KR102410582B1 - Apparatus, method and computer program for augmenting learning data for harmful words - Google Patents

Abstract

A device for augmenting learning data for harmful words, by inputting a specific utterance included in the learning data into a pre-learned ethical model, and a plurality of vocabulary and phrases included in the specific utterance to belong to each of a plurality of preset unethical classes a harmful word detection unit calculating a probability value of , and detecting harmful words included in the specific utterance based on the plurality of probability values; a harmful word classification unit for classifying the harmful word into at least one of the plurality of unethical classes based on the plurality of probability values for the harmful word; and a harmful word enhancer configured to generate new learning data by augmenting the harmful word by applying a predefined syllable transformation method to the harmful word based on the probability value of at least one unethical class classified in the harmful word.
It is a technology developed through the Seoul Innovation Challenge 2021 (Preliminary) (SC210060) "Development of Ethics Determination Engine for Open Domain Conversational Chatbot Business" by Seoul Industry Promotion Agency.

Description

Apparatus, method and computer program for augmenting learning data for harmful words

The present invention relates to an apparatus, method and computer program for augmenting learning data for harmful words.

Recently, due to the controversy over sexual harassment and hate speech by the artificial intelligence chatbot 'Iruda', the problem of ethical language expression and recognition of interactive artificial intelligence has emerged.

Artificial intelligence absorbs various biases and unethical expressions inherent in data in the process of learning human conversations in natural language. In particular, in the online space where anonymity is guaranteed, the use of unethical language such as profanity and hate speech is more free, so if such conversations are used as raw materials for learning, artificial intelligence will inevitably be biased.

Meanwhile, compared to English, Korean unethical datasets and related studies for machine learning are still lacking. In English, relatively diverse data are publicly available, and standards for hate speech are subdivided. Generally, hate speech is defined as an expression that insults, attacks, or insults the identity of an individual or group as a target.

However, the only currently published data on hate speech in Korean is the Korean Hate Speech Detection Dataset, which collects comments on entertainment articles from domestic portal sites. In the set, 10,000 comments were classified according to three criteria: Hate, Offensive, and None, and whether gender bias was included or not was determined.

In this regard, recently, there is a technology for detecting unethical utterances such as profanity, insult, and sexism included in specific utterances using artificial intelligence. The prior art detects unethical utterances included in specific utterances using an artificial intelligence-based learning model.

However, the prior art learning model does not have a detailed classification of unethical utterances that can be classified according to various criteria, so it is not easy to flexibly apply to various services. For example, the first learning model is sensitive to utterances related to 'prejudice', but is relatively insensitive to utterances related to 'sexual harassment', and the second learning model is sensitive to utterances related to 'sexual harassment', They are relatively insensitive to utterances related to 'slander'. In addition, since the prior art learning model discriminates the presence or absence of an unethical expression binary, detection accuracy or flexibility in data utilization may be reduced. Furthermore, classification criteria emphasizing only specific attributes may be limited.

It is a technology developed through the Seoul Innovation Challenge 2021 (Preliminary) (SC210060) "Development of Ethics Determination Engine for Open Domain Conversational Chatbot Business" by Seoul Industry Promotion Agency.

Korean Patent Publication No. 10-0662337 (Registered on December 21, 2006) Korean Patent Publication No. 10-2011-0056999 (published on May 31, 2011)

The present invention is to solve the problems of the prior art described above, and to provide a learning data augmentation apparatus for learning so that an ethical model for detecting harmful words can be flexibly applied to various services.

Another object of the present invention is to provide a learning data augmentation device that learns to accurately detect various types of harmful words, such as harmful words with deformed forms and words with some forms omitted.

However, the technical problems to be achieved by the present embodiment are not limited to the technical problems described above, and other technical problems may exist.

As a means for achieving the above-described technical problem, an embodiment of the present invention provides an apparatus for augmenting learning data for harmful words, inputting a specific utterance included in the learning data into a pre-learned ethical model to the specific a harmful word detection unit for calculating a plurality of probability values that vocabulary and phrases included in the utterance belong to each of a plurality of preset unethical classes, and detecting harmful words included in the specific utterance based on the plurality of probability values; a harmful word classification unit for classifying the harmful word into at least one of the plurality of unethical classes based on the plurality of probability values for the harmful word; and a harmful word enhancer for generating new learning data by augmenting the harmful word by applying a predefined syllable transformation method to the harmful word based on the probability value of at least one unethical class classified in the harmful word In, it is possible to provide a learning data augmentation device.

Another embodiment of the present invention provides a method for augmenting learning data for harmful words, by inputting a specific utterance included in the learning data into a pre-learned ethical model, and a plurality of preset vocabulary and phrases included in the specific utterance calculating a plurality of probability values to belong to each of the unethical classes of ; detecting harmful words included in the specific utterance based on the plurality of probability values; classifying the harmful word into at least one of the plurality of unethical classes based on the plurality of probability values for the harmful word; and generating new learning data by augmenting the harmful words by applying a predefined syllable transformation method to the harmful words based on the probability value of at least one unethical class classified in the harmful words, A method of augmenting learning data may be provided.

Another embodiment of the present invention is a computer program stored in a computer-readable recording medium including a sequence of instructions for augmenting learning data for harmful words, wherein the computer program is stored in the learning data when executed by a computing device. By inputting the included specific utterances into the pre-learned ethical model, a plurality of probability values for the vocabulary and phrases included in the specific utterances to belong to each of a plurality of preset unethical classes are calculated, and based on the plurality of probability values, the specific utterances are calculated. detects a harmful word included in , classifies the harmful word into at least one of the plurality of unethical classes based on the plurality of probability values for the harmful word, and at least one unethical class classified into the harmful word It is possible to provide a computer program stored in a computer-readable recording medium comprising a sequence of instructions for generating new learning data by augmenting the harmful word by applying a predefined syllable transformation to the harmful word based on the probability value of .

The above-described problem solving means are merely exemplary, and should not be construed as limiting the present invention. In addition to the exemplary embodiments described above, there may be additional embodiments described in the drawings and detailed description.

According to any one of the above-described problem solving means of the present invention, a harmful word included in a specific utterance can be detected based on the detailed classification standard for the harmful word, and the harmful word can be classified according to the detailed classification standard . An ethical model can be trained based on the harmful words classified according to the detailed classification criteria. Therefore, it is possible to provide a learning data augmentation device in which an ethical model for detecting harmful words can be applied to various services and used flexibly.

In addition, at least one harmful word may be augmented based on the harmful word included in the specific utterance. The augmented harmful word may have a shape modified according to a predefined syllable modification method, or some shapes may be omitted. It is possible to learn an ethics model using the augmented harmful words, and thus, it is possible to provide a learning data augmentation apparatus capable of accurately detecting various types of harmful words included in a specific utterance through the learned ethical model.

1 is a block diagram of an apparatus for augmenting learning data according to an embodiment of the present invention.
2 is an exemplary diagram for explaining a process of pre-learning an ethics model according to an embodiment of the present invention.
3 is an exemplary view for explaining an ethical model according to an embodiment of the present invention.
4 and 5 are exemplary views for explaining a process of augmenting a harmful word according to an embodiment of the present invention.
6 and 7 are exemplary views for explaining a process of augmenting harmful phrases according to an embodiment of the present invention.
8 is a flowchart of a learning data augmentation method according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement them. However, the present invention may be embodied in several different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification, when a part is "connected" with another part, this includes not only the case of being "directly connected" but also the case of being "electrically connected" with another element interposed therebetween. . In addition, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components, unless otherwise stated, and one or more other features However, it is to be understood that the existence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded in advance.

In this specification, a "part" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. In addition, one unit may be implemented using two or more hardware, and two or more units may be implemented by one hardware.

Some of the operations or functions described as being performed by the terminal or device in this specification may be instead performed by a server connected to the terminal or device. Similarly, some of the operations or functions described as being performed by the server may also be performed in a terminal or device connected to the server.

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

1 is a block diagram of an apparatus for augmenting learning data according to an embodiment of the present invention. Referring to FIG. 1 , the learning data augmentation apparatus 100 may include a harmful word detection unit 110 , a harmful word classification unit 120 , a harmful word augmentation unit 130 , and a learning unit 140 . However, the above components 110 to 140 are merely illustrative of components that can be controlled by the learning data augmentation apparatus 100 .

Each component of the learning data augmentation apparatus 100 of FIG. 1 , for example, the harmful word detection unit 110 , the harmful word classification unit 120 , the harmful word augmentation unit 130 , and the learning unit 140 simultaneously Alternatively, they may be connected at intervals of time.

The learning data augmentation apparatus 100 may detect harmful words included in a specific utterance based on the detailed classification criteria for the harmful words, and classify the corresponding harmful words according to the detailed classification criteria. For example, the learning data augmentation apparatus 100 uses at least one of a plurality of preset unethical classes including 'profanity', 'insult', 'violence/danger' and 'obscenity' for harmful words detected in a specific utterance. can be classified as either one.

The learning data augmentation apparatus 100 may learn an ethical model based on harmful words classified according to detailed classification criteria. Accordingly, the learning data augmentation apparatus 100 may be flexibly utilized by applying the learned ethical model to various services.

The learning data augmentation apparatus 100 may augment at least one harmful word based on a harmful word included in a specific utterance. For example, the learning data augmentation apparatus 100 augments the harmful word by changing the shape of the harmful word, omitting some forms, or changing the order of words constituting the harmful word according to a predefined syllable modification method can do. The learning data augmentation apparatus 100 may learn an ethical model using the augmented harmful words.

Accordingly, the learning data augmentation apparatus 100 may accurately detect various types of harmful words included in a specific utterance through the learned ethics model.

Hereinafter, each configuration of the learning data augmentation apparatus 100 will be described.

The harmful word detection unit 110 may input a specific utterance included in the learning data into a pre-learned ethical model to calculate a plurality of probability values that the vocabulary and phrase included in the specific utterance will belong to each of a plurality of preset unethical classes. . For example, multiple unethical classes are subdivided into profanity, insult, violence/threat, obscenity, crime/promoting, gender, age, racial origin, sexual orientation, disability, religion, political affiliation, and other hate, for a total of 13 unethical classes. It can be set as an ethics class.

Specifically, the harmful word detection unit 110 may divide a specific utterance into words. The harmful word detection unit 110 may input each divided word into the pre-learned ethical model to calculate a plurality of probability values for at least one word, and detect harmful words included in a specific utterance based on the plurality of probability values. can do.

For example, the harmful word detection unit 110 may divide a specific utterance, for example, 'crime committed in cyberspace' into 'cyber', 'in space', 'doing' and 'crime', respectively, and Each word, 'cyber', 'in space', 'doing' and 'crime' can be entered into the pre-learned ethical model. The harmful word detection unit 110 may calculate a probability value that at least one word, for example, 'crime' belongs to a total of 13 unethical classes, respectively, and detect whether 'crime' is a harmful word from the calculated probability value. can

The harmful word classification unit 120 may classify the harmful word into at least one of a plurality of unethical classes based on a plurality of probability values for the harmful word. For example, when a harmful word is detected in a specific utterance, for example, when the word 'crime' is detected as a harmful word, the harmful word classification unit 120 may determine that 'crime' belongs to 13 unethical classes. Each probability value Based on this, 'crime' can be classified as a crime-promoting class among 13 unethical classes.

In this way, the learning data augmentation apparatus 100 may detect harmful words included in a specific utterance based on the unethical class subdivided into categories, and a probability value of how much the detected harmful words belong to each subdivided category. By calculating, harmful words included in specific utterances can be classified into any one of the subdivided categories. Accordingly, the learning data augmentation apparatus 100 may be applied to services in various fields, and may sensitively detect corresponding harmful words in each service field.

The harmful word enhancer 130 may generate new learning data by augmenting the harmful word by applying a predefined syllable transformation method to the harmful word based on the probability value of at least one unethical class classified into the harmful word. For example, when a harmful word is detected in a specific utterance, for example, when the word 'crime' is detected as a harmful word, the harmful word augmentation unit 130 is based on the probability value of the crime promoting class into which the 'crime' is classified. Harmful words for 'crime' can be reinforced by applying a predefined syllable transformation method to 'crime'.

The predefined syllable modification method according to the present invention may include, for example, the first to fifth methods. The first method may separate the initial, neutral, or final consonant included in the harmful word, and the second method may replace the vowel of the harmful word with a similar vowel. For example, according to the first method, by separating the initial, neuter, or final consonant included in 'crime', it is possible to generate 'vㅓㅁㅗㅣ', and according to the second method, the vowel of 'crime' is replaced with a similar vowel. So you can create a 'beom'.

In the third method, consonants and vowels of the harmful words may be substituted with similar characters, in the fourth method, characters may be inserted between the harmful words, and in the fifth method, the arrangement order of the harmful words may be changed. For example, according to the third method, 'beomjo|' can be created by replacing the consonants and vowels of 'crime' with similar characters, and according to the fourth method, 'crime @ crime' by inserting a letter between 'crime' ', and by changing the arrangement order of 'crime' according to the fifth method, 'criminal' may be generated.

When the detected harmful word is a word, and the probability value of any one unethical class having the highest probability value of at least one unethical class classified in the word is greater than or equal to a preset probability value, the first method to At least one of the fifth methods may be applied.

For example, the harmful word augmentation unit 130 compares the probability value and the preset probability value of the crime promotion class for which the highest probability value among the crime promotion class, other hate class, and violence threat class classified for the harmful word 'crime' is calculated. can be compared The harmful word enhancer 130 may apply at least one of the predefined syllable transformation method and the first to fifth methods according to the comparison result of the probability value of the crime promotion class and the predetermined probability value.

Also, the harmful word enhancer 130 may apply a predefined syllable modification method according to the letter form of the word, that is, at least one of the first to fifth methods. The harmful word enhancer 130 may apply all of the first to fifth methods when all the letters of a word consist only of a leading consonant and a neutral.

The harmful word reinforcement unit 130 may apply the first method, the fourth method, and the fifth method when all the letters of a word are made up of a beginning, a neutral, and a final. When at least one is present, the first to third methods may be applied.

The harmful word augmentation unit 130 may build a harmful word dictionary in which an augmented word for each word and a probability value of at least one unethical class classified in the word are mapped and stored. For example, the harmful word augmentation unit 130 is a word augmented to the word 'crime', Probability values of unethical classes, crime promotion class/0.7, other hate classes/0.5, and violent threat class/0.4 are mapped to build a stored hazard dictionary.

When the detected harmful word is a phrase, the harmful word enhancer 130 may determine whether the phrase is included in the harmful word dictionary. When the detected harmful word is a phrase and the phrase is not included in the harmful word dictionary, the harmful word reinforcement unit 130 may enhance the phrase by applying at least one of the first to fifth methods to the phrase. have.

For example, when the detected harmful word is a phrase, for example, when the phrase 'Let's block the harmful environment' is detected as a harmful word, the harmful word enhancement unit 130 may divide the phrase into each word, It can be determined whether the divided word word is included in the harmful word dictionary. When a word divided from the corresponding phrase is not included in the harmful word dictionary, the harmful word reinforcement unit 130 enhances the corresponding phrase based on the syllable transformation method and the first to fifth methods predefined for the corresponding phrase. can do.

When the sum of the probability values of the plurality of upper unethical classes of the predetermined number among the probability values of at least one unethical class classified in the phrase is equal to or greater than a preset threshold, the harmful word reinforcement unit 130 is configured to modify a phrase defined in the phrase. You can reinforce the phrase by applying .

For example, the harmful word augmentation unit 130 is the detected harmful phrase, for example, among the probability values of the unethical class for 'Let's block the harmful environment' The probability value of the class and the probability value of other hate classes may be summed, and the sum of the probability values of the top three unethical classes may be compared with a preset threshold. When the sum of the probability values of the top three unethical classes, 0.7, is greater than a preset threshold (eg, 0.5), the harmful word reinforcement unit 130 applies a predefined phrase transformation method to the detected harmful phrase to improve the corresponding phrase. can be augmented

The predefined phrase modification method according to the present invention may include sixth to eighth methods. The sixth method may replace words included in the phrase with similar words, and the seventh method may change the arrangement order of words in the phrase. The eighth method may delete a word included in the phrase. For example, according to the sixth method, the word included in 'Let's block the harmful environment' may be replaced with a similar word to generate 'Let's exclude the harmful environment', and according to the seventh method, 'Block the harmful environment' 'Let's block harmful environment' can be created by changing the arrangement order of words in 'Let's do' have.

The harmful word enhancer 130 may input a phrase modified according to a predefined method of changing the phrase into an ethical model to calculate a plurality of probability values that the modified phrase belongs to each of a plurality of preset unethical classes.

For example, the harmful word augmentation unit 130 inputs the modified phrases, 'let's exclude harmful environments', 'let's cut off environmental hazards' and 'let's cut harm' into the ethical model, so that each modified phrase has 13 ratios. Probability values belonging to each ethical class can be calculated.

When the sum of the predetermined number of the upper plurality of unethical class probability values is equal to or greater than a predetermined threshold, the harmful word enhancer 130 is a modified phrase for each phrase, the probability value of at least one unethical class classified into the modified phrase is mapped It is possible to build a dictionary of harmful phrases that have been saved.

For example, the harmful word augmentation unit 130 determines that the sum (eg, 0.7) of the probability values of the top three unethical classes with respect to the modified phrase, 'let's exclude the harmful environment' is higher than a preset threshold (eg, 0.5). In the large case, a modified phrase for the phrase 'Let's block the harmful environment', 'Let's exclude the harmful environment' and the probability value of the crime promotion class (0.7), the probability value of the violent threat class (0.4), and the probability value of the other hate class (0.3) ) can be mapped and stored to build a dictionary of harmful phrases.

The learning unit 140 may pre-lear a generation model (Generator) to transform at least one token among a plurality of tokens generated from a specific utterance into a fake token, and a discriminator is a fake included in the plurality of tokens. It can be pre-trained to discriminate tokens. For example, the learner 140 may pre-lear the generation model and the discrimination model using a transformer encoder.

The ethical model according to the present invention may include a discrimination model and a feed forward neural network. For example, the learning unit 140 may generate an ethics model by adding a recurrent neural network to the pre-trained discrimination model. The learning unit 140 may perform fine tuning by classifying the pre-trained discrimination model for each token.

The learning unit 140 may learn the ethical model by updating the recurrent neural network based on the harmful word dictionary and the harmful phrase dictionary. For example, the learning unit 140 may re-learn the ethical model by using the augmented harmful words.

Accordingly, the learning data augmentation apparatus 100 may accurately detect various types of harmful words included in a specific utterance. For example, the learning data augmentation apparatus 100 may accurately detect harmful words in which a form included in a specific utterance is changed or some forms are omitted.

2 is an exemplary diagram for explaining a process of pre-learning an ethical model according to an embodiment of the present invention. Referring to FIG. 2 , the training data augmentation apparatus 100 may pre-lear the generation model 230 and the discrimination model 250 .

For example, the training data augmentation apparatus 100 may generate a plurality of tokens 210 from a specific utterance, and input the plurality of tokens 210 to the generation model 230 to include the plurality of tokens 210 . One or more tokens can be transformed into fake tokens. The training data augmentation apparatus 100 may pre-lear the generation model 230 to generate a plurality of tokens 240 including fake tokens from the plurality of tokens 210 .

Specifically, the training data augmentation apparatus 100 may determine a set of positions to be masked with respect to the plurality of tokens 210 generated from a specific utterance. For example, a position to be masked among the plurality of tokens 210 is an integer between 1 and n, and may be expressed as Equation 1 below.

<Equation 1>

In Equation 1, m means a position to perform masking within the plurality of tokens 210 , and k means the number of tokens to perform masking in the plurality of tokens 210 . For example, the number k to be masked may be 0.15n, and the training data augmentation apparatus 100 may perform masking on about 15% of the tokens in the total of the plurality of tokens 210 . A token at a position determined to be masked within the plurality of tokens 210 may be expressed as in Equation 2 below.

<Equation 2>

For example, the training data augmentation apparatus 100 may mask the first token and the third token from the plurality of tokens 210 including {the, chef, cooked, the, meal}, and the first A plurality of tokens 220, {[MASK], chef, [MASK], the, meal} in which the token and the third token are masked may be generated.

For example, the training data augmentation apparatus 100 may input a plurality of masked tokens 220 into the generation model 230 to predict what the masked tokens are. The generative model 230 can be expressed by Equation 3 below.

<Equation 3>

In Equation 3, e(·) means embedding, and through Equation 3, the output layer and the embedding layer of the generative model 230 may share a weight.

For example, the training data augmentation apparatus 100 may predict the meaning of some masked tokens included in the plurality of masked tokens 220 through the generation model 230 using Equation 3 . The training data augmentation apparatus 100 may generate a fake token by predicting the meaning of some masked tokens from the plurality of masked tokens 220 . That is, the training data augmentation apparatus 100 may generate a plurality of tokens 240 including fake tokens by inputting a plurality of masked tokens 220 into the generation model 230 .

For example, the training data augmentation apparatus 100 includes a fake token by inputting a plurality of masked tokens 220, {[MASK], chef, [MASK], the, meal} to the generation model 230 . It is possible to generate a plurality of tokens 240, {the, chef, ate, the, meal}. Generation of a fake token may be expressed by Equation 4 below.

<Equation 4>

Specifically, the training data augmentation apparatus 100 may predict and output the masked first token through the generation model 230 as the same 'the' token as the first, and display the masked third token different from the first. You can output the replaced token by predicting with the 'ate' token. That is, the learning data augmentation apparatus 100 may generate a plurality of tokens 240 including fake tokens in which the third token is substituted through the generation model 230 .

For example, the training data augmentation apparatus 100 may determine whether a fake token is included by inputting a plurality of tokens 240 including a fake token into the determination model 250 . That is, the training data augmentation apparatus 100 may pre-lear the determination model 250 to determine the fake tokens included in the plurality of tokens 240 including the fake tokens.

Specifically, the determination model 250 distinguishes whether each token included in the plurality of tokens 240 including a fake token is a normal (original) token or a fake (replaced) token through the generation model 230. It can be pre-trained by binary classification. Here, the general token means the same token as the first, and the fake token means a token replaced through the generation model 230 unlike the first.

For example, the training data augmentation apparatus 100 may pre-train the generative model 230 and the discriminant model 250 using a loss function. The loss function may be expressed as in Equation 5 below.

<Equation 5>

Referring to Equation 5, the loss function may be a mean square error (MSE). For example, the training data augmentation apparatus 100 may perform pre-learning on the generative model 230 and the discriminant model 250 using a loss function as in Equation 5 .

3 is an exemplary view for explaining an ethical model according to an embodiment of the present invention. Referring to FIG. 3 , the training data augmentation apparatus 100 may generate the ethical model 300 based on the pre-learning completed discrimination model 320 .

For example, the training data augmentation apparatus 100 may generate the ethical model 300 by adding the recurrent neural network 330 to the pre-trained discrimination model 320 . The training data augmentation apparatus 100 may perform fine tuning in such a way that the pre-trained discrimination model 320 classifies classes for each token.

For example, the learning data augmentation apparatus 100 inputs a plurality of tokens 310 generated from a specific utterance into the ethics model 300 so that each token is one or more of a plurality of preset unethical classes 340 . can be classified. Here, a plurality of preset unethical classes 340 are, for example, profanity / insult / threat of violence / obscenity / crime promotion / gender / age / racial origin / sexual orientation / disability / religion / political orientation / other hate and Likewise, it can be subdivided into a total of 13 unethical classes 340 .

For example, the learning data augmentation apparatus 100 inputs a plurality of tokens 310 generated from a specific utterance, for example, {the, chef, cooked, the, meal} into the ethical model 300, each token, As an example, a probability value that the {meal} token belongs to the 13 unethical classes 340 may be calculated, respectively. The learning data augmentation apparatus 100 may classify each token, for example, {meal} token, into other hate classes based on the probability value.

4 and 5 are exemplary views for explaining a process of augmenting a harmful word according to an embodiment of the present invention. Referring to FIG. 4 , when the detected harmful word is a word, the learning data augmentation apparatus 100 provides the highest probability value (x) of the unethical class among the probability values of the unethical class for the word and a preset probability value (eg, 0.5). ) can be compared (S410).

For example, the learning data augmentation apparatus 100 is the highest crime promotion class among the detected harmful words, the probability value of the crime promoting class for 'crime (eg 0.7), and the probability value (eg, 0.4) of other hate classes. A probability value (eg, 0.7) may be compared with a preset probability value (eg, 0.5).

The learning data augmentation apparatus 100 may determine whether all letters of a word are made up of only the initial and neutral characters when the probability value of the unethical class for the word, x, is greater than a preset probability value, for example, 0.5 (S411) ( S420).

For example, when the learning data augmentation apparatus 100 has a probability value (eg, 0.7) of a crime promotion class for 'crime' is greater than a preset probability value (eg, 0.5), all letters of 'crime' are initial and neutral It can be determined whether it consists only of

The learning data augmentation apparatus 100 may apply all of the first to fifth methods among the predefined syllable transformation methods when all letters of a word consist only of a leading consonant and a neutral (S421) (S430).

For example, the learning data augmentation apparatus 100 may separate the initial, neutral, or final consonants included in 'idiot' when the letters of a word consist only of a beginning and a neutral, such as 'idiot', and 'idiot' You can replace the vowels included in 'Fool' with similar vowels, replace the consonants and vowels included in 'Fool' with similar characters, insert a letter between 'Fool', and change the arrangement order of 'Fool'. By changing, it is possible to augment learning data, such as 'bao', 'bye', 'vrbo', 'ba@bo' and 'boba' for 'idiot'.

On the other hand, when all the letters of a word do not consist of only the initial consonant and neuter (S422), the learning data augmentation apparatus 100 may determine whether all the letters of the word are composed of the initial consonant, loyalty, and final consonant (S440).

For example, when all letters, such as 'crime', do not consist of only the initial and neuter, the learning data augmentation apparatus 100 may determine whether 'crime' consists of the initial, neuter, and final consonants.

The learning data augmentation apparatus 100 may apply the first method, the fourth method, and the fifth method among the predefined syllable transformation methods when all the letters of a word are made up of a leading consonant, a middle consonant, and a final consonant (S441) (S450). ).

For example, the learning data augmentation apparatus 100 may separate the initial consonant, neutral or final consonant included in the 'penalty' when all letters, such as 'penal', consist of a leading consonant, a middle consonant, and a final consonant, and between the 'penalty'. It is possible to insert characters in the 'penalty' and change the arrangement order of 'penalty' to augment learning data such as 'hehㅕㅇ ㅅㅓㄹ', 'punishment@punishment' and 'penalty' for 'penalty'.

On the other hand, the learning data augmentation apparatus 100 may determine whether at least one letter consisting only of a leading consonant and a neuter exists in the letters of the word when all the letters of the word do not consist of the initial consonant, the middle consonant and the final consonant (S442). (S460).

For example, the learning data augmentation apparatus 100 may include at least one letter consisting of only a beginning and a neutral in the letters included in 'Unrooted' when all letters do not consist of a leading, neutral, and final consonant as in 'Unrooted'. can determine whether

The learning data augmentation apparatus 100 may apply the first to third methods among the predefined syllable transformation methods when there is at least one letter consisting only of a leading consonant and a neutral in the letter of a word (S461) (S470). ).

For example, the learning data augmentation apparatus 100 may separate the initial, neutral, or final consonant included in 'bad' when there is at least one letter consisting only of a leading consonant and a neutral, such as 'bad', and 'bad' Vowels included in ' can be replaced with similar vowels, and consonants and vowels included in 'bad' are replaced with similar characters, so that 'bad' It is possible to augment learning data such as '.

Referring to FIG. 5 , when the detected harmful word is a word, the learning data augmentation apparatus 100 provides a probability value (x) of the highest unethical class among the probability values of the unethical class for the word and a preset probability value (eg, 0.5). and 0.3) can be compared (S510).

For example, the learning data augmentation apparatus 100 is the highest among the detected harmful words, the probability value of the crime-promoting class for 'idiot' (eg 0.2), and the probability value (eg, 0.4) of the other hate class. A probability value (eg, 0.4) may be compared with a preset probability value (eg, 0.5 and 0.3).

The learning data augmentation apparatus 100 determines whether all the letters of the word consist of only the initial and neutral characters when the probability value of the unethical class for the word, x, is a preset probability value, for example, less than 0.5 but greater than 0.3 (S511). It can be done (S520).

For example, if the learning data augmentation apparatus 100 has a probability value (eg, 0.4) of other hate classes for 'idiot' is less than a preset probability value, for example, less than 0.5 and greater than 0.3, all the letters of 'idiot' are It can be judged whether it consists of only the initial and neutral consonants.

The learning data augmentation apparatus 100 may apply all of the first to fifth methods among the predefined syllable transformation methods when all the letters of a word consist of only a leading consonant and a neutral (S521) (S530).

For example, the learning data augmentation apparatus 100 may separate the initial, neutral, or final consonants included in 'idiot' when the letters of a word consist only of a beginning and a neutral, such as 'idiot', and 'idiot' You can replace the vowels included in 'Fool' with similar vowels, replace the consonants and vowels included in 'Fool' with similar characters, insert a letter between 'Fool', and change the arrangement order of 'Fool'. By changing, it is possible to augment learning data, such as 'bao', 'bye', 'vrbo', 'ba@bo' and 'boba' for 'idiot'.

On the other hand, when all the letters of the word are not made up of only the initial consonant and the neutral (S522), the learning data augmentation apparatus 100 provides the highest probability value of the unethical class among the probability values of the unethical class for the word (x) and A preset number (eg, three) of a plurality of upper unethical class probability values may be compared with a preset probability value (eg, 0.3) and a preset threshold (eg, 0.3) (S540)

For example, the learning data augmentation apparatus 100 provides a probability value of a crime promotion class for 'punishment' (eg 0.2) and a preset probability value ( Example: 0.3), the probability values of the top 3 crime-promoting classes (eg 0.2) for 'punishment', the probability values of the religious class (eg 0.1), and the probability values of other hate classes (eg 0.1) and the A set threshold (eg, 0.3) can be compared.

The learning data augmentation apparatus 100 is a probability value of an unethical class for a word, x is a preset probability value, for example, less than 0.3, but a preset number, for example, the sum of the top three unethical class probability values is a preset threshold, an example is greater than 0.3 (S540), all of the first to fifth methods among the predefined syllable modification methods may be applied (S550).

For example, the learning data augmentation apparatus 100 has a probability value of the crime promotion class for 'punishment', where 0.2 is a predetermined probability value, for example, less than 0.3, but a predetermined number, for example, the probability value of the top three crime promotion classes ( Example: 0.2), the sum of the probability value of the religious class (eg 0.1) and the probability value of other hate classes (eg 0.1), when 0.4 is a preset threshold, for example, greater than 0.3 Alternatively, the last name can be separated, the vowel included in 'penalty' can be replaced with a similar vowel, the consonants and vowels included in the 'penalty' can be replaced with a similar character, and a letter is inserted between the 'penalty' By changing the arrangement order of 'penalty', learning data such as 'heh ㅇㅇㅇㅇ', 'penalty', 'penalty% punishment' and 'penalty' for 'penalty' can be augmented.

6 and 7 are exemplary views for explaining a process of augmenting harmful phrases according to an embodiment of the present invention. Referring to FIG. 6 , when the detected harmful word is a phrase, the learning data augmentation apparatus 100 provides the highest probability value (x) of the unethical class among the probability values of the unethical class for the corresponding phrase and a preset probability value (eg, 0.5). ) can be compared (S610).

For example, the learning data augmentation apparatus 100 has the highest among the detected harmful phrases, the probability value (eg, 0.7) of the crime-promoting class for 'let's exclude the harmful environment', and the probability value (eg, 0.4) of other hate classes. A probability value (eg, 0.7) of the crime promotion class may be compared with a predetermined probability value (eg, 0.5).

The learning data augmentation apparatus 100 may determine whether the corresponding phrase is included in the harmful word dictionary when the probability value of the unethical class for the word, x, is greater than a preset probability value, for example, 0.5 (S611) ( S620).

For example, when the learning data augmentation apparatus 100 has a probability value (eg, 0.7) of the crime promotion class for 'exclude harmful environment' is greater than a preset probability value (eg, 0.5), 'exclude harmful environment' It may be determined whether each phrase included in ', 'harm', 'environment' and 'exclude' are included in the harmful dictionary.

When the corresponding phrase is not included in the harmful word dictionary (S621), the learning data augmentation apparatus 100 applies at least one of the first to fifth methods among the syllable transformation methods predefined for the phrase to generate the phrase. It can be augmented (S630).

For example, if the learning data augmentation apparatus 100 does not include each phrase, 'harm', 'environment', and 'let's exclude' in the harmful word dictionary, each phrase, 'harmful', 'environment' and separating the initial, neuter, or final consonant included in ‘let’s exclude’, replacing vowels with similar vowels, replacing consonants and vowels with similar letters, inserting letters between each phrase, or changing the arrangement order within each phrase can be changed

Referring to FIG. 7 , when the detected harmful word is a phrase, the learning data augmentation apparatus 100 provides a preset number (eg, three) of a plurality of upper unethical class probability values for the corresponding phrase and a preset threshold ( Example: 0.3) can be compared (S710).

For example, the learning data augmentation apparatus 100 may include a detected harmful phrase, a probability value of the top three crime promotion classes for 'let's exclude a harmful environment' (eg 0.3), a probability value of a violent threat class (eg 0.1) and the sum (eg, 0.6) of the probability values (eg, 0.2) of other hate classes may be compared with a preset threshold (eg, 0.3).

The learning data augmentation apparatus 100 performs a preset number of phrases, for example, when the sum of the top three unethical class probability values is greater than a preset threshold, for example 0.3 (S711), a phrase transformation method defined in the phrase It can be applied to augment the phrase (S720).

For example, the learning data augmentation apparatus 100 provides a probability value of the top three crime promotion classes for 'let's exclude harmful environments' (eg 0.3), a probability value of a violent threat class (eg 0.1), and other hate classes. When the sum of probability values (eg 0.2) (eg 0.6) is greater than a preset threshold (eg 0.3), words included in 'Let's exclude harmful environments' can be substituted with similar words, and word order can be changed, and by deleting the words included in each phrase, 'Let's exclude harmful environments' for 'Let's exclude harmful environments' can be augmented

The learning data augmentation apparatus 100 may input the modified phrase into the ethical model to calculate a plurality of probability values that the modified phrase belongs to each of a plurality of preset unethical classes ( S730 ).

For example, the learning data augmentation apparatus 100 inputs the modified phrases, 'let's exclude harmful environment', 'block environmental harm' and 'let's block harm' into each ethical model, so that each phrase contains 13 unethical sentences. Each probability value belonging to a class can be calculated.

The learning data augmentation apparatus 100 may compare the sum of a predetermined number (eg, three) of a plurality of upper unethical class probability values for the modified phrase and a predetermined threshold (eg, 0.3) ( S740 ).

For example, the learning data augmentation apparatus 100 is a modified phrase, for example, a probability value (eg 0.2) of the top three crime promotion classes for 'let's exclude a harmful environment', a probability value of a violence threat class (eg: 0.1) and the sum (eg, 0.5) of the probability values (eg, 0.2) of other hate classes may be compared with a preset threshold (eg, 0.3).

The learning data augmentation apparatus 100 provides a preset number of modified phrases, for example, when the sum of the top three unethical class probability values is greater than a preset threshold, for example, 0.3 (S741), each phrase transformed into a modified phrase, transformation Probability values of at least one unethical class classified to the old phrases are mapped to build a stored harmful phrase dictionary (S750).

For example, the learning data augmentation apparatus 100 is a modified phrase, for example, a probability value (eg 0.2) of the top three crime promotion classes for 'let's exclude a harmful environment', a probability value of a violence threat class (eg: If the sum (eg 0.5) of the probability value (eg 0.2) of 0.1) and other hate classes is greater than a preset threshold (eg 0.3), a modified phrase in the harmful phrase for 'Let's block the harmful environment', ' Probability values of crime promotion class (eg 0.2), threat of violence class (eg 0.1), and probability values of other hate classes (eg 0.2) for 'let's exclude harmful environment' and modified phrases are mapped and stored. You can build a dictionary.

8 is a flowchart of a learning data augmentation method according to an embodiment of the present invention. The learning data augmentation method illustrated in FIG. 8 includes steps processed in time series according to the embodiments illustrated in FIGS. 1 to 7 . Therefore, even if omitted below, it is also applied to the method of augmenting learning data for harmful words in the apparatus for augmenting learning data according to the embodiment shown in FIGS. 1 to 7 .

In step S810, the training data augmentation apparatus inputs a specific utterance included in the training data into the pre-learned ethical model to calculate a plurality of probability values that vocabulary and phrases included in the specific utterance will belong to each of a plurality of preset unethical classes. have.

In step S820, the learning data augmentation apparatus may detect harmful words included in a specific utterance based on a plurality of probability values.

In step S830, the learning data augmentation apparatus may classify the harmful word into at least one of a plurality of unethical classes based on a plurality of probability values for the harmful word.

In step S840, the learning data augmentation apparatus may generate new learning data by augmenting the harmful word by applying a predefined syllable transformation method to the harmful word based on the probability value of at least one unethical class classified into the harmful word.

In the above description, steps S810 to S840 may be further divided into additional steps or combined into fewer steps according to an embodiment of the present invention. In addition, some steps may be omitted as needed, and the order between the steps may be switched.

The method for augmenting learning data for harmful words in the learning data augmentation apparatus described through FIGS. 1 to 8 is a computer program stored in a computer-readable recording medium executed by a computer or a record including instructions executable by the computer It can also be implemented in the form of a medium. In addition, the method for augmenting learning data for harmful words in the learning data augmentation apparatus described with reference to FIGS. 1 to 8 may be implemented in the form of a computer program stored in a computer-readable recording medium that is executed by a computer.

A computer-readable recording medium may be any available medium that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. In addition, the computer-readable recording medium may include a computer storage medium. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The description of the present invention described above is for illustration, and those of ordinary skill in the art to which the present invention pertains can understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. For example, each component described as a single type may be implemented in a dispersed form, and likewise components described as distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the following claims rather than the above detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. do.

100: learning data augmentation device
110: harmful fish detection unit
120: harmful fish classification unit
130: harmful fish reinforcement unit
140: study unit

Claims (25)

In the device for augmenting learning data for harmful words,
A specific utterance included in the learning data is input to the previously learned ethical model to calculate a plurality of probability values that the vocabulary and phrase included in the specific utterance will belong to each of a plurality of preset unethical classes, and based on the plurality of probability values a harmful word detection unit detecting harmful words included in the specific utterance;
a harmful word classification unit for classifying the harmful word into at least one of the plurality of unethical classes based on the plurality of probability values for the harmful word; and
A harmful word reinforcement unit for generating new learning data by augmenting the harmful word by applying a predefined syllable transformation method to the harmful word based on the probability value of at least one unethical class classified in the harmful word
includes,
The predefined syllable transformation method is,
a first method of separating the initial, neutral, or final consonant included in the harmful fish; a second method of replacing a vowel of the harmful word with a similar vowel; a third method of replacing consonants and vowels of the harmful words with similar characters; a fourth method of inserting a character between the harmful words; and at least one of a fifth method of changing the arrangement order of the harmful words,
The harmful word reinforcement unit,
When the detected harmful word is a word, and the probability value of any one unethical class having the highest probability value of at least one unethical class classified in the word is greater than or equal to a preset probability value, the first method to the fifth method Reinforcing the word by applying at least one method of
constructing a harmful word dictionary in which the augmented word and the probability value of at least one unethical class classified in the word are mapped and stored;
When the detected harmful word is a phrase and the phrase is not included in the harmful word dictionary, the phrase is reinforced by applying at least one of the first to fifth methods to the phrase,
When the sum of the probability values of a predetermined number of upper plural unethical classes among the probability values of at least one unethical class classified in the phrase is equal to or greater than a predetermined threshold, the phrase is converted to the phrase by applying a predefined phrase transformation method to the phrase A learning data augmentation device that is augmented.
The method of claim 1,
The harmful word detection unit,
dividing the specific utterance into words, inputting each divided word into the pre-learned ethical model, calculating the plurality of probability values for at least one word, and detecting harmful words included in the specific utterance , a learning data augmentation device.
delete delete The method of claim 1,
The harmful word reinforcement unit,
When all the letters of the word consist only of a leading consonant and a neutral, all of the first to fifth methods are applied, and when all the letters of the word consist of a leading consonant, a neutral and a final consonant, the first method and the first method Method 4 and the fifth method are applied, and when at least one letter consisting of only a leading consonant and a neutral exists in the letters of the word, the first method to the third method are applied, the learning data augmentation apparatus.
delete delete delete The method of claim 1,
The predefined phrase modification method comprises:
a sixth method of substituting a word included in the phrase with a similar word; a seventh method of changing the arrangement order of words within the phrase; and at least one of an eighth method of deleting a word included in the phrase.
10. The method of claim 9,
The harmful word reinforcement unit,
A plurality of probability values are calculated that the modified phrase belongs to each of the plurality of preset unethical classes by inputting a phrase modified according to the predefined phrase modification method into the ethical model,
When the sum of the predetermined number of the upper plurality of unethical class probability values is equal to or greater than a predetermined threshold, the probabilities of the modified phrase and at least one unethical class classified in the modified phrase are mapped and stored for each of the phrases. To build a, learning data augmentation device.
11. The method of claim 10,
Learning unit to learn the above ethical model
further comprising,
The learning unit pre-learns a generation model (Generator) to transform at least one token among a plurality of tokens generated from the specific utterance into a fake token, and a discrimination model (Discriminator) selects the fake tokens included in the plurality of tokens A learning data augmentation device that pre-learns to discriminate.
12. The method of claim 11,
The ethical model includes the discriminant model and a feed forward neural network,
The learning unit,
The learning data augmentation apparatus of claiming that the ethical model is learned by updating the recurrent neural network based on the harmful word dictionary and the harmful phrase dictionary.
A method of augmenting learning data for harmful words performed by a learning data augmentation device, the method comprising:
inputting a specific utterance included in the training data into a pre-learned ethical model, and calculating a plurality of probability values that the vocabulary and phrase included in the specific utterance will belong to each of a plurality of preset unethical classes;
detecting harmful words included in the specific utterance based on the plurality of probability values;
classifying the harmful word into at least one of the plurality of unethical classes based on the plurality of probability values for the harmful word; and
generating new learning data by augmenting the harmful word by applying a predefined syllable transformation method to the harmful word based on the probability value of at least one unethical class classified in the harmful word
includes,
The predefined syllable transformation method is,
a first method of separating the initial, neutral, or final consonant included in the harmful fish; a second method of replacing a vowel of the harmful word with a similar vowel; a third method of replacing consonants and vowels of the harmful words with similar characters; a fourth method of inserting a character between the harmful words; and at least one of a fifth method of changing the arrangement order of the harmful words,
The step of generating new learning data by augmenting the harmful words,
When the detected harmful word is a word, and the probability value of any one unethical class having the highest probability value of at least one unethical class classified in the word is greater than or equal to a preset probability value, the first method to the fifth method augmenting the word by applying at least one of the methods;
constructing a harmful word dictionary in which the augmented word and the probability value of at least one unethical class classified in the word are mapped to each of the words;
when the detected harmful word is a phrase and the phrase is not included in the harmful word dictionary, augmenting the phrase by applying at least one of the first to fifth methods to the phrase; and
When the sum of the probability values of a predetermined number of upper plural unethical classes among the probability values of at least one unethical class classified in the phrase is equal to or greater than a predetermined threshold, the phrase is converted to the phrase by applying a predefined phrase transformation method to the phrase A method of augmenting learning data, comprising the step of augmenting.
14. The method of claim 13,
The step of detecting the harmful fish,
dividing the specific utterance into words;
calculating the plurality of probability values for at least one word by inputting each divided word into the pre-learned ethical model; and
detecting harmful words included in the specific utterance
That comprising a, learning data augmentation method.
delete delete 14. The method of claim 13,
The step of generating new learning data by augmenting the harmful words,
When all the letters of the word consist only of a leading consonant and a neutral, all of the first to fifth methods are applied, and when all the letters of the word consist of a leading consonant, a neutral and a final consonant, the first method and the first method applying the fourth method and the fifth method, and applying the first to third methods when there is at least one letter consisting only of a leading consonant and a neutral in the letters of the word
Which further comprises, the learning data augmentation method.
delete delete delete 14. The method of claim 13,
The predefined phrase modification method comprises:
a sixth method of substituting a word included in the phrase with a similar word; a seventh method of changing the arrangement order of words within the phrase; and at least one of an eighth method of deleting a word included in the phrase.
22. The method of claim 21,
The step of generating new learning data by augmenting the harmful words,
calculating a plurality of probability values that the modified phrase belongs to each of the plurality of preset unethical classes by inputting a phrase modified according to the predefined phrase modification method into the ethical model; and
When the sum of the predetermined number of the upper plurality of unethical class probability values is equal to or greater than a predetermined threshold, the probabilities of the modified phrase and at least one unethical class classified in the modified phrase are mapped and stored for each of the phrases. steps to build
Which further comprises, the learning data augmentation method.
23. The method of claim 22,
learning the ethical model
further comprising,
The step of learning the ethical model is,
pre-learning, by a generation model (Generator), to transform at least one token among a plurality of tokens generated from the specific utterance into a fake token; and
Pre-training a discriminator to discriminate the fake tokens included in the plurality of tokens
That comprising a, learning data augmentation method.
24. The method of claim 23,
The ethical model includes the discriminant model and a feed forward neural network,
The step of learning the ethical model is,
Learning the ethical model by updating the recurrent neural network based on the harmful word dictionary and the harmful phrase dictionary
Which further comprises, the learning data augmentation method.
A computer program stored in a computer-readable recording medium including a sequence of instructions for augmenting learning data for harmful words,
When the computer program is executed by a computing device,
By inputting a specific utterance included in the learning data into a pre-learned ethical model, a plurality of probability values that the vocabulary and phrase included in the specific utterance will belong to each of a plurality of preset unethical classes are calculated,
Detecting harmful words included in the specific utterance based on the plurality of probability values,
Classifying the harmful word into at least one of the plurality of unethical classes based on the plurality of probability values for the harmful word,
Based on the probability value of at least one unethical class classified in the harmful word, a predefined syllable transformation method is applied to the harmful word to augment the harmful word to generate new learning data,
The predefined syllable modification method is,
a first method of separating the initial, neutral, or final consonant included in the harmful fish; a second method of substituting a vowel of the harmful word with a similar vowel; a third method of replacing consonants and vowels of the harmful words with similar characters; a fourth method of inserting a character between the harmful words; and at least one of a fifth method of changing the arrangement order of the harmful words,
Generating new learning data by augmenting the harmful words,
When the detected harmful word is a word, and the probability value of any one unethical class having the highest probability value of at least one unethical class classified in the word is greater than or equal to a preset probability value, the first method to the fifth method Reinforcing the word by applying at least one method of
For each word, the augmented word and the probability value of at least one unethical class classified in the word are mapped to build a stored harmful dictionary,
When the detected harmful word is a phrase and the phrase is not included in the harmful word dictionary, the phrase is reinforced by applying at least one of the first to fifth methods to the phrase,
When the sum of the probability values of a predetermined number of upper plural unethical classes among the probability values of at least one unethical class classified in the phrase is equal to or greater than a predetermined threshold, the phrase is converted to the phrase by applying a predefined phrase transformation method to the phrase A computer program stored on a computer-readable medium comprising a sequence of augmenting instructions.

Images