KR20220044011A - Method and system for text style transfer, and learning method for implementing same - Google Patents

Abstract

According to one aspect according to the technical idea of the present disclosure, provided is a method for converting a text style, which includes the steps of: obtaining input text; obtaining, through an encoder, a latent representation for the obtained input text; obtaining the style embedding corresponding to a target style through a style embedding model; coupling the obtained style embedding to the latent representation; and obtaining, via the decoder, output text for the combined latent representation.

Description

Text style transformation method and system, and learning method for its implementation

The technical idea of the present disclosure relates to a text style conversion method and system, and a learning method for implementing the same.

Text style conversion is the task of modifying text with specified style attributes. Given input text having a particular style (eg positive), text style transformation aims to change the input text to obtain a different style (eg negative).

In this text style conversion, it is important to reflect the given target style, maintain the content of parts that are not related to the style, and generate natural sentences.

According to the conventional method, a model for text style transformation was provided using a supervised learning method using a parallel data set including a pair of source sentences and target sentences. However, this approach may make it impossible to obtain parallel data sets that achieve a one-to-one correspondence with a given style as a whole.

In addition, a separation approach to separate the style component and the content component within the input text is being studied, but it can be difficult to divide the sentence into the style component and the content component since the components are not mutually exclusive. In addition, such separation may cause loss of information.

SUMMARY OF THE INVENTION An object of the present invention is to provide a method and system for improving sentence restoration ability and style accuracy when converting text styles.

An object of the present invention is to provide a method and system capable of converting the style of input text by reflecting the strength of the style.

In order to achieve the above object, a text style conversion method according to an aspect according to the technical spirit of the present disclosure includes: obtaining input text; obtaining, through an encoder, a latent representation for the obtained input text; acquiring a style embedding corresponding to a target style through the style embedding model; coupling the obtained style embeddings to the latent representation; and obtaining, via the decoder, output text for the combined latent representation.

According to an embodiment, obtaining the latent representation comprises: inputting the input text into the encoder; and obtaining, through a neural network of the encoder, the latent representation corresponding to the input text.

According to an embodiment, the obtaining of the style embedding may include: inputting style information corresponding to the target style into the style embedding model; and obtaining the style embedding corresponding to the target style through a neural network of the style embedding model.

According to an embodiment, obtaining the output text comprises: inputting the combined latent representation into the decoder; and obtaining the output text corresponding to the combined latent expression through the neural network of the decoder, wherein the output text may be text reconstructed by transforming the input text according to the target style.

According to an embodiment, the latent expression and the style embedding may have a vector form, and the combining may include obtaining the combined latent expression through a vector sum of the latent expression and the style embedding. .

According to an embodiment, obtaining the combined latent expression includes: applying a style strength to the style embedding; and obtaining the combined latent expression through a vector sum of the style embedding to which the style intensity is applied and the latent expression, wherein the step of obtaining the output text includes an output text that changes according to the style intensity. It may include the step of obtaining.

A text style conversion system according to an aspect according to the technical spirit of the present disclosure includes: an encoder including a neural network providing a latent expression vector from input text; a style embedding model comprising a neural network providing a style embedding vector corresponding to a target style; and at least one computing device configured to implement a decoder comprising a neural network that provides output text using a combined latent representation vector provided by a combination of the latent representation vector and the style embedding vector.

A text style conversion system according to an aspect according to the technical spirit of the present disclosure includes an encoder and a decoder for reconstructing style-converted output text from input text, and a style embedding model that provides a style embedding vector for style conversion. The learning method of the text style conversion system may include: controlling learning of the encoder and the decoder based on a reconstruction result of the output text; and controlling learning of the style embedding model based on the classification result of the style.

According to an embodiment, the controlling of the learning of the encoder and the decoder may include updating the neural network included in the encoder and the neural network included in the decoder according to a reconstruction loss function based on the reconstruction result of the output text. may include

According to an embodiment, the controlling of the learning of the encoder and the decoder may include: providing, by a neural network of the encoder, a latent expression vector using the input text; providing, by the neural network of the style embedding model, a style embedding vector using style information of the input text; obtaining a combined latent expression vector through the combination of the latent expression vector and the style embedding vector; and reconstructing, by the neural network of the decoder, the output text using the combined latent expression vector.

According to an embodiment, the controlling of the learning of the style embedding model may include: providing, by the neural network of the encoder, a latent expression vector using the input text; calculating a degree of similarity with the latent expression for each of a plurality of style embeddings corresponding to the set number of styles; classifying the style of the input text based on the calculated similarities; and controlling learning of the neural network of the style embedding model according to a classification loss function based on the classification result.

According to the technical idea of the present disclosure, by separating the style embedding model for style expression from the encoder and the decoder for sentence reconstruction, the neural network structure of the encoder and the decoder can be simplified, and the text reconstruction ability can be improved.

In addition, as the style embedding model is separated, the degree of style conversion may be changed by setting and adjusting style strength during style conversion. Accordingly, since more continuous style conversion is possible compared to the conventional method, the utility of the text style conversion system can be improved.

Effects according to the technical spirit of the present disclosure are not limited to the above-mentioned effects, and other effects not mentioned may be clearly understood by those of ordinary skill in the art to which the present invention belongs from the description below. There will be.

In order to more fully understand the drawings cited in this disclosure, a brief description of each drawing is provided.
1 is a schematic block diagram of a text style conversion system according to an exemplary embodiment of the present disclosure.
2 is a diagram for explaining learning operations of an encoder and a decoder of a sentence reconstruction module according to an exemplary embodiment of the present disclosure.
3 is a flowchart illustrating a learning operation of a style embedding model of a style embedding module and a style classifier according to an exemplary embodiment of the present disclosure.
4 is a flowchart illustrating a text style conversion method according to an exemplary embodiment of the present disclosure.
5 is a diagram for explaining a text style conversion operation of a text style conversion system according to an exemplary embodiment of the present disclosure.
6 illustrates examples of output text in which the style of input text is converted according to a text style conversion method according to an exemplary embodiment of the present disclosure.
7 is an example of comparing output text according to a text style conversion method according to an exemplary embodiment of the present disclosure with output text according to a conventional method.
8 is a schematic block diagram illustrating a control configuration of a device constituting the text style conversion system of FIG. 1 .

Exemplary embodiments according to the technical spirit of the present disclosure are provided to more fully explain the technical spirit of the present disclosure to those of ordinary skill in the art, and the following embodiments are modified in various other forms may be, and the scope of the technical spirit of the present disclosure is not limited to the following embodiments. Rather, these embodiments are provided to more fully and complete the present disclosure, and to fully convey the technical spirit of the present invention to those skilled in the art.

Although the terms first, second, etc. are used in this disclosure to describe various members, regions, layers, regions, and/or components, these members, parts, regions, layers, regions, and/or components refer to these terms It is self-evident that it should not be limited by These terms do not imply a specific order, upper and lower, or superiority, and are used only to distinguish one member, region, region, or component from another member, region, region, or component. Accordingly, a first member, region, region, or component to be described below may refer to a second member, region, region, or component without departing from the teachings of the present disclosure. For example, without departing from the scope of the present disclosure, a first component may be referred to as a second component, and similarly, the second component may also be referred to as a first component.

Unless defined otherwise, all terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the concepts of this disclosure belong, including technical and scientific terms. In addition, commonly used terms as defined in the dictionary should be construed as having a meaning consistent with their meaning in the context of the relevant technology, and unless explicitly defined herein, in an overly formal sense. shall not be interpreted.

In cases where certain embodiments may be implemented differently, a specific process sequence may be performed differently from the described sequence. For example, two processes described in succession may be performed substantially simultaneously, or may be performed in an order opposite to the described order.

In the accompanying drawings, variations of the illustrated shapes can be expected, for example depending on manufacturing technology and/or tolerances. Therefore, the embodiments according to the technical spirit of the present disclosure should not be construed as being limited to the specific shape of the region shown in the present disclosure, but should include, for example, a change in shape resulting from a manufacturing process. The same reference numerals are used for the same components in the drawings, and duplicate descriptions thereof are omitted.

As used herein, the term 'and/or' includes each and every combination of one or more of the recited elements.

Hereinafter, embodiments according to the technical spirit of the present disclosure will be described in detail with reference to the accompanying drawings.

1 is a schematic block diagram of a text style conversion system according to an exemplary embodiment of the present disclosure.

Text style transfer refers to an operation of providing new output text by reflecting a new style or a desired style while preserving the content of input text. As an example, text style transformation may include, given input text that includes a particular style (eg, positive), changing the input text to obtain a different style (eg, negative).

With the recent development of artificial intelligence technologies such as machine learning and deep learning, various studies are being conducted to implement a model or system for text style conversion through artificial intelligence-based learning.

Based on this, referring to FIG. 1 , the text style conversion system 10 may be implemented as a system that performs the above-described text style conversion operation. Such text style conversion system 10 may include one or more computing devices. When the text style conversion system 10 is implemented with two or more computing devices, the configurations shown in FIG. 1 are divided into the two or more computing devices, and the two or more computing devices may be connected through a network.

The text style conversion system 10 according to an exemplary embodiment of the present disclosure may include a sentence reconstruction module 11 and a style embedding module 13 .

The sentence reconstruction module 11 may be a module that performs an operation of providing an output text through reconstruction (reconstruction) of the input text, and a learning operation for this. For example, the sentence reconstruction module 11 may include an encoder 111 and a decoder 113 .

The encoder 111 may process the input text to provide a latent representation corresponding to the input text. The latent representation is a compressed representation of the input text, and may have, for example, a vector form. Accordingly, the latent expression may also be referred to as a latent expression vector.

The decoder 113 may provide the output text using the latent representation provided from the encoder 111 or a latent representation in which style embedding provided from the style embedding module 13 is combined with the latent representation. . In summary, the decoder 113 may provide the reconstructed output text by converting the input text into a style corresponding to the style embedding.

Each of the encoder 111 and the decoder 113 may be implemented as hardware including a neural network, software, or a combination thereof. According to an embodiment, the encoder 111 and the decoder 113 may be implemented as a Transformer-based sequence-to-sequence model in order to capture various meanings of the same word included in text. For example, the same word may have subtly different meanings in emotional text, and the encoder 111 and decoder 113 may capture the aforementioned nuances by using relational information in a Transformer-based model. For the Transformer-based sequence-to-sequence model, see "Google's multilingual neural machine translation system: Enablindg zero-shot translation. Transactions of the Association for Computational Linguistics, 5:339-351 (Johnson et al. 2017)", "Attention is all You need.

The neural network of the encoder 111 may be trained to provide the latent representation from the input text, and the neural network of the decoder 113 may be trained to provide the output text from the latent representation. According to an embodiment, the encoder 111 and the decoder 113 may be implemented as autoencoders, and may be trained to provide the same output text as the input text.

In the conventional case, it was implemented so that both learning for sentence reconstruction and learning for style transformation were performed in one module (model) for text style transformation. Accordingly, the structure of the model became complicated, and there were disadvantages that only discrete conversion was possible when the style was changed.

On the other hand, since the encoder 111 and the decoder 113 included in the sentence reconstruction module 11 according to an embodiment of the present disclosure can be learned (trained) only by the reconstruction loss of the sentence (text), the structure is different than in the prior art. can be simplified. In addition, learning for the reconstruction of the sentence may be intensively performed, and as a result, the ability to restore the sentence may be improved compared to the related art.

Details related to the learning operation of the encoder 111 and the decoder 113 will be described later with reference to FIG. 2 .

Meanwhile, the style embedding module 13 may provide style embedding based on the input style. The style embedding module 13 may include a style embedding model 131 , a similarity calculator 133 , and a style classifier 135 .

The style embedding model 131 may process the input style to provide a style embedding corresponding to the input style. The style embedding may be provided in a vector form. In this case, the style embedding may also be referred to as a style embedding vector. The input style may correspond to style information (or style label) indicating a style (target style) to be converted, but is not limited thereto. For example, the input style may represent an overall attitude or atmosphere of a sentence, such as positive or negative.

As mentioned above, the style embeddings may be combined with the latent representation provided from the encoder 111 . The decoder 113 may reconstruct the text whose style is converted according to the style embedding by providing the output text using the latent expression combined with the style embedding.

Also, according to an embodiment of the present disclosure, style strength may be set when style embeddings are combined, and accordingly, a degree of style transformation may vary. For example, as the style intensity increases, the degree of style conversion of the sentence may be stronger, and the style of the sentence may be converted into a similar style or the opposite style according to the sign of the style intensity value. This will be described in more detail later with reference to FIGS. 5 to 7 .

According to an embodiment, the style embedding model 131 may be implemented as hardware including a neural network, software, or a combination thereof. The similarity calculator 133 and the style classifier 135 may correspond to a configuration for learning the style embedding model 131 .

A learning operation using the components included in the style embedding module 13 will be described later with reference to FIG. 3 .

2 is a diagram for explaining learning operations of an encoder and a decoder of a sentence reconstruction module according to an exemplary embodiment of the present disclosure.

The learning operation according to the embodiment of FIGS. 2 to 3 may be performed by a learning device. The learning device may include at least one computing device constituting the text style conversion system 10 of FIG. 1 or a separate computing device for learning.

Referring to FIG. 2 , when input text x is input to the encoder 111 , the encoder 111 may provide (output) a latent representation (or latent representation vector; z) based on the input text x. . Meanwhile, an input style (s) corresponding to the style of the input text (x) may be input to the style embedding model 131 . The style embedding model 131 may provide a style embedding (or a style embedding vector; se) based on the input style (s).

The latent expression z provided from the encoder 111 and the style embedding se provided from the style embedding model 131 may be combined with each other by a combiner 112 . For example, the combined latent expression (z*) may correspond to a vector sum of the latent expression (z) and the style embedding (se), but is not limited thereto. The combined latent representation z* may be input to the decoder 113 . The decoder 113 may provide an output text (x') based on the combined latent representation (z*).

According to an embodiment, the encoder 111 and the decoder 113 are implemented as autoencoders, so that the output text (x') may be learned to be the same as the input text (x). The encoder 111 and the decoder 113 may be trained (trained) such that a reconstruction loss is minimized. According to an embodiment, a label smoothing regularization technique may be used to improve reliability and performance during training. The reconstruction loss (L _rec ) may be calculated according to the reconstruction loss function disclosed in Equation 1 below.

[Equation 1]

는 각각 어휘에 대한 예측 확률 분포와 실제 확률 분포를 나타낼 수 있다. E_θe 는 인코더를 나타내고, D_θd는 디코더를 나타낸다.Here, v represents the size of the vocabulary, and ε represents a smoothing parameter. with p

may represent the predicted probability distribution and the actual probability distribution for each vocabulary. E _θe denotes an encoder, and D _θd denotes a decoder.

Each of the encoder 111 and the decoder 113 may perform a learning operation by updating the weights of the output layer and the hidden layer of the neural network based on the reconstruction loss L _rec , so that the reconstruction loss L _rec is minimized. In this case, the operation of updating the weight may be performed in a direction opposite to the direction in which the operation of generating the output text is performed according to the back-propagation algorithm (eg, the output layer and the hidden layer of the neural network of the decoder 113 ). , the output layer of the neural network of the encoder 111, and the order of the hidden layer).

On the other hand, this reconstruction loss (L _rec ) only affects the learning of the encoder 111 and the decoder 113 , and may not affect the style embedding model 131 . In addition, since the encoder 111 and the decoder 113 are learned only by the reconstruction loss (L _rec ), intensive learning for the reconstruction (reconstruction) of the sentence may be possible.

3 is a flowchart illustrating a learning operation of a style embedding model of a style embedding module and a style classifier according to an exemplary embodiment of the present disclosure.

Referring to FIG. 3 , an input style (s) indicating the style of the input text (x) may be input to the style embedding model 131 .

The style embedding model 131 may output a style embedding set S including a plurality of style embeddings (S ₁ , S ₂ , ??, S _k ; k is the number (type) of styles). Each of the plurality of style embeddings included in the style embedding set S may correspond to a kind of common representation applicable to various texts having a corresponding style. In this case, any one of the plurality of style embeddings may correspond to a style embedding corresponding to the input style (s). For example, when the type of style is 'positive' and 'negative', the style embedding set S is a first style embedding S ₁ corresponding to 'positive' and a first style embedding corresponding to 'negative'. The second style embedding (S ₂ ) may be included. When the input style (s) corresponds to 'positive', the first style embedding (S ₁ ) among the plurality of style embeddings may be a style embedding corresponding to the input style (s).

Meanwhile, the input text x may be input to the encoder 111 . The encoder 111 may provide (output) a latent representation z based on the input text x. Since the portion related to the style expression is not separated from the input text x, the style characteristic of the input text x may be reflected in the latent expression z.

The similarity calculator 133 calculates the similarity (SIM(z,S)) between each of the style embeddings (S ₁ , S ₂ , ??, S _k ) included in the style embedding set (S) and the latent expression (z). can be calculated. For example, the similarity calculator 133 may calculate the similarity by calculating a dot product of the style embedding and the latent expression.

The style classifier 135 may classify the style embedding (s') having the highest similarity as the style embedding of the input style (s) by comparing the calculated similarities. For example, in the above example, when the style of the input text (x) corresponds to 'positive', the similarity (SIM(z,S ₁ )) between the first style embedding (S ₁ ) and the latent expression (z) is the second It may be higher than the similarity (SIM(z,S ₂ )) between the style embedding (S ₂ ) and the latent expression (z).

The style classifier 135 may calculate the classification loss for the classified style embedding according to the classification loss function disclosed according to Equation 2 below.

[Equation 2]

where C _θc denotes the style classifier 135, si denotes the style of the input text x, and sim _{z ,Si} denotes the similarity between the latent expression and the style embedding.

Each of the style embedding model 131 and the style classifier 135 updates the neural network (updating the weights of the output layer and the hidden layer) so that the classification loss for the classified style embedding is minimized based on the classification loss (L _se ). A learning operation can be performed. In this case, the operation of updating the weight may be performed in a direction opposite to the direction in which the style classification operation is performed according to a back-propagation algorithm.

On the other hand, this classification loss (L _se ) only affects the learning of the style embedding model 131 and the style classifier 135 , but also affects the encoder 111 and the decoder 113 included in the sentence reconstruction module 11 . may not reach In addition, since the style embedding model 131 is learned only by the classification loss L _se , intensive learning for style transformation may be possible.

1 to 3 , the text style conversion system 10 according to an embodiment of the present disclosure separates the sentence reconstruction module 11 and the style embedding module 13 to reconstruct (restore) the sentence. Learning for style expression and learning for style expression may be performed separately from each other. Accordingly, the structures of the encoder 111 and the decoder 113 of the sentence reconstruction module 11 may be simplified, and the sentence reconstruction ability may be improved.

A text style conversion operation using the learned sentence reconstruction module 11 and the style embedding module 13 according to the above-described embodiments will be described below with reference to FIGS. 4 to 7 .

4 is a flowchart illustrating a text style conversion method according to an exemplary embodiment of the present disclosure. 5 is a diagram for explaining a text style conversion operation of a text style conversion system according to an exemplary embodiment of the present disclosure.

4 and 5 , the text style conversion method may include obtaining an input text x ( S300 ), and obtaining a latent representation for the input text through the encoder 111 ( S310).

The input text x may be input to any one of at least one computing device constituting the text style conversion system 10 . For example, the input text x may be input through a communication interface or various known input means.

The encoder 111 may provide a latent representation (or latent representation vector) corresponding to the input text x through a neural network.

The text style conversion method includes the steps of combining the style embedding output from the style embedding model 131 to the latent expression (S320), and through the decoder 113, obtaining the output text for the combined latent expression. may be included (S330).

The style embedding model 131 may acquire an input style (s) including information (or label) of a style to be converted. Like the input text (x), the input style (s) may be input into any of said at least one computing. Meanwhile, the input style (s) may correspond to style information (or style label) of the target style to be converted, but may also be style information indicating the style of the input text (x) according to an embodiment. The style embedding model 131 may provide a style embedding (or a style embedding vector) corresponding to the input style (s) through a neural network.

The text style conversion system 10 may obtain a combined latent expression by combining the latent expression and the style embedding. As described above, the text style conversion system 10 may include a combiner 112 for calculating a vector sum of the latent expression and the style embedding, but is not limited thereto.

The decoder 113 may reconstruct the text in which the style of the input text x is converted by providing the output text y corresponding to the combined latent expression through the neural network.

According to an embodiment, the text style conversion system 10 may apply a style strength to the style embedding, and may combine the style embedding to which the style strength is applied and the latent expression. As described above, the degree of style conversion may vary according to the value of the style strength, and thus more diverse output texts may be provided according to the application of the style strength. The style strength w may be applied based on Equation 3 below.

[Equation 3]

Here, z* may mean a combined latent expression, z means a latent expression, and se may mean style embedding.

Based on Equation 3, the style of the output text y may be changed according to the sign of the style strength. For example, if the input style (s) is the same as the style of the input text (x) and the style strength has a negative value, the output text (y) can be reconstructed into a sentence having a style opposite to that of the input text (x). There will be.

Since it is impossible to apply style strength according to the conventional text style conversion method, the style of the output text is discretely converted according to the set style type. However, according to the present disclosure, various types of output text may be provided even for the same style according to the adjustment of style strength, so that the style of the output text may be more continuously converted.

Hereinafter, examples of text style conversion according to an embodiment of the present disclosure will be described with reference to FIGS. 6 to 7 .

6 illustrates examples of output text in which the style of input text is converted according to a text style conversion method according to an exemplary embodiment of the present disclosure.

The text style conversion system 10 may convert input text having a negative style into output text having a positive style. For example, an input text such as "so, no treatment and no medication to help me deal with my condition." may include a 'no' part indicating a negative style. The text style conversion system 10 converts the input text into positive style output text by changing the negative style-related part (word) of the input text to the positive style word ('best', 'great', etc.). can

In this case, the input style input to the style embedding module 13 may be style information (or style label) corresponding to 'positive'. Alternatively, the input style may correspond to 'negative' corresponding to the style of the input text, and the style strength w may have a negative value when the latent expression of the input text and style embedding are combined.

Alternatively, the text style conversion system 10 may convert input text having a positive style into output text having a negative style. For example, 'thorough' and 'genuinely' parts representing positive styles may exist in input text such as "he is very thorough and genuinely cares for his customers." The text style conversion system 10 changes the part (word) related to the positive style of the input text to the word ('never', 'lazy', 'not', etc.) indicating the negative style, thereby outputting the input text in a negative style. can be converted to text.

In this case, the input style input to the style embedding module 13 may be style information (or style label) corresponding to 'negative'. Alternatively, the input style may correspond to 'negative' corresponding to the style of the input text, and the style strength w may have a negative value when the latent expression of the input text and style embedding are combined.

Also, the text style conversion system 10 may provide different output texts according to the style strength w. The text style conversion system 10 may provide the output text in which the degree of style conversion is increased as the value of the style strength w is increased. In each of the examples shown in FIG. 6 , it can be seen that as the style strength w increases from 8 to 10, the conversion degree of the output text generally increases.

In the text style conversion system 10 according to an embodiment of the present disclosure, the sentence reconstruction module 11 for reconstructing a sentence and the style embedding module 13 for style expression are separately implemented, thereby increasing style strength during style conversion. By applying it, more various types of output text can be obtained.

7 is an example of comparing output text according to a text style conversion method according to an exemplary embodiment of the present disclosure with output text according to a conventional method.

In the case of the conventional text style conversion method, both learning of sentence reconstruction and style embedding can be performed in one model. Accordingly, the complexity of the model increases, and there is a problem that only discrete style conversion is possible during style conversion. As shown in FIG. 7 , in the conventional method, only the output text identical to the input text is provided when converting to the same style.

On the other hand, in the case of the text style conversion method according to an embodiment of the present disclosure, since the sentence reconstruction module 11 and the style embedding module 13 are separately implemented, style conversion using more diverse expressions may be possible during style conversion. 7 , in the case of the text style conversion method according to an embodiment of the present disclosure, output text including an expression different from the input text may be provided even if the input text is converted into the same style.

Table 1 below shows evaluation results of the conventional text style conversion method and the text style conversion method according to an embodiment of the present disclosure.

[Table 1]

The 'Accuracy' item in Table 1 is an item indicating the sameness of input text and output text, and a lower value indicates a higher degree of style conversion. That is, it can be seen that the text style conversion method according to the present disclosure effectively converts the text style compared to the conventional method. Meanwhile, according to the present disclosure, it is possible to adjust the style intensity, and as the style intensity decreases, the degree of style conversion decreases, so the value of the 'Accuracy' item may increase.

In addition, the 'Self-BLEU (bilingual evaluation understudy)' item is an item for the identity of the non-styled part of the text. The higher the value, the better the output text with the non-styled part of the input text is reconstructed. means to be That is, it can be seen that the text style conversion method according to the present disclosure maintains a portion of the input text that is not related to the style better than the conventional method.

8 is a schematic block diagram illustrating a control configuration of a device constituting the text style conversion system of FIG. 1 .

Referring to FIG. 8 , a device 800 according to an embodiment of the present disclosure may correspond to any one of at least one computing device constituting the text style conversion system 10 described above in FIG. 1 .

The device 800 may include a processor 810 and a memory 830 . However, the components of the device 800 are not limited to the above-described example. For example, the device 800 may include more or fewer components than the aforementioned components. In addition, the processor 810 may be at least one, and the memory 830 may also be at least one. Also, the processor 810 and the memory 830 may be combined into one chip.

According to an embodiment of the present disclosure, the processor 810 may perform learning of the encoder 111 and the decoder 113 of the sentence reconstruction module 11 by using the input text and style information of the input text. Also, the processor 810 may perform learning of the style embedding module 13 using the input text and style information.

According to an embodiment, the processor 810 obtains the output text obtained by converting the style of the input text into the target style by using the learned sentence reconstruction module 11 and the style embedding module 13, It is possible to obtain the output text in which the style of the input text is variously converted according to the style strength.

The processor 810 includes hardware such as a CPU, an application processor (AP), an integrated circuit, a microcomputer, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), and/or a neural processing unit (NPU). can do.

According to an embodiment of the present disclosure, the memory 830 may store programs and data necessary for the operation of the device 800 .

Also, the memory 830 may store at least one of data generated or acquired through the processor 810 . For example, the memory 830 may store text for training or a loss function.

The memory 830 may be configured of a storage medium such as ROM, RAM, flash memory, SSD, HDD, or a combination of storage media.

Since the descriptions of the above embodiments are merely those given with reference to the drawings for a more thorough understanding of the present disclosure, they should not be construed as limiting the technical spirit of the present disclosure.

In addition, it will be apparent to those of ordinary skill in the art to which the present disclosure pertains that various changes and modifications can be made without departing from the basic principles of the present disclosure.

Claims (15)

obtaining input text;
obtaining, through an encoder, a latent representation of the obtained input text;
acquiring a style embedding corresponding to the target style through the style embedding model;
coupling the obtained style embeddings to the latent representation; and
obtaining, via a decoder, output text for the combined latent representation;
How to convert text styles. According to claim 1,
The step of obtaining the latent expression comprises:
inputting the input text into the encoder; and
obtaining, via a neural network of the encoder, the latent representation corresponding to the input text;
How to convert text styles. 3. The method of claim 2,
The step of obtaining the style embedding comprises:
inputting style information corresponding to the target style into the style embedding model; and
Acquiring the style embedding corresponding to the target style through a neural network of the style embedding model,
How to convert text styles. 4. The method of claim 3,
Obtaining the output text comprises:
inputting the combined latent representation to the decoder; and
obtaining, via the neural network of the decoder, the output text corresponding to the combined latent representation,
The output text is text reconstructed by converting the input text according to the target style.
How to convert text styles. According to claim 1,
the latent representation and the style embedding have a vector form,
The combining step is
obtaining the combined latent representation through a vector sum of the latent representation and the style embedding,
How to convert text styles. 6. The method of claim 5,
The step of obtaining the combined latent expression comprises:
applying a style strength to the style embedding; and
obtaining the combined latent expression through a vector sum of the style embedding to which the style strength is applied and the latent expression;
Obtaining the output text comprises:
to obtain an output text that changes according to the style strength,
How to convert text styles. an encoder comprising a neural network providing a latent representation vector from the input text;
a style embedding model comprising a neural network providing a style embedding vector corresponding to a target style; and
at least one computing device configured to implement a decoder comprising a neural network that provides output text using a combined latent representation vector provided by the combination of the latent representation vector and the style embedding vector;
Text style conversion system. 8. The method of claim 7,
The combined latent expression vector is,
provided by the vector sum of the latent expression vector and the style embedding vector,
Text style conversion system. 9. The method of claim 8,
Any one of the at least one computing device,
applying a style strength to the style embedding vector;
implemented to provide the combined latent representation through a vector sum of the style embedding vector to which the style strength is applied and the latent representation vector,
Text style conversion system. 10. The method of claim 9,
The decoder is
providing output text that changes according to the style strength;
Text style conversion system. 10. The method of claim 9,
Any one of the at least one computing device,
Provides a plurality of style embedding vectors to which different style strengths are applied,
the decoder provides a plurality of output texts based on the plurality of style embedding vectors and the latent representation;
at least some of the plurality of output texts are different from each other;
Text style conversion system. A learning method of a text style transformation system comprising an encoder and a decoder for reconstructing style-converted output text from input text, and a style embedding model for providing a style embedding vector for style transformation,
controlling learning of the encoder and the decoder based on the reconstruction result of the output text; and
Including the step of controlling the learning of the style embedding model based on the classification result of the style,
How to learn text style conversion system. 13. The method of claim 12,
Controlling the learning of the encoder and the decoder comprises:
Updating the neural network included in the encoder and the neural network included in the decoder according to a reconstruction loss function based on the reconstruction result of the output text,
How to learn text style conversion system. 14. The method of claim 13,
Controlling the learning of the encoder and the decoder comprises:
providing, by the neural network of the encoder, a latent expression vector using the input text;
providing, by the neural network of the style embedding model, a style embedding vector using style information of the input text;
obtaining a combined latent expression vector through the combination of the latent expression vector and the style embedding vector; and
Reconstructing, by the neural network of the decoder, the output text using the combined latent expression vector;
How to learn text style conversion system. 13. The method of claim 12,
The step of controlling the learning of the style embedding model,
providing, by the neural network of the encoder, a latent expression vector using the input text;
calculating a degree of similarity with the latent expression for each of a plurality of style embeddings corresponding to the set number of styles;
classifying the style of the input text based on the calculated similarities; and
Controlling the learning of the neural network of the style embedding model according to a classification loss function based on the classification result,
How to learn text style conversion system.

Images