distance to the word embedding corresponding to the disease name, such as the word embedding vector diabetes (422), the word embedding vector hypertension (434), and the word embedding vector hypotension (424).

[0132] Referring to Figure 7, based on the first reference word embedding vector 'Seoul' (521), the electronic device measures the distance from the word embedding vector Sydney (416). can be calculated. In addition, based on the first reference word embedding vector 'Seoul' (521), the electronic device can calculate the distance to the word embedding vector pneumonia (427). can be calculated, but this is only an example and is not limited to the examples mentioned above.

[0133] In addition, for example, based on the second reference word embedding vector 'obesity' (522), the electronic device can identify the distance to the word embedding corresponding to the disease name, such as the word embedding vector diabetes (422), the word embedding vector hypertension (434), and the word embedding vector hypotension (424). In addition, the electronic device can identify the distance to the word embedding corresponding to the address, such as the word embedding vector Tokyo (412), and the word embedding vector Beijing (413).

[0134] Referring to Figure 7, based on the second reference word embedding vector 'obesity' (522), the electronic device calculates the distance to the word embedding vector diabetes. can be calculated. Similarly, the electronic device can calculate the distance to other candidate word embedding vectors, regardless of the type of the candidate word embedding vector, based on the second reference word embedding vector 'obesity' (522).

[0135] According to one embodiment of the present disclosure, an electronic device can calculate the distance between a word embedding vector and each candidate word embedding vector and store the calculated distance in a memory of the electronic device. If distance information is stored in the memory, the electronic device can calculate the distance between a reference word embedding and a candidate word embedding using the stored distance information. By using the distance information stored in the memory, the electronic device can avoid repeatedly calculating the distance between each word embedding. By using the distance information stored in the memory, the electronic device can reduce the amount of computational effort of the electronic device. By using the distance information stored in the memory, the electronic device can increase processing speed and improve performance of the electronic device.

[0136] FIG. 8A is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to select a set number of candidate word embeddings as replacement candidates for a string.

[0137] According to one embodiment of the present disclosure, the electronic device can select alternative candidates for a string representing personal information among candidate word embeddings based on an identified distance between a reference word embedding and each candidate word embedding.

[0138] According to one embodiment of the present disclosure, an electronic device can set a reference number. According to one embodiment of the present disclosure, the reference number for replacing a string representing personal information may be preset. Furthermore, according to one embodiment of the present disclosure, the electronic device can receive a reference number input from a user. According to one embodiment of the present disclosure, the reference number is pre-stored in the memory of the electronic device, and the electronic device can utilize the pre-stored number of replacement candidates.

[0139] According to one embodiment of the present disclosure, an electronic device may compare distances between a reference word embedding and each candidate word embedding. The electronic device may list each candidate word embedding in descending order of distance between the reference word embedding and each candidate word embedding. The electronic device may select each candidate word embedding as a replacement candidate from among a plurality of candidate word embeddings in descending order of distance from the reference word embedding. The electronic device may select a reference number of candidate word embeddings as replacement candidates from among the candidate word embeddings in descending order of identified distance.

[0140] Referring to FIG. 8A, the electronic device may set the number of replacement candidates to 3. The electronic device may compare the distances between all candidate word embeddings based on the first reference word embedding vector 'Seoul' (521). The electronic device may select three candidate word embeddings having the closest distance to the first reference word embedding vector 'Seoul' (521). For example, the electronic device may select the word embedding vector Tokyo (412), the word embedding vector Berlin (415), and the word embedding vector New York (417) having a small distance from Seoul (521). The candidate word embeddings selected by the electronic device may be replacement candidates (810) for the word Seoul corresponding to the first reference word embedding vector 'Seoul' (521).

[0141] FIG. 8b is a diagram for explaining an operation of an electronic device according to one embodiment of the present disclosure to select a set number of candidate word embeddings as replacement candidates for each reference word embedding.

[0142] According to one embodiment of the present disclosure, the electronic device may select, for each of a plurality of reference word embeddings, alternative candidates for a string representing personal information from among the candidate word embeddings based on an identified distance between the reference word embedding and each of the candidate word embeddings.

[0143] Referring to FIG. 8B, the electronic device can set the number of alternative candidates to 3. The electronic device can compare the distances between all candidate word embeddings based on the first reference word embedding vector 'Seoul' (521). The electronic device can select the three candidate word embeddings that have the closest distances to the first reference word embedding vector 'Seoul' (521).

[0144] Also, referring to FIG. 8B, the electronic device can compare the distances between all candidate word embeddings based on the second reference word embedding vector 'obesity' (522). The electronic device can select three candidate word embeddings having the closest distances to the second reference word embedding vector 'obesity' (522). For example, the electronic device can select the word embedding vector diabetes (422), the word embedding vector depression (428), and the word embedding vector anxiety disorder (429) having small distances to obesity (522). The candidate word embeddings selected by the electronic device for the second reference word embedding vector 'obesity' (522) can be replacement candidates (820) for the word obesity corresponding to the second reference word embedding vector 'obesity' (522).

[0145] FIG. 9A is a diagram for explaining an operation of an electronic device according to one embodiment of the present disclosure to select replacement candidates based on a reference distance.

[0146] According to one embodiment of the present disclosure, an electronic device can determine a reference distance for a plurality of reference word embedding vectors. Based on the reference distance, the electronic device can select replacement candidates based on each reference word embedding vector.

[0147] According to one embodiment of the present disclosure, an electronic device can identify a reference distance. The electronic device can select a candidate word embedding with a reference order from among candidate word embeddings. For example, the electronic device can identify a first distance corresponding to a predetermined order in descending order of identified distances between a first reference word embedding vector and each candidate word embedding. The electronic device can identify a second distance corresponding to a predetermined order in descending order of identified distances between a second reference word embedding vector and each candidate word embedding. The electronic device can compare the first distance and the second distance and select a larger distance as the reference distance.

[0148] For example, referring to FIG. 9A, the electronic device can select a set number of candidate word embedding vectors in order of proximity based on the first reference word embedding vector 'Seoul' (521). For example, when the set number of criteria is 3, the electronic device can select the word embedding vector Tokyo (412) which is the closest in distance based on the first reference word embedding vector 'Seoul' (521), the word embedding vector New York (417) which is the second closest in distance based on the first reference word embedding vector 'Seoul' (521), and the word embedding vector Berlin (415) which is the third closest in distance based on the first reference word embedding vector 'Seoul' (521). At this time, since the set number is 3, the distance with the candidate word embedding vector which is the third closest in distance to the first reference word embedding vector can be identified as the first distance. In other words, the third candidate word embedding, which is the reference order, may be the word embedding vector Berlin (415) that has the longest distance from the first reference word embedding vector among the word embedding vector Tokyo (412), the word embedding vector New York (417), and the word embedding vector Berlin (415). Accordingly, the distance between the first reference word embedding vector 'Seoul' (521) and the word embedding vector Berlin (415) is identified as the first distance.

[0149] In addition, for example, the electronic device can select a set number of candidate word embedding vectors in order of proximity based on the second reference word embedding vector 'obesity' (522). For example, when the set number of criteria is 3, the electronic device can select the word embedding vector diabetes (422) having the closest distance based on the second reference word embedding vector 'obesity' (522), the word embedding vector depression (428) having the second closest distance based on the second reference word embedding vector 'obesity' (522), and the word embedding vector anxiety disorder (429) having the third closest distance based on the second reference word embedding vector 'obesity' (522). In this case, since the set number is 3, the distance to the candidate word embedding vector having the third closest distance to the second reference word embedding vector can be identified as the second distance. In other words, the third candidate word embedding, which is the reference order, may be the word embedding vector anxiety disorder (429) that has the greatest distance from the second reference word embedding vector among the word embedding vector diabetes (422), word embedding vector depression (428), and word embedding vector anxiety disorder (429). Accordingly, the distance between the second reference word embedding vector 'obesity' (522) and the word embedding vector anxiety disorder (429) is identified as the second distance.

[0150] According to one embodiment of the present disclosure, when the first and second distances differ, the degree of privacy protection or the degree of privacy loss for each reference word embedding may differ. Accordingly, the electronic device can compare the first and second distances and set the greater distance as the reference distance. By using a single reference distance for multiple reference word embeddings, the electronic device can maintain a similar degree of privacy protection and degree of privacy loss.

[0151] According to one embodiment of the present disclosure, the electronic device may compare a first distance and a second distance. Referring to FIG. 9A, the electronic device may select the second distance, which is the distance between the second reference word embedding vector 'obesity' (522) and the word embedding vector 'anxiety disorder' (429), as the reference distance, since the second distance is greater than the first distance, which is the distance between the first reference word embedding vector 'Seoul' (521) and the word embedding vector Berlin (415).

[0152] According to one embodiment of the present disclosure, an electronic device may select, from among candidate word embeddings, candidate word embeddings within a reference distance from a first reference word embedding vector as replacement candidates for a first string. The electronic device may select, from among candidate word embeddings, candidate word embeddings within a reference distance from a second reference word embedding vector as replacement candidates for a second string.

[0153] Referring to FIG. 9A, the electronic device may select candidate word embeddings having a distance smaller than the reference distance from the first reference word embedding vector 'Seoul' (521). For example, the electronic device may select the word embedding vector Tokyo (412), the word embedding vector Beijing (413), the word embedding vector Berlin (415), the word embedding vector Sydney (416), and the word embedding vector New York (417) as alternative candidates (910).

[0154] Additionally, referring to FIG. 9A, the electronic device may select candidate word embeddings having a distance smaller than the reference distance based on the second reference word embedding vector 'obesity' (522). For example, the electronic device may select the word embedding vector 'diabetes' (422), the word embedding vector 'depression' (428), and the word embedding vector 'anxiety disorder' (429) as alternative candidates (820).

[0155] FIG. 9b is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to delete a candidate word embedding from a replacement candidate based on a distance from a reference word embedding vector.

[0156] According to one embodiment of the present disclosure, the electronic device can delete a candidate word embedding from the replacement candidates based on a distance between a reference word embedding vector and the candidate word embedding.

[0157] According to one embodiment of the present disclosure, an electronic device can identify a distance between a reference word embedding vector and a candidate word embedding vector. The electronic device can calculate a cosine similarity between the reference word embedding vector and the candidate word embedding vector.

[0158] According to one embodiment of the present disclosure, an electronic device is a value of cosine similarity If the value is 0, the word corresponding to the reference word embedding and the word corresponding to the candidate word embedding can be identified as having no clear correlation or semantic connection between the two words. In addition, the electronic device can use the value of cosine similarity The closer the value is to -1, the more contrasting the meanings of the word corresponding to the reference word embedding and the word corresponding to the candidate word embedding are identified. Therefore, if the cosine similarity value is between 0 and -1, the electronic device can remove the candidate word embedding vector from the list of replacement candidates.

[0159] For example, referring to FIG. 9B, the electronic device can calculate the cosine similarity between the first reference word embedding vector 'Seoul' (521) and the word embedding vector stroke (425). Since the cosine similarity between the first reference word embedding vector 'Seoul' (521) and the word embedding vector stroke (425) is between 0 and -1, the electronic device can exclude the word embedding vector stroke (425) from the replacement candidate (930).

[0160] FIG. 10 is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to replace a string representing personal information with another string.

[0161] According to one embodiment of the present disclosure, an electronic device can replace a string representing personal information with another string. The electronic device can obtain a replacement candidate for the string representing the personal information.

[0162] For example, referring to FIG. 10, an electronic device may obtain first replacement candidates (1010) for a first character string 'Seoul' (511) representing first personal information from an original text (510). The first replacement candidates (1010) may be candidate word embeddings selected based on a first reference word embedding vector. The electronic device may randomly select one of the first replacement candidates (1010). The electronic device may obtain a character string corresponding to one of the selected first replacement candidates (1010). For example, the electronic device may obtain the character string 'Tokyo' corresponding to the word embedding vector 'Tokyo'. The electronic device may replace the first character string 'Seoul' (511) with the replacement character string 'Tokyo'.

[0163] Additionally, for example, the electronic device may obtain second replacement candidates (1020) for the second character string 'obesity' (512) representing the second personal information in the original text (510). The second replacement candidates (1020) may be candidate word embeddings selected based on a second reference word embedding vector. The electronic device may randomly select one of the second replacement candidates (1020). The electronic device may obtain a character string corresponding to one of the selected second replacement candidates (1020). For example, the electronic device may obtain the character string 'depression' corresponding to the word embedding vector 'depression'. The electronic device may replace the second character string 'obesity' (512) with the replacement character string 'depression'.

[0164] FIG. 11 is a diagram illustrating an operation of an electronic device according to one embodiment of the present disclosure to replace a string representing personal information with another string.

[0165] According to one embodiment of the present disclosure, an electronic device can obtain original text such as, "Hello, my name is Hong Gil-dong. I was recently diagnosed with diabetes and need to control my diet and exercise for three months. Please call +82)10-1234-1234."

[0166] According to one embodiment of the present disclosure, an electronic device can extract first personal information, 'Hong Gil-dong' (1110), from a text. The first personal information may represent a name. The electronic device can select candidate words to replace the first string 'Hong Gil-dong' (1110) representing the first personal information. The electronic device can determine candidate words such as Kim Cheol-su and Park Young-hee as replacement word candidates for the first string. The electronic device can randomly select one of the replacement word candidates for the first string. The electronic device can replace the first string with 'Kim Cheol-su' (1114), which is a string corresponding to one of the replacement word candidates.

[0167] According to one embodiment of the present disclosure, an electronic device can extract second personal information, "diabetes" (1120), from a text. The second personal information may represent health information. The electronic device can select candidate words to replace the second string "diabetes" (1120) representing the second personal information. The electronic device can determine weight loss, indigestion, etc. as candidate words to replace the second string. The electronic device can randomly select one of the candidate words to replace the second string. The electronic device can replace the second string with "indigestion" (1124), which is a string corresponding to one of the candidate words to replace the second string.

[0168] According to one embodiment of the present disclosure, an electronic device can extract third personal information, '+82)10-1234-1234' (1130), from a text. The third personal information may represent a contact information. The electronic device can select candidate words to replace the third string '+82)10-1234-1234' (1130) representing the third personal information. The electronic device can determine '+82)10-1111-2222', etc. as a replacement word candidate for the third string. The electronic device can randomly select one of the replacement word candidates for the third string. The electronic device can replace the third string with '+82)10-1111-2222' (1134), which is a string corresponding to one of the replacement word candidates.

[0169] According to one embodiment of the present disclosure, an electronic device can provide alternative text by replacing a string representing personal information with a replacement string. The electronic device can provide alternative text such as, "Hello, my name is Kim Cheol-su. I was recently diagnosed with indigestion and need to control my diet and exercise for the next three months. Please call +82)10-1111-2222." By replacing personal information with a string of similar meaning, the electronic device can protect the personal information while maintaining the meaning of the original personal information in the text as much as possible.

[0170] According to one embodiment of the present disclosure, an electronic device can utilize alternative text to train an artificial intelligence model. Furthermore, the electronic device can utilize alternative text for data analysis.

[0171] FIG. 12 is a block diagram illustrating an electronic device according to an embodiment of the present disclosure.

[0172] In one embodiment, the electronic device (1200) may include at least one processor (1210), a memory (1220), and a communication interface (1230). The configuration of the electronic device (1200) illustrated in FIG. 12 is merely an example, and examples of electronic devices performing an embodiment of the present disclosure are not limited to the configuration illustrated in FIG. 12. In one embodiment, one or more of the configurations illustrated in FIG. 12 may be deleted or modified, or a configuration not illustrated in FIG. 12 may be added to the electronic device (1200).

[0173] The processor (1210) can control the overall operations of the electronic device (1200). The processor (1210) can include a processing circuit. For example, the processor (1210) can control the overall operations of the electronic device (1200) to replace a string representing personal information with another string by executing one or more commands of a program stored in the memory (1220). There can be one or more processors (1210).

[0174] The processor (1210) may be configured as at least one of, for example, a central processing unit (CPU), a microprocessor, a graphic processing unit (GPU), an application specific integrated circuits (ASICs), a digital signal processor (DSPs), a digital signal processing device (DSPDs), a programmable logic device (PLDs), a field programmable gate array (FPGAs), an application processor (AP), a neural processing unit (NPU), or an artificial intelligence processor designed with a hardware structure specialized for processing an artificial intelligence model, but is not limited thereto.

[0175] The processor (1210) may obtain a text including a character string representing personal information. The processor (1210) may convert the character string into a reference word embedding using a word embedding model. The processor (1210) may identify a distance between each candidate word embedding and the reference word embedding. Based on the identified distance, the processor (1210) may select replacement candidates for the character string from among the candidate word embeddings. The processor (1210) may replace the character string representing the personal information with a character string corresponding to one of the replacement candidates. Since a description related to the operations of the processor has already been described in the description of the previous drawings, a repeated description will be omitted.

[0176] In one embodiment, there may be one or more processors (1210). When there is one or more processors (1210), the operations of the present disclosure may be performed by one or more processors individually or collectively executing instructions and / or programs stored in the memory (1220). When a method according to one embodiment of the present disclosure includes multiple operations, the multiple operations may be performed by one processor (1210) or by multiple processors (1210).

[0177] For example, when the first operation, the second operation, and the third operation are performed by the method according to one embodiment, the first operation, the second operation, and the third operation may all be performed by the first processor, or some of the first to third operations may be performed by the first processor (e.g., a general-purpose processor) and the remaining operations may be performed by the second processor (e.g., an AI-dedicated processor). Here, operations for training / inference of an AI model may be performed by an AI-dedicated processor, which is an example of the second processor. However, the embodiments of the present disclosure are not limited thereto.

[0178] One or more processors according to the present disclosure may be implemented as a single-core processor or a multi-core processor. If a method according to an embodiment of the present disclosure includes multiple operations, the multiple operations may be performed by a single core or by multiple cores included in one or more processors.

[0179] The memory (1220) can store one or more instructions and one or more programs that cause the electronic device (1200) to operate to replace a string representing personal information with another string.

[0180] For example, the memory (1220) may store a personal information replacement model (1240) readable (or executable) by the processor (1210). In one embodiment, the memory (1220) may store instructions that, when individually or in combination, are executed by the processor (1210), cause the electronic device (1200) to perform at least some of the operations of the electronic device (1200) described above with reference to FIGS. 1 to 11. For example, the processor (1210) may perform at least some of the operations described in FIGS. 1 to 11 by executing one or more instructions or codes stored in the memory (1220).

[0181] The memory (1220) may include various types of memory. The memory (1220) may include a flash memory type, a hard disk type, a multimedia card micro type, a card type memory (e.g., SD or XD memory, etc.), and may include a non-volatile memory including at least one of a ROM (Read-Only Memory), an EEPROM (Electrically Erasable Programmable Read-Only Memory), a PROM (Programmable Read-Only Memory), a magnetic memory, a magnetic disk, and an optical disk, and a volatile memory such as a RAM (Random Access Memory) or an SRAM (Static Random Access Memory).

[0182] The communication interface (1230) can perform data communication with other electronic devices under the control of the processor (1210). The communication interface (1230) can include a communication circuit.

[0183] The communication interface (1230) may perform data communication between the electronic device (1200) and another electronic device (e.g., a server, etc.) using at least one of data communication methods including, for example, wired LAN (e.g., Ethernet), wireless LAN (e.g., Wi-Fi), cellular network (e.g., 4G, 5G, etc.), Bluetooth, BLE (Bluetooth Low Energy), ZigBee, infrared communication (IrDA, infrared Data Association), NFC (Near Field Communication), RF communication, and various other types of known wireless / wired communication technologies.

[0184] The electronic device (1200) can transmit and receive data for replacing a string representing personal information with another string through a communication interface (1230) with another electronic device (e.g., a server, etc.).

[0185] Meanwhile, although not illustrated in FIG. 12, the electronic device (1200) may further include additional components to perform the operations described in the aforementioned embodiments. For example, the electronic device (1200) may further include a display, a microphone, an input / output interface, and the like.

[0186] The display can output a video signal to the screen of the electronic device (1200) under the control of the processor (1210). The display can output a video signal processed in the process of the electronic device (1200) replacing a character string expressing personal information with another character string. For example, the display can display an input field for text input, a text list for selecting an input text file, text search results, selected text, etc., and can display the result of replacing a character string expressing personal information within a text with another character string.

[0187] FIG. 13 is a block diagram exemplarily illustrating the configuration of a server according to one embodiment of the present disclosure.

[0188] The processor (1310), memory (1320), and communication interface (1330) of the server (1300) of FIG. 13 may correspond to the basic functions described in the processor (1210), memory (1220), and communication interface (1230) of the electronic device (1200) of FIG. 12. Therefore, for the sake of brevity, repeated descriptions are omitted.

[0189] Meanwhile, the server (1300) may be a computing device comprised of hardware elements with higher performance specifications than the electronic device (1200), enabling it to process complex operations and tasks for processing large amounts of data generated during the process of operating a personal information replacement model. Accordingly, each component of the server (1300) may have similar functions to each component of the electronic device (1200), but may have higher specifications in terms of performance (e.g., computational volume, computational speed, etc.).

[0190] The server (1300) may receive a personal information detection task request using a personal information replacement model from a client device (e.g., a user device), perform a personal information replacement task, and provide a result text in which a string expressing personal information is replaced to the client device as a response.

[0191] According to one aspect of the present disclosure, a method performed by an electronic device may be provided. The method may include obtaining a text including a character string representing personal information. The method may include converting the character string into a reference word embedding using a word embedding model. The method may include identifying a distance between each candidate word embedding and the reference word embedding. The candidate word embedding may be a word embedding converted using the word embedding model of a word included in a vocabulary used for training the word embedding model. The method may include selecting replacement candidates for the character string from among the candidate word embeddings based on the identified distance. The method may include replacing the character string representing the personal information with a character string corresponding to one of the replacement candidates.

[0192] According to one embodiment of the present disclosure, the word embedding model may be trained using the vocabulary as training data. The vocabulary may be characterized as a set of character strings representing personal information.

[0193] According to one embodiment of the present disclosure, the step of obtaining a text including a character string representing the personal information may include the step of obtaining an index of the character string representing the personal information from metadata of the text. The step of obtaining the text including the character string representing the personal information may include the step of extracting the character string from the text based on the index.

[0194] According to one embodiment of the present disclosure, the step of converting the string into a reference word embedding using the word embedding model may include the step of converting a first string representing first personal information included in the text into a first reference word embedding vector using the word embedding model. The step of converting the string into a reference word embedding using the word embedding model may include the step of converting a second string representing second personal information included in the text into a second reference word embedding vector using the word embedding model.

[0195] According to one embodiment of the present disclosure, the step of identifying the distance between candidate word embeddings converted from words used for training the word embedding model and the reference word embedding may include the step of calculating at least one of an angular difference or a length difference between the reference word embedding and each of the candidate word embeddings.

[0196] According to one embodiment of the present disclosure, the step of identifying a distance between candidate word embeddings converted from words used for training the word embedding model and the reference word embedding may include a step of calculating a distance between the first reference word embedding vector and each of the candidate word embeddings. The step of identifying a distance between each of the candidate word embeddings and the reference word embedding may include a step of calculating a distance between the second reference word embedding vector and each of the candidate word embeddings.

[0197] According to one embodiment of the present disclosure, the step of selecting replacement candidates for the string from among the candidate word embeddings based on the identified distance may include the step of setting the number of replacement candidates. The step of selecting replacement candidates for the string from among the candidate word embeddings based on the identified distance may include the step of selecting the set number of candidate word embeddings as replacement candidates for the string from among the candidate word embeddings in descending order of the identified distance.

[0198] According to one embodiment of the present disclosure, the step of selecting replacement candidates for the string from among the candidate word embeddings based on the identified distance may include the step of identifying a first distance corresponding to a predetermined order in descending order of the identified distance between the first reference word embedding vector and each of the candidate word embeddings. The step of selecting replacement candidates for the string from among the candidate word embeddings based on the identified distance may include the step of identifying a second distance corresponding to the predetermined order in descending order of the identified distance between the second reference word embedding vector and each of the candidate word embeddings. The step of selecting replacement candidates for the string from among the candidate word embeddings based on the identified distance may include the step of comparing the first distance and the second distance and selecting a larger distance as a reference distance. Based on the identified distance, the step of selecting replacement candidates for the string from among the candidate word embeddings may include the step of selecting candidate word embeddings that are within the reference distance from the first reference word embedding vector as replacement candidates for the first string from among the candidate word embeddings. Based on the identified distance, the step of selecting replacement candidates for the string from among the candidate word embeddings may include the step of selecting candidate word embeddings that are within the reference distance from the second reference word embedding vector as replacement candidates for the second string from among the candidate word embeddings.

[0199] According to one embodiment of the present disclosure, the step of replacing the string representing the personal information with a string corresponding to one of the replacement candidates may include the step of selecting one of the replacement candidates. The step of replacing the string representing the personal information with a string corresponding to one of the replacement candidates may include the step of obtaining a word corresponding to the selected replacement candidate. The step of replacing the string representing the personal information with a string corresponding to one of the replacement candidates may include the step of replacing the string representing the personal information with the obtained word.

[0200] According to one embodiment of the present disclosure, the personal information may include at least one of name, date of birth, contact information, financial information, health information, employment information, and sensitive information.

[0201] According to one aspect of the present disclosure, an electronic device may be provided. The electronic device may include at least one processor including processing circuitry; and a memory storing instructions. The instructions may be individually or collectively executed by the at least one processor, thereby enabling the electronic device to obtain a text including a character string representing personal information. The instructions may be individually or collectively executed by the at least one processor, thereby enabling the electronic device to convert the character string into a reference word embedding using a word embedding model. The instructions may be individually or collectively executed by the at least one processor, thereby enabling the electronic device to identify a distance between each candidate word embedding and the reference word embedding. The candidate word embedding may be a word embedding converted using the word embedding model from a word included in a vocabulary used for training the word embedding model. The electronic device may select replacement candidates for the string from among the candidate word embeddings based on the identified distance by individually or collectively executing the instructions by the at least one processor. The electronic device may replace the string representing the personal information with a string corresponding to one of the replacement candidates by individually or collectively executing the instructions by the at least one processor.

[0202] According to one embodiment of the present disclosure, the instructions are individually or collectively executed by the at least one processor, thereby enabling the electronic device to obtain an index of a character string representing the personal information from metadata of the text. The instructions are individually or collectively executed by the at least one processor, thereby enabling the electronic device to extract the character string from the text based on the index.

[0203] According to one embodiment of the present disclosure, when the instructions are individually or collectively executed by the at least one processor, the electronic device can convert a first character string representing first personal information included in the text into a first reference word embedding vector using the word embedding model. When the instructions are individually or collectively executed by the at least one processor, the electronic device can convert a second character string representing second personal information included in the text into a second reference word embedding vector using the word embedding model.

[0204] According to one embodiment of the present disclosure, the instructions are individually or collectively executed by the at least one processor, whereby the electronic device can calculate at least one of an angular difference or a length difference between the reference word embedding and each of the candidate word embeddings.

[0205] According to one embodiment of the present disclosure, the electronic device can calculate a distance between the first reference word embedding vector and each of the candidate word embeddings by individually or collectively executing the instructions by the at least one processor. The electronic device can calculate a distance between the second reference word embedding vector and each of the candidate word embeddings by individually or collectively executing the instructions by the at least one processor.

[0206] According to one embodiment of the present disclosure, the electronic device can set the number of replacement candidates by individually or collectively executing the instructions by the at least one processor. The electronic device can select the set number of candidate word embeddings from among the candidate word embeddings in descending order of the identified distances as replacement candidates for the string by individually or collectively executing the instructions by the at least one processor.

[0207] According to one embodiment of the present disclosure, when the instructions are individually or collectively executed by the at least one processor, the electronic device can identify a first distance corresponding to a predetermined order in descending order of the identified distances between the first reference word embedding vector and each of the candidate word embeddings. When the instructions are individually or collectively executed by the at least one processor, the electronic device can identify a second distance corresponding to the predetermined order in descending order of the identified distances between the second reference word embedding vector and each of the candidate word embeddings. When the instructions are individually or collectively executed by the at least one processor, the electronic device can compare the first distance and the second distance and select a larger distance as a reference distance. The electronic device may select, as replacement candidates for a first string, candidate word embeddings within a reference distance from the first reference word embedding vector among the candidate word embeddings, as replacement candidates for a first string, by individually or collectively executing the instructions by the at least one processor. The electronic device may select, as replacement candidates for a second string, candidate word embeddings within a reference distance from the second reference word embedding vector among the candidate word embeddings, as replacement candidates for a second string, by individually or collectively executing the instructions by the at least one processor.

[0208] According to one embodiment of the present disclosure, the electronic device can select one of the replacement candidates by individually or collectively executing the instructions by the at least one processor. The electronic device can obtain a word corresponding to the selected replacement candidate by individually or collectively executing the instructions by the at least one processor. The electronic device can replace the character string representing the personal information with the obtained word by individually or collectively executing the instructions by the at least one processor.

[0209] According to one aspect of the present disclosure, a computer-readable recording medium having recorded thereon a program for executing any one of the aforementioned and hereinafter described methods for operating an electronic device and / or an electronic device can be provided.

[0210] Meanwhile, embodiments of the present disclosure may also be implemented in the form of a recording medium containing computer-executable instructions, such as program modules, executed by a computer. Computer-readable media may be any available media that can be accessed by a computer, and include both volatile and nonvolatile media, removable and non-removable media. Furthermore, computer-readable media may include computer storage media and communication media. Computer storage media include both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, program modules, or other data. Communication media may typically include computer-readable instructions, data structures, or other data in a modulated data signal, such as program modules.

[0211] Additionally, a computer-readable storage medium may be provided in the form of a non-transitory storage medium. Here, the term "non-transitory storage medium" simply means a tangible device that does not contain signals (e.g., electromagnetic waves). This term does not distinguish between cases where data is permanently stored in the storage medium and cases where data is temporarily stored. For example, a "non-transitory storage medium" may include a buffer in which data is temporarily stored.

[0212] According to one embodiment, the method according to various embodiments disclosed in the present document may be provided as included in a computer program product. The computer program product may be traded as a product between a seller and a buyer. The computer program product may be distributed in the form of a machine-readable storage medium (e.g., compact disc read-only memory (CD-ROM)), or may be distributed online (e.g., downloaded or uploaded) through an application store or directly between two user devices (e.g., smartphones). In the case of online distribution, at least a portion of the computer program product (e.g., a downloadable app) may be temporarily stored or temporarily generated in a machine-readable storage medium, such as the memory of a manufacturer's server, an application store's server, or an intermediary server.

[0213] The above description of the present disclosure is provided for illustrative purposes only, and those skilled in the art will readily appreciate that modifications to other specific forms can be made without altering the technical spirit or essential features of the present disclosure. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, components described as being single may be implemented in a distributed manner, and similarly, components described as being distributed may be implemented in a combined manner.

[0214] The scope of the present disclosure is indicated by the claims described below rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present disclosure.

Claims

1. In a method performed by an electronic device, A step of obtaining a text (10; 310; 510) containing a character string (311; 511, 512) representing personal information; A step of converting the above string (311; 511, 512) into a reference word embedding (521, 522) using a word embedding model; A step of identifying the distance between candidate word embeddings (411-417, 421-429) converted from words used for training the word embedding model and the reference word embedding (521, 522); A step of selecting replacement candidates for the string (311; 511, 512) among the candidate word embeddings (411-417, 421-429) based on the identified distance; and A step of replacing the string (311; 511, 512) representing the above personal information with a string corresponding to one of the above replacement candidates; A method comprising:

2. In paragraph 1, The above word embedding model is, It is trained using a vocabulary as training data, The above vocabulary is, A method characterized in that it is a set of strings representing personal information.

3. In any one of paragraphs 1 and 2, The step of converting the string into a reference word embedding using the above word embedding model is as follows: A step of converting a first character string (511) representing first personal information included in the text into a first reference word embedding vector (521) using the word embedding model; and A step of converting a second character string (512) representing second personal information included in the text into a second reference word embedding vector (522) using the word embedding model; A method comprising:

4. In any one of paragraphs 1 to 3, The step of identifying the distance between the candidate word embeddings converted from words used in training the above word embedding model and the reference word embedding is as follows: A step of calculating at least one of an angular difference or a length difference between the above reference word embedding (521, 522) and each of the above candidate word embeddings (411-417, 421-429); A method comprising:

5. In paragraph 3, The step of identifying the distance between the candidate word embeddings converted from words used in training the above word embedding model and the reference word embedding is as follows: A step of calculating the distance between the first reference word embedding vector (521) and each of the candidate word embeddings (411-417, 421-429); and A step of calculating the distance between the second reference word embedding vector (522) and each of the candidate word embeddings (411-417, 421-429); A method comprising:

6. In any one of paragraphs 1 to 5, Based on the identified distance, the step of selecting replacement candidates for the string among the candidate word embeddings is: a step of setting the number of the above alternative candidates; and A step of selecting the set number of candidate word embeddings from among the candidate word embeddings (411-417, 421-429) in descending order of the identified distances as replacement candidates for the string (511, 512); A method comprising:

7. In paragraph 3 or paragraph 5, Based on the identified distance, the step of selecting replacement candidates for the string among the candidate word embeddings is: A step of identifying a first distance corresponding to a predetermined order in descending order of the identified distance between the first reference word embedding vector (521) and each of the candidate word embeddings (411-417, 421-429); A step of identifying a second distance corresponding to the predetermined order in descending order of the identified distance between the second reference word embedding vector (522) and each of the candidate word embeddings (411-417, 421-429); A step of comparing the first distance and the second distance and selecting a larger distance as a reference distance; A step of selecting candidate word embeddings (411-417, 421-429) that are within the reference distance from the first reference word embedding vector (521) as replacement candidates for the first string (511); and A step of selecting candidate word embeddings (411-417, 421-429) that are within the reference distance based on the second reference word embedding vector (522) as replacement candidates for the second string (512); A method comprising:

8. In electronic devices, At least one processor comprising processing circuitry; and Contains memory that stores instructions, The electronic device, whereby the instructions are individually or collectively executed by the at least one processor, Obtain a text (10; 310; 510) containing a character string (311; 511, 512) representing personal information, Using the word embedding model, the above string (311; 511, 512) is converted into a reference word embedding (521, 522), Identify the distance between the candidate word embeddings (411-417, 421-429) converted from words used for training the above word embedding model and the reference word embedding (521, 522), Based on the identified distance, select replacement candidates for the string (311; 511, 512) among the candidate word embeddings (411-417, 421-429), Replace the string (311; 511, 512) representing the above personal information with a string corresponding to one of the above replacement candidates. Electronic devices.

9. In paragraph 8, The above word embedding model is, It is trained using a vocabulary as training data, The above vocabulary is, Characterized by being a set of strings representing personal information, Electronic devices.

10. In any one of paragraphs 8 to 9, The electronic device, whereby the instructions are individually or collectively executed by the at least one processor, Using the above word embedding model, the first character string (511) representing the first personal information included in the text is converted into a first reference word embedding vector (521), Using the above word embedding model, the second character string (512) representing the second personal information included in the text is converted into a second reference word embedding vector (522). Electronic devices.

11. In any one of paragraphs 8 to 10, The electronic device, whereby the instructions are individually or collectively executed by the at least one processor, Computing at least one of the angle difference or length difference between the above reference word embedding (521, 522) and each of the above candidate word embeddings (411-417, 421-429), Electronic devices.

12. In paragraph 10, The electronic device, whereby the instructions are individually or collectively executed by the at least one processor, Calculate the distance between the first reference word embedding vector (521) and each of the candidate word embeddings (411-417, 421-429), Calculating the distance between the second reference word embedding vector (522) and each of the candidate word embeddings (411-417, 421-429), Electronic devices.

13. In any one of paragraphs 8 to 12, The electronic device, whereby the instructions are individually or collectively executed by the at least one processor, Set the number of the above alternative candidates, Selecting the candidate word embeddings (411-417, 421-429) in the order of the smallest identified distance as the candidate word embeddings of the set number as replacement candidates for the string (511, 512). Electronic devices.

14. In paragraph 10 or 12, The electronic device, whereby the instructions are individually or collectively executed by the at least one processor, Identifying a first distance corresponding to a predetermined order in descending order of the identified distance between the first reference word embedding vector (521) and each of the candidate word embeddings (411-417, 421-429), Identifying a second distance corresponding to the predetermined order in descending order of the identified distance between the second reference word embedding vector (522) and each of the candidate word embeddings (411-417, 421-429), By comparing the first distance and the second distance, the larger distance is selected as the reference distance, Among the above candidate word embeddings (411-417, 421-429), candidate word embeddings within the reference distance based on the first reference word embedding vector (521) are selected as replacement candidates for the first string (511). Among the above candidate word embeddings (411-417, 421-429), candidate word embeddings within the reference distance based on the second reference word embedding vector (522) are selected as replacement candidates for the second string (512). Electronic devices.

15. A computer-readable recording medium having recorded thereon a program for performing the method of any one of clauses 1 to 7 on a computer.

Images