Data augmentation device

Abstract

The present invention relates to a data augmentation device capable of augmenting data on a text cuboid. The data augmentation device comprises: a data collection module that collects text data; a preprocessing module that generates matrix data of a predefined method on the basis of a plurality of pieces of the text data; and a data augmentation module that generates augmented data by performing data augmentation on the matrix data by using a pre-trained neural network, wherein the matrix data may include at least one embedding vector generated for each keyword included in the text data.

Description

Data Augmentation Device

[0001] The present invention relates to a data augmentation device.

[0002] Specifically, the present invention relates to a data augmentation device capable of augmenting data on a text cuboid.

[0003] Recently, with technological advancements, research and development on data augmentation for training data classifiers are steadily progressing. Advances in artificial intelligence (AI) technology are also driving the development of technologies that augment similar text data using large-scale language models (e.g., ChatGPT) powered by neural networks.

[0004] However, most data augmentation technologies perform data augmentation on text data itself, and there is no data augmentation technology targeting matrix data in the form of preprocessed text data or matrix data itself.

[0005] Therefore, there is a need for data augmentation techniques targeting such matrix data or matrix data itself.

[0006] The problem to be solved by the present invention is to provide a data augmentation device capable of performing data augmentation in a manner suitable for text cuboid, which is matrix data in a predefined manner.

[0007] Specifically, the problem to be solved by the present invention is to provide a data augmentation device that can increase augmentation efficiency by augmenting data on a text cuboid rather than the text data itself and achieve excellent text classification results even with a small amount of learning data.

[0008] The objectives of the present invention are not limited to those mentioned above. Other objectives and advantages of the present invention not mentioned above can be understood through the following description and will be more clearly understood through the embodiments of the present invention. Furthermore, it will be readily apparent that the objectives and advantages of the present invention can be realized by the means and combinations thereof set forth in the claims.

[0009] A data augmentation device according to some embodiments of the present invention includes a data collection module that collects text data, a preprocessing module that generates matrix data in a predefined manner based on a plurality of the text data, and a data augmentation module that performs data augmentation on the matrix data using a pre-trained neural network to generate augmented data, wherein the matrix data may include at least one embedding vector generated for each keyword included in the text data.

[0010] In addition, the preprocessing module can extract at least one keyword from the text data, generate an embedding vector corresponding to the keyword, and generate the matrix data based on the embedding vector.

[0011] Additionally, the preprocessing module can generate an embedding vector for the keyword using a pre-learned transformation model.

[0012] Additionally, the above transformation model may include a pre-trained BERT (Bidirectional Encoder Representations from Transformers) model based on a neural network.

[0013] In addition, the data augmentation module may include a location information generation unit that generates location information related to the location of a target keyword to be augmented in the matrix data, and a decoding unit that generates the augmented data based on a random distribution and the location information.

[0014] In addition, the position information includes a plurality of position vectors, the decoding unit includes a plurality of decoding layers, and the position information generation unit can input each of the plurality of position vectors as an input value of each of the plurality of decoding layers.

[0015] In addition, the data augmentation module may further include, in the learning step, an encoding unit that converts the embedding vector for the target keyword and the location information into a latent representation, and a determination unit that controls a first learning process related to learning of the encoding unit and a second learning process related to learning of the decoding unit.

[0016] Additionally, the discrimination unit may include a first discrimination unit that trains the encoding unit in a direction that reduces an error between the latent expression and the random distribution during the first learning process.

[0017] In addition, the determination unit may include a second determination unit that trains the decoding unit in a direction to reduce an error between an output value of the encoding unit that uses the augmented data and the position information as input values ​​and an embedding vector for the target keyword and an output value of the encoding unit that uses the position information and the position information as input values, when the decoding unit generates augmented data in the second learning process.

[0018] Additionally, the random distribution may include a Gaussian random distribution.

[0019] A data augmentation device according to some embodiments of the present invention has a novel effect of performing data augmentation in a manner suitable for text cuboids, which are matrix data in a predefined format. Specifically, the data augmentation device according to some embodiments of the present invention enhances data on the text cuboid, rather than the text data itself, thereby increasing augmentation efficiency and achieving excellent text classification results even with a small amount of training data. Therefore, the data augmentation device according to some embodiments of the present invention can be utilized for various high-value-added services and profit generation.

[0020] Furthermore, data augmentation devices according to some embodiments of the present invention can enhance data augmentation efficiency by performing data augmentation using a data augmentation model based on a Generative Adversarial Network (GAN), thereby demonstrating higher performance when training a text classifier. This enables high-performance data augmentation with a small amount of data, minimizing human intervention in text data generation.

[0021] In addition, the data augmentation device according to some embodiments of the present invention can augment data in real time by maintaining only a small amount of data without storing a large amount of text data, thereby reducing the data storage space.

[0022] In addition to the above-described contents, the specific effects of the present invention are described together with the specific matters for carrying out the invention below.

[0023] FIG. 1 illustrates a data augmentation system according to some embodiments of the present invention.

[0024] FIG. 2 is a block diagram of a data augmentation device according to some embodiments of the present invention.

[0025] FIG. 3 is a diagram illustrating the operation of a preprocessing module according to some embodiments of the present invention.

[0026] FIG. 4 is a diagram illustrating a neural network structure of a data augmentation module according to some embodiments of the present invention.

[0027] Figure 5 is a diagram illustrating an autoencoder, which is an example of a neural network.

[0028] FIG. 6 is a detailed block diagram of a data augmentation module according to some embodiments of the present invention.

[0029] FIG. 7 is a drawing for explaining the location information generation process of a location information generation unit according to some embodiments of the present invention.

[0030] FIG. 8 is a diagram illustrating a first learning process of a data augmentation module according to some embodiments of the present invention.

[0031] FIG. 9 is a diagram illustrating a second learning process of a data augmentation module according to some embodiments of the present invention.

[0032] FIG. 10 is a diagram illustrating the execution steps of a data augmentation module according to some embodiments of the present invention.

[0033] The terms and words used in this specification and claims should not be interpreted based on their general or dictionary meanings. In accordance with the principle that inventors can define the concepts of terms and words to best describe their inventions, they should be interpreted in a way that is consistent with the technical concept of the present invention. Furthermore, the embodiments described in this specification and the configurations depicted in the drawings are merely examples of how the present invention can be realized and do not fully represent the technical concept of the present invention. Therefore, it should be understood that various equivalents, modifications, and applicable examples may exist as of the time of filing.

[0034] The terms first, second, A, B, etc. used in this specification and claims may be used to describe various components, but the components should not be limited by these terms. These terms are used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be referred to as the second component, and similarly, the second component may also be referred to as the first component. The term "and / or" includes any combination of a plurality of related listed items or any item among a plurality of related listed items.

[0035] The terminology used in this specification and claims is for the purpose of describing specific embodiments only and is not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. It should be understood that terms such as "comprise" or "have" in this application do not preclude the presence or addition of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification.

[0036] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0037] Terms defined in commonly used dictionaries should be interpreted to have a meaning consistent with their meaning in the context of the relevant technology, and will not be interpreted in an idealized or overly formal sense unless expressly defined in this application.

[0038] In addition, each configuration, process, procedure or method included in each embodiment of the present invention may be shared within a scope that is not technically inconsistent with each other.

[0039] Hereinafter, a data augmentation device according to some embodiments of the present invention will be described in detail with reference to FIGS. 1 to 10.

[0040]

[0041] FIG. 1 illustrates a data augmentation system according to some embodiments of the present invention.

[0042] Referring to FIG. 1, a data augmentation system (1) according to some embodiments of the present invention may include an external database (100), a data augmentation device (200), and a communication network (300).

[0043] An external database (100) may refer to a database that stores, manages, and transmits multiple text data. In other words, the external database (100) may be an entity that stores, manages, and transmits multiple text data.

[0044] Text data may include at least one text, including words, sentences, paragraphs, etc. The text data may include at least one text expressed in various languages, such as Korean, English, Chinese, and Japanese. However, embodiments of the present invention are not limited thereto.

[0045] The external database (100) may be in the form of a workstation, a data center, an internet data center (IDC), a direct attached storage (DAS) system, a storage area network (SAN) system, a network attached storage (NAS) system, and a redundant array of inexpensive disks (RAID) system, but the embodiments of the present invention are not limited thereto.

[0046] The data augmentation device (200) can perform data augmentation based on text data. In other words, the data augmentation device (200) can perform data augmentation based on at least one piece of text data to generate augmented data.

[0047] As some examples, the data augmentation device (200) can perform data augmentation on the matrix data after converting text data into matrix data (text cuboid) in a predetermined manner.

[0048] Specifically, first, the data augmentation device (200) can convert text data into matrix data in a predetermined format. At this time, the matrix data in a predetermined format may be referred to as a text cuboid. For example, the data augmentation device (200) can extract keywords from text data, generate an embedding vector corresponding to each keyword, and generate matrix data based on the generated embedding vector. At this time, the data augmentation device (200) can generate the embedding vector using a pre-learned conversion model.

[0049] Next, when matrix data is generated, the data augmentation device (200) can perform data augmentation based on the matrix data to generate augmented data. As some examples, the data augmentation device (200) can generate location information regarding the location of a keyword in the matrix data, and generate augmented data based on the generated location information and a random distribution. At this time, the location information may include a plurality of location vectors, the decoding unit may include a plurality of decoding layers, and the data augmentation device (200) can generate augmented data by inputting each of the plurality of location vectors as an input value of the plurality of decoding layers. Meanwhile, the data augmentation device (200) can be trained to reduce an error between augmented data output based on the location information and the random distribution, and a value output based on the location information and the embedding vector.

[0050] The detailed operation process of the data augmentation device (200) will be described later.

[0051] Meanwhile, the data augmentation device (200) may be in the form of a workstation, a data center, an internet data center (IDC), a direct attached storage (DAS) system, a storage area network (SAN) system, a network attached storage (NAS) system, and a redundant array of inexpensive disks (RAID) system, but the embodiments of the present invention are not limited thereto.

[0052] The communication network (300) serves to connect the external database (100) and the data augmentation device (200). That is, the communication network (300) refers to a communication network that provides a connection path so that the data augmentation device (200) can transmit and receive data from the external database (100). The communication network (300) may include wired networks such as LANs (Local Area Networks), WANs (Wide Area Networks), MANs (MetRoFolitan Area Networks), and ISDNs (Integrated Service Digital Networks), or wireless networks such as wireless LANs, CDMA, Bluetooth, and satellite communication, but the scope of the present invention is not limited thereto.

[0053] Hereinafter, the operation of the data augmentation device (200) will be described in more detail with reference to FIG. 2.

[0054]

[0055] FIG. 2 is a block diagram of a data augmentation device according to some embodiments of the present invention.

[0056] Referring to FIG. 2, a data augmentation device (200) according to some embodiments of the present invention may include a data collection module (210), a preprocessing module (220), and a data augmentation module (230).

[0057] The data collection module (210) can receive text data (hereinafter referred to as "TD"). As some examples, the data collection module (210) can receive text data (TD) from an external database (100 in FIG. 1).

[0058] Text data (TD) may include at least one text, including words, sentences, paragraphs, etc. In this case, the text data (TD) may include at least one text expressed in various languages, such as Korean, English, Chinese, and Japanese. However, embodiments of the present invention are not limited thereto.

[0059] The data collection module (210) can transmit the received text data (TD) to other components within the data augmentation device (200). For example, the data collection module (210) can transmit the received text data (TD) to a preprocessing module (220), etc., but the embodiments of the present invention are not limited thereto.

[0060] The preprocessing module (220) can generate matrix data (hereinafter referred to as “MD”) based on the received text data (TD).

[0061] As some examples, the preprocessing module (220) can convert text data (TD) into matrix data (MD) in a predetermined manner or format. The matrix data (MD) may also be referred to as a text cuboid or a text matrix.

[0062] For example, the preprocessing module (220) can extract keywords from text data (TD), generate an embedding vector corresponding to each keyword, and generate matrix data (MD) based on the generated embedding vector.

[0063] Matrix data (MD) may include at least one embedding vector generated for each keyword included in text data (TD). In other words, matrix data (MD) may include a matrix composed of embedding vectors for each keyword included in text data (TD). In other words, matrix data (MD) may have a matrix form, and the size of the matrix may be determined based on the number of keywords included in the text data (TD) and the length of the embedding vector (e.g., 128).

[0064] At this time, the data augmentation device (200) can generate an embedding vector using a pre-learned transformation model.

[0065] Hereinafter, the matrix data (MD) generation process of the preprocessing module (220) according to some embodiments of the present invention will be described in more detail with reference to FIG. 3.

[0066]

[0067] FIG. 3 is a diagram illustrating the operation of a preprocessing module according to some embodiments of the present invention.

[0068] Referring to FIGS. 2 and 3, the preprocessing module (220) can convert text data (TD) into matrix data (MD) in a predetermined manner or format.

[0069] Specifically, first, the preprocessing module (220) can extract a keyword (hereinafter referred to as “KEY”) from at least one text data (TD). As some examples, the preprocessing module (220) can extract a keyword (KEY) from the text data (TD) using a predefined keyword extraction algorithm. At this time, the keyword (KEY) can have the form of a predefined word including a noun, a verb, etc. As an example of FIG. 3, the preprocessing module (220) can extract words such as “Dog, Jump, Fence, Chase, Squirrel, Playful” included in the first text data (TD_1) as a first keyword (KEY_1), and can extract words such as “Cat, Run, House” included in the second text data (TD_2) as a second keyword (KEY_2).

[0070] Next, the preprocessing module (220) can generate an embedding vector (hereinafter referred to as “EV”) corresponding to the extracted keyword (KEY). Referring to FIG. 3 as an example, the preprocessing module (220) can convert each keyword (e.g., Dog, etc.) included in the first keyword (KEY_1) into an embedding vector (EV). Similarly, the preprocessing module (220) can also convert each keyword (e.g., Cat, etc.) included in the second keyword (KEY_2) into an embedding vector (EV). For example, the preprocessing module (220) can generate the embedding vector (EV) using a pre-learned conversion model (hereinafter referred to as “CM”). In this case, the conversion model (CM) may be a pre-learned model based on a neural network. For example, the transformation model (CM) may include a Bidirectional Encoder Representations from Transformers (BERT) model, but embodiments of the present invention are not limited thereto.

[0071] Next, the preprocessing module (220) can generate matrix data (MD) based on the generated embedding vector (EV) for each keyword. In other words, the preprocessing module (220) can configure matrix data (MD) based on the generated embedding vector (EV) for each keyword.

[0072] At this time, the matrix data (MD) may be in the form of a matrix composed of an embedding vector for each keyword included in the text data (TD). In other words, the matrix data (MD) may have a matrix form, and the size of the matrix may be determined according to the number of keywords included in the text data (TD) and the length of the embedding vector (e.g., 128). In Fig. 3, for the convenience of explanation, the number of keywords is 4 and the length of the embedding vector is 12, but it is obvious that the embodiment of the present invention is not limited thereto.

[0073]

[0074] Referring again to FIG. 2, the preprocessing module (220) can transfer the generated matrix data (MD) to the data augmentation module (230).

[0075] The data augmentation module (230) can generate augmented data (hereinafter referred to as “AD”) based on matrix data (MD).

[0076] As some examples, the data augmentation module (230) may perform data augmentation on matrix data (MD) to generate augmented data (AD). That is, the data augmentation module (230) may perform data augmentation on matrix data (MD) obtained by converting text data (TD) into a vector form, rather than on the text data (TD) itself, to generate augmented data (AD).

[0077] At this time, the data augmentation module (230) can generate augmented data (AD) using a pre-trained neural network.

[0078] Specifically, the data augmentation module (230) can be trained based on machine learning. To explain in more detail, deep learning, a type of machine learning, is a technology that learns by going down to a deep level in multiple stages based on data. In other words, deep learning refers to a set of machine learning algorithms that extract core data from multiple data sets by increasing the level.

[0079] As some examples, the data augmentation module (230) may utilize various known deep learning structures. For example, the data augmentation module (230) may utilize structures such as a convolutional neural network (CNN), a recurrent neural network (RNN), a deep belief network (DBN), a graph neural network (GNN), a generative adversarial network (GAN), a transformer, and an autoencoder.

[0080] Specifically, a Convolutional Neural Network (CNN) is a model that mimics the function of the human brain, based on the assumption that when recognizing an object, humans extract its basic features, then perform complex computations in the brain to recognize the object based on the results. CNNs can include, but are not limited to, well-known structures such as LeNet, AlexNet, VGGNet, GoogleNet, and ResNet.

[0081] RNN (Recurrent Neural Network) is widely used in natural language processing, etc., and is an effective structure for processing time-series data that changes over time. It can construct an artificial neural network structure by stacking layers at each moment.

[0082] A DBN (Deep Belief Network) is a deep learning structure constructed by stacking multiple layers of Restricted Boltzman Machines (RBMs), a deep learning technique. By repeatedly training RBMs (Restricted Boltzman Machines), a certain number of layers can be created, creating a DBN (Deep Belief Network) with that number of layers.

[0083] GNN (Graphic Neural Network, hereinafter referred to as GNN) represents an artificial neural network structure implemented in a way that derives similarity and feature points between modeling data by using modeling data modeled based on data mapped between specific parameters.

[0084] A Generative Adversarial Network (GAN) is an artificial neural network structure that uses a generative neural network and a discriminative neural network to generate new data in a similar form to the input data. GANs may include the well-known DCGAN (Deep Convolutional GAN), CGAN (Conditional GAN), WGAN (Wasserstein GAN), StyleGAN (Style-Based GAN), CycleGAN, etc., but embodiments of the present invention are not limited thereto.

[0085] Transformer is an artificial neural network with an attention-based encoder-decoder structure that can understand the overall meaning between input and output sequences. Transformer uses the attention mechanism to ensure that all elements of the input sequence influence the output sequence, allowing both the encoder and decoder to consider the entire sequence. Transformer can use natural language, time-series data, and even patched images as input.

[0086] An autoencoder is a deep learning architecture that extracts and reconstructs data features. Typically, an autoencoder comprises an encoder, which compresses input values, and a decoder, which restores the compressed data. The encoder transforms the input values ​​into a low-dimensional latent representation, and the decoder reconstructs the latent representation to the same dimensionality as the input values. Each encoder and decoder can be configured as a multilayer perceptron (MLP). When training an autoencoder, input data is input, and weights and biases are trained to minimize the difference between the output and the input values. This trained autoencoder can effectively extract input data features and reconstruct noisy input data. Autoencoders are primarily used in fields such as data compression, dimensionality reduction, noise removal, and data generation, and can also be utilized in areas such as image recognition, natural language processing, and speech recognition.

[0087] Meanwhile, the artificial neural network learning of the data augmentation module (230) can be achieved by adjusting the weights of the connection lines between nodes (and adjusting the bias value, if necessary) to produce the desired output for a given input. Furthermore, the artificial neural network can continuously update the weight values ​​through learning. Furthermore, methods such as backpropagation can be used for artificial neural network learning.

[0088] At this time, machine learning methods of artificial neural networks may include unsupervised learning, semi-supervised learning, and supervised learning. In addition, the data augmentation module (230) may be controlled to automatically update the artificial neural network structure for outputting post-learning analysis data according to settings.

[0089] Hereinafter, with reference to FIG. 4, the neural network structure (neural network structure) of the data augmentation module (230) according to some embodiments of the present invention will be described in detail, with reference to FIG. 5, the structure of the auto-encoder used by the data augmentation module (230) according to some embodiments of the present invention will be described in detail, and with reference to FIG. 6, the data augmentation process of the data augmentation module (230) will be described in more detail.

[0090]

[0091] FIG. 4 is a diagram illustrating a neural network structure of a data augmentation module according to some embodiments of the present invention.

[0092] Referring to FIGS. 2 and 4, a data augmentation module (230) according to some embodiments of the present invention may include an input layer (input) having matrix data (MD) as an input node, an output layer (Output) having augmented data (AD) as an output node, and M hidden layers arranged between the input layer and the output layer.

[0093] Here, weights can be assigned to the edges connecting the nodes of each layer. These weights or the presence or absence of edges can be added, removed, or updated during the learning process. Therefore, the weights of the nodes and edges between the k input nodes and i output nodes can be updated during the learning process.

[0094] Before the data augmentation module (230) performs learning, all nodes and edges can be set to initial values. However, when information is input cumulatively, the weights of the nodes and edges change, and in this process, a matching can be achieved between the parameters input as learning factors (matrix data (MD)) and the values ​​assigned to the output nodes (augmented data (AD)).

[0095] Additionally, when utilizing a cloud server, the data augmentation module (230) can receive and process a large number of parameters. Accordingly, the data augmentation module (230) can perform learning based on a large amount of data.

[0096] The weights of nodes and edges between input nodes and output nodes constituting the data augmentation module (230) can be updated by the learning process of the data augmentation module (230). In addition, it goes without saying that the parameters input or output from the data augmentation module (230) can be further expanded to various data other than matrix data (MD) and augmented data (AD).

[0097]

[0098] Figure 5 is a diagram illustrating an autoencoder, which is an example of a neural network.

[0099] Referring to FIG. 5, an autoencoder (hereinafter referred to as “AE”) includes an encoder neural network (hereinafter referred to as “EN”) and a decoder neural network (hereinafter referred to as “DN”), and may include a middle layer (ML) positioned between the encoder (EN) and the decoder (DN).

[0100] An autoencoder (AE) is a type of deep neural network model that compresses data (i.e., input data) received through an encoder (EN) to reduce the data, and then converts the reduced data to the same size as the input data from the encoder (EN) using a decoder (DN) and outputs it, thereby making the output data of the autoencoder (AE) identical to the input data.

[0101] An autoencoder (AE) learns the features of input data in an unsupervised manner. To achieve this, the autoencoder (AE) transforms the input data via an encoder (EN) into low-dimensional data (latent representation) that effectively represents the features. This data can then be reconstructed back into the original data via a decoder (DN).

[0102] An autoencoder (AE) can learn patterns inherent in the original data with the goal of minimizing the reconstruction error, which corresponds to the difference between the original data (X1, X2, X3, X4), i.e., the input data, and the reconstructed data (X1', X2', X3', X4), i.e., the output data.

[0103]

[0104] FIG. 6 is a detailed block diagram of a data augmentation module according to some embodiments of the present invention.

[0105] Referring to FIG. 6, the data augmentation module (230) may include a location information generation unit (231), an encoding unit (232), a decoding unit (233), and a determination unit (234). The determination unit (234) may include a first determination unit (234a) and a second determination unit (234b).

[0106] At this time, the location information generation unit (231) and the decoding unit (233) can be used in both the learning phase and the inference phase of the data augmentation module (230), and the encoding unit (232) and the determination unit (234) can be used only in the learning phase of the data augmentation module (230).

[0107] The location information generation unit (231) can generate location information related to the location of a target keyword to be augmented in matrix data (MD). At this time, the location information can include information related to the location of a target keyword, which is a keyword to be augmented among each keyword in the matrix data (MD), on the matrix data (MD).

[0108] Hereinafter, the location information generation process of the location information generation unit (231) will be described in detail with reference to FIG. 7.

[0109]

[0110] FIG. 7 is a drawing for explaining the location information generation process of a location information generation unit according to some embodiments of the present invention.

[0111] Referring to FIG. 7, the location information generation unit (231) can generate location information (Location Data, hereinafter referred to as "LD") related to the location of a target keyword (Target Keyword, hereinafter referred to as "TK") to be augmented in matrix data (MD). At this time, the location information (LD) can include information related to the location of the target keyword (TK) on the matrix data (MD).

[0112] Specifically, first, the location information generation unit (231) can search for a target keyword (TK) that is the target of data augmentation in the matrix data (MD). At this time, the target keyword (TK) may be input by the administrator of the data augmentation device (200 of FIGS. 1 and 2) or may be arbitrarily determined from among the keywords of the matrix data (MD). For convenience of explanation, FIG. 7 illustrates a case where "Jump," a keyword existing in the fourth row of the matrix data (MD), is the target keyword (TK).

[0113] Next, the location information generation unit (231) can generate a line vector (hereinafter referred to as “LV”) based on the matrix data (MD). At this time, the line vector (LV) can have the same number of rows as the matrix data (MD), and the value of each row can have a binary value. For example, the line vector (LV) can have a value of 0 or 1 for each row of the matrix data (MD). For example, the line vector (LV) can have a value of 1 for a row of a keyword included in keyword data that includes the target keyword (TK), and can have a value of 0 for a row of a keyword included in keyword data that does not include the target keyword (TK). Taking FIG. 3 and FIG. 7 as examples, the location information generation unit (231) can generate a line vector (LV) by setting the value of the row in which each keyword "Dog, Jump, Fence, Chase, Squirrel, Playful" included in the first text data (TD_1) that includes the target keyword (TK) "Jump" is located to 1, and setting the value of the row in which the keyword (e.g., Cat) included in other text data (e.g., TD_2) that does not include the target keyword (TK) "Jump" is located to 0. In FIG. 7, for convenience of explanation, the row set to a value of 1 is hatched, and the row set to a value of 0 is not hatched.

[0114] Next, the location information generation unit (231) can generate location information (LD) based on the line vector (LV). At this time, the location information (LD) may be information indicating the location of the target keyword (TK) on the line vector (LV). For example, the location information (LD) may be data that expresses information related to which location the target keyword (TK) belongs among a plurality of rows (values) of the line vector (LV) as a vector. For example, the location information generation unit (231) can determine location information (LD) related to the target keyword (TK) from the line vector (LV) through a predefined location information determination algorithm. At this time, the location information (LD) may include a plurality of location vectors. Although FIG. 7 illustrates a case where the length of the location vector of the location information (LD) is 5, it is obvious that the embodiment of the present invention is not limited thereto.

[0115]

[0116] Referring again to FIG. 6, the location information generation unit (231) can transmit the generated location information to the encoding unit (232) and the decoding unit (233) in the “learning phase” and to the decoding unit (233) in the “execution phase”.

[0117] The encoding unit (232) can be used in the learning stage of the data augmentation module (230). For example, the encoding unit (232) can generate a latent representation based on an embedding vector for a target keyword to be augmented and location information for the target keyword in the first learning process. As another example, the encoding unit (232) can generate a first latent representation based on the augmented data output by the decoding unit (233) and location information for the target keyword in the second learning process, and can generate a second latent representation based on an embedding vector for a target keyword to be augmented and location information for the target keyword. At this time, the encoding unit (232) can perform the role of an encoder (EN of FIG. 5) in an autoencoder.

[0118] The decoding unit (233) can be utilized in the learning and execution stages of the data augmentation module (230). As an example, the decoding unit (233) can generate augmented data based on a random distribution and positional information for a target keyword in the second learning process and execution stage. At this time, the decoding unit (233) can perform the role of a decoder (DN of FIG. 5) in an autoencoder. Meanwhile, the decoding unit (233) can include a plurality of decoding layers. The random distribution can include a Gaussian random distribution, but embodiments of the present invention are not limited thereto.

[0119] The discrimination unit (234) can control the learning process of the data augmentation module (230). As some examples, the first discrimination unit (234a) can control the first learning process related to the learning of the encoding unit (232), and the second discrimination unit (234b) can control the second learning process related to the learning of the decoding unit (233). In this case, the discrimination unit (234) can perform the role of a discriminator in the GAN model.

[0120] Hereinafter, the first learning process of the data augmentation module (230) will be described in detail with reference to FIG. 8, the second learning process of the data augmentation module (230) will be described in detail with reference to FIG. 9, and the execution steps of the data augmentation module (230) will be described in detail with reference to FIG. 10.

[0121]

[0122] FIG. 8 is a diagram illustrating a first learning process of a data augmentation module according to some embodiments of the present invention.

[0123] Referring to Fig. 8, the encoding unit (232) can be learned by the first discrimination unit (234a) in the first learning process.

[0124] Specifically, first, the encoding unit (232) can generate a latent representation (hereinafter referred to as "LR") based on an embedding vector (EV) for a target keyword to be enhanced and position information (LD) regarding the target keyword. In other words, the encoding unit (232) can compress the embedding vector (EV) for a target keyword to be enhanced and position information (LD) to derive a latent representation (LR).

[0125] The first discrimination unit (234a) can train the encoding unit (232) based on the latent representation (LR) generated by the encoding unit (232) and the Gaussian random distribution (hereinafter referred to as “GRD”).

[0126] As some examples, the first discriminator (234a) may train the encoder (232) in a direction that reduces the error between the latent representation (LR) and the Gaussian random distribution (GRD). For example, the first discriminator (234a) may train the encoder (232) so that the latent representation (LR) output from the encoder (232) follows the Gaussian random distribution (GRD). For example, the first discriminator (234a) may calculate a validity score by calculating the error between the latent representation (LR) and the Gaussian random distribution (GRD), and train the encoder (232) in a direction that increases the calculated validity score. In this case, the validity score may include the degree of agreement between the latent representation (LR) and the Gaussian random distribution (GRD).

[0127]

[0128] FIG. 9 is a diagram illustrating a second learning process of a data augmentation module according to some embodiments of the present invention.

[0129] Referring to Fig. 9, the decoding unit (233) can be learned by the second discrimination unit (234b) in the second learning process.

[0130] Specifically, first, the decoding unit (233) can generate augmented data (AD) based on a Gaussian random distribution (GRD) and positional information (LD) regarding a target keyword. In other words, the decoding unit (233) can generate augmented data (AD) by restoring original data through a Gaussian random distribution (GRD) and positional information (LD) regarding a target keyword.

[0131] Next, the encoding unit (232) can compress the augmented data (AD) and the embedding vector (EV) for the target keyword to be augmented. As some examples, the encoding unit (232) can generate a first latent representation (LR1) based on the augmented data (AD) and the position information (LD), and can generate a second latent representation (LR2) based on the embedding vector (EV) and the position information (LD) for the target keyword. At this time, in the second learning process, the encoding unit (232) learned in the first learning process can be used. In other words, the first learning process can be referred to as a pre-learning process for learning the encoding unit (232) used in the second learning process.

[0132] The second discrimination unit (234b) can train the decoding unit (233) based on the first latent representation (LR1) and the second latent representation (LR2).

[0133] In some examples, the second discriminator (234b) may train the decoding unit (233) to reduce the error between the first latent representation (LR1) and the second latent representation (LR2). For example, the second discriminator (234b) may train the decoding unit (233) such that the first latent representation (LR1) in which the augmented data (AD) restored from the decoding unit (233) is compressed follows the second latent representation (LR2) in which the Gaussian random distribution (GRD) is compressed. For example, the second discriminator (234b) may calculate a validity score by calculating the error between the first latent representation (LR1) and the second latent representation (LR2), and train the decoding unit (233) in a direction in which the calculated validity score increases. In this case, the validity score may include the degree of agreement between the first latent representation (LR1) and the second latent representation (LR2).

[0134]

[0135] FIG. 10 is a diagram illustrating the execution steps of a data augmentation module according to some embodiments of the present invention.

[0136] Referring to Fig. 10, in the execution step of the data augmentation module (230), the location information generation unit (231) and decoding unit (233) included in the data augmentation module (230) may be utilized. In other words, in the execution step of the data augmentation module (230), the encoding unit (232) and the determination unit (234) included in the data augmentation module (230) may be omitted.

[0137] Specifically describing the execution steps of the data augmentation module (230), first, the location information generation unit (231) can generate location information (LD) regarding the target keyword (TK) to be augmented based on matrix data (MD).

[0138] More specifically, the location information generation unit (231) searches for a target keyword (TK) that is a target of data augmentation in the matrix data (MD) as described above in FIG. 7, generates a line vector (LV), and then generates location information (LD) related to the location of the corresponding target keyword (TK) based on the line vector (LV). The generated location information (LD) may include a plurality of location vectors. Although FIG. 10 illustrates a case in which the length of the location vector of the location information (LD) is 5, as in FIG. 7, it is obvious that the embodiment of the present invention is not limited thereto.

[0139] The decoding unit (233) can restore data based on a Gaussian random distribution (GRD) and position information (LD) to generate augmented data (AD). The decoding unit (233) may include a plurality of decoding layers (hereinafter referred to as “DLs”) as illustrated in FIG. 10. For convenience of explanation, the decoding unit (233) is illustrated in FIG. 10 as including a first decoding layer (DL1) to a fifth decoding layer (DL5), but the embodiment of the present invention is not limited thereto, and the number of decoding layers (DLs) may be freely modified.

[0140] At this time, the location information generation unit (231) can input each of the plurality of location vectors included in the location information (LD) as input values ​​of each of the plurality of decoding layers (DL1 to DL5) included in the decoding unit (233). As an example of Fig. 10, the first decoding layer (DL1) can generate the first restored data (1st Restored Data, hereinafter referred to as “RD1”) using a Gaussian random distribution (GRD) and the first location vector of the location information (LD) as input values, and the second decoding layer (DL2) can generate the second restored data (2nd Restored Data, hereinafter referred to as “RD2”) using the first restored data (RD1) and the second location vector of the location information (LD) as input values. Thereafter, the third decoding layer (DL3) to the fifth decoding layer (DL5) can perform the restoration process in a similar manner.

[0141] In this manner, when the fifth decoding layer (DL5) finally generates augmented data (AD), the augmented data (AD) can be added to the matrix data (MD). As a result of the data augmentation, the number of embedding vectors for the target keyword (TK) can increase.

[0142] The above description is merely an example of the technical idea of ​​the present embodiment, and those skilled in the art to which the present embodiment pertains may make various modifications and variations without departing from the essential characteristics of the present embodiment. Therefore, the present embodiments are not intended to limit the technical idea of ​​the present embodiment, but to explain it, and the scope of the technical idea of ​​the present embodiment is not limited by these embodiments. The protection scope of the present embodiment should be interpreted by the following claims, and all technical ideas within a scope equivalent thereto should be interpreted as being included in the scope of rights of the present embodiment.

Claims

1. Data collection module that collects text data; A preprocessing module for generating matrix data in a predefined manner based on a plurality of the above text data; and A data augmentation module that generates augmented data by performing data augmentation on the matrix data using a pre-trained neural network, The above matrix data includes at least one embedding vector generated for each keyword included in the text data. Data augmentation device.

2. In paragraph 1, The above preprocessing module, Extracting at least one keyword from the text data, Generate an embedding vector corresponding to the above keyword, Generating the matrix data based on the above embedding vector Data augmentation device.

3. In paragraph 2, The above preprocessing module, Generate an embedding vector for the above keywords using a pre-trained transformation model. Data augmentation device.

4. In paragraph 3, The above transformation model includes a BERT (Bidirectional Encoder Representations from Transformers) model that is pre-trained based on a neural network. Data augmentation device.

5. In paragraph 1, The above data augmentation module, In the above matrix data, a location information generation unit that generates location information related to the location of the target keyword to be enhanced, and A decoding unit that generates the augmented data based on a random distribution and the location information. Data augmentation device.

6. In paragraph 5, The above location information includes multiple location vectors, The above decoding unit includes a plurality of decoding layers, The above location information generation unit inputs each of the plurality of location vectors as an input value of each of the plurality of decoding layers. Data augmentation device.

7. In paragraph 5, The above data augmentation module, in the learning phase, An encoding unit that converts the embedding vector for the target keyword and the location information into a latent representation, Further comprising a determination unit that controls a first learning process related to learning of the encoding unit and a second learning process related to learning of the decoding unit. Data augmentation device.

8. In paragraph 7, The above-mentioned judgment unit, in the first learning process, A first discriminator that trains the encoding unit in a direction that reduces the error between the latent expression and the random distribution. Data augmentation device.

9. In paragraph 7, The above-mentioned judgment unit, in the second learning process, When the decoding unit generates augmented data using the random distribution and the location information as input values, A second determination unit that trains the decoding unit in a direction that reduces the error between the output value of the encoding unit that uses the augmented data and the location information as input values ​​and the embedding vector for the target keyword and the output value of the encoding unit that uses the location information as input values. Data augmentation device.

10. In paragraph 5, The above random distribution includes a Gaussian random distribution. Data augmentation device.

Images