KR102131641B1 - Detection metohd of fake news using grammatic transformation on neural network, computer readable medium and apparatus for performing the method - Google Patents

Abstract

Disclosed is a method for searching for fake news using grammatical transformation on a neural network, and a recording medium and apparatus for performing the same. The method of searching for fake news using grammatical transformation on a neural network includes embedding a proposition sentence in which a true or false is determined by a news sentence as a word vector, and inputting the word vector into a predetermined natural language processing neural network to context (context) generating a vector, inputting the context vector into a predetermined natural language processing neural network, generating a candidate sentence having the same meaning but different grammar, and comparing the candidate sentence with the news sentence And determining whether the propositional sentence is true or false and searching whether the news sentence corresponds to fake news.

Description

DETECTION METOHD OF FAKE NEWS USING GRAMMATIC TRANSFORMATION ON NEURAL NETWORK, COMPUTER READABLE MEDIUM AND APPARATUS FOR PERFORMING THE METHOD}

The present invention relates to a fake news search method using a grammatical modification on a neural network, a recording medium and a device for performing the same, and more specifically, to generate a sentence based on deep learning, and to use this to determine whether the news is authentic The present invention relates to a method for searching for fake news using grammatical transformation on a neural network, and a recording medium and device for performing the same.

With the advent of the mobile era in which high-speed mobile communication networks and smartphones are widely spread, the use of various social networking services (SNS) is rapidly increasing. In particular, in recent years, the use of SNS such as blogs, KakaoTalk, Line, Facebook, Twitter, Instagram and Tumblr has skyrocketed. Meanwhile, the delivery of information or news through various SNS is also explosively increasing.

However, it is also a reality that whenever there is a political event such as an election, a lot of misinformation or news spread through various SNS. In addition, more seriously, it is a frequent occurrence of misinformation or news propagation through various SNS intentionally for a specific purpose. For this reason, the delivery of information or news through SNS may become an important social issue in the future. It can be said that this is high.

On the other hand, fake news, which is news deliberately manipulated with a specific purpose as described above, is mostly searched by a person, and the authenticity is determined. This requires a lot of time and effort, so it does not require human judgment and requires a new fake news search model that can select fake news in real time.

The present invention generates a candidate sentence that modifies the grammar of a proposition sentence based on deep learning, and compares the candidate sentence with the news sentence to search for fake news using a grammar modification on the neural network to determine whether the news is authentic or not. Provided is a recording medium and apparatus for performing this.

The method of searching for fake news using grammatical transformation on the neural network according to the present invention includes embedding a proposition sentence in which a true or false is determined by a news sentence into a word vector, and the neural network processing the word vector into a predetermined natural language Generating a context vector by inputting it into a context vector, generating the candidate sentence which is a sentence having the same meaning but different grammar by inputting the context vector into a predetermined natural language processing neural network, and the candidate sentence and the And comparing the news sentences to determine whether the propositional sentence is true or false and searching whether the news sentence corresponds to fake news.

Meanwhile, the step of embedding a proposition sentence in which a true or false is determined by the news sentence into a word vector comprises constructing the proposition sentence using a one-hot encoding method. And converting each word into a vector.

In addition, inputting the word vector into a predetermined natural language processing neural network to generate a context vector may include inputting the word vector into an LSTM (Long Short Term Memory) neural network to generate the context vector. Can.

In addition, the step of generating a candidate sentence that is a sentence having the same meaning but a different grammar by inputting the context vector into a predetermined natural language processing neural network is a mechanism for attracting a sequence-to-sequence learning model. Generating an attention vector by calculating a weighted sum of the context vector according to the comparison, and comparing the attention vector and a hidden state vector generated when the context vector is generated to compare a matching vector The predicting step and inputting the matching vector into a LSTM (Long Short Term Memory) neural network may include generating the candidate sentence.

Further, comparing the candidate sentence and the news sentence to determine whether the propositional sentence is true or false, and searching for whether the news sentence corresponds to fake news, sequences of vectors corresponding to each word constituting the candidate sentence Selecting and outputting a given number of words by inputting to a softmax function of a sequence to sequence learning model, and combining output words of the softmax function using a beam search decoder And generating a final candidate sentence and comparing the final candidate sentence with the news sentence to determine true or false of the propositional sentence and searching for whether the news sentence corresponds to fake news.

In addition, comparing the final candidate sentence with the news sentence to determine whether the propositional sentence is true or false and searching for whether the news sentence corresponds to fake news includes embedding the final candidate sentence and the news sentence, Computing the cosine similarity between the embedded final candidate sentence and the news sentence, and determining whether the propositional sentence is true or false by using the cosine similarity.

In addition, a computer-readable recording medium having a computer program recorded thereon may be used to perform a method for searching for fake news using a grammar modification on the neural network.

On the other hand, the fake news search apparatus using the grammar transformation on the neural network according to the present invention is a word embedding unit that embeds a proposition sentence in which a true or false is determined by a news sentence as a word vector, and the word vector is predetermined Matching to generate a context vector by inputting into the natural language processing neural network of the unit, and matching to generate a candidate sentence that has the same meaning as the propositional sentence but having a different grammar by inputting the context vector into a predetermined natural language processing neural network And a reasoning unit for comparing the candidate sentence and the news sentence to determine whether the propositional sentence is true or false, and searching whether the news sentence corresponds to fake news.

Meanwhile, the word embedding unit may convert each word constituting the proposition sentence into a vector using a one-hot encoding method.

Also, the context generation unit may generate the context vector by inputting the word vector into a LSTM (Long Short Term Memory) neural network.

In addition, the matching unit calculates a weighted sum of the context vector according to the attention mechanism of the sequence-to-sequence learning model to generate an attention vector, and the attention vector and the A matching vector is predicted by comparing a hidden state vector generated when the context vector is generated, and the candidate sentence can be generated by inputting the matching vector into an LSTM (Long Short Term Memory) neural network.

In addition, the inference unit selects and outputs a given number of words by inputting a vector corresponding to each word constituting the candidate sentence into a softmax function of a sequence-to-sequence learning model. , Generates a final candidate sentence combining the output words of the softmax function using a beam search decoder, compares the final candidate sentence with the news sentence to determine true or false of the propositional sentence, and determines the news You can search whether the sentence corresponds to fake news.

In addition, the inference unit embeds the final candidate sentence and the news sentence, calculates cosine similarity between the embedded final candidate sentence and the news sentence, and uses the cosine similarity to determine the propositional sentence by the news sentence. You can discriminate between true and false.

According to the present invention, no human judgment is required, and real-time fake news screening is possible. In addition, by selecting fake news, it is possible to prevent the spread of fake news and prevent social confusion that may arise from fake news.

1 is a block diagram of a fake news search apparatus using grammar transformation on a neural network according to an embodiment of the present invention.
FIG. 2 is a diagram illustrating an LSTM-based sequence-to-sequence learning model constructed by a fake news search apparatus using grammar transformation on a neural network according to an embodiment of the present invention shown in FIG. 1.
3 is a flowchart of a fake news search method using grammatical transformation on a neural network according to an embodiment of the present invention.

Advantages and features of the present invention, and methods for achieving them will be clarified with reference to embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in various different forms, and only the present embodiments allow the disclosure of the present invention to be complete, and common knowledge in the art to which the present invention pertains. It is provided to fully inform the holder of the scope of the invention, and the invention is only defined by the scope of the claims. The same reference numerals refer to the same components throughout the specification.

The terminology used herein is for describing the embodiments and is not intended to limit the present invention. In the present specification, the singular form also includes the plural form unless otherwise specified in the phrase. As used herein, "comprises" and/or "comprising" does not exclude the presence or addition of one or more other components, steps and actions.

1 is a block diagram of a fake news search apparatus using grammar transformation on a neural network according to an embodiment of the present invention.

Referring to FIG. 1, the fake news search apparatus 1 using a grammar modification on a neural network according to an embodiment of the present invention includes a word embedding unit 10, a context generating unit 30, a matching unit 50, and a reasoning unit It may include (70).

The fake news search apparatus 1 using grammatical transformation on a neural network according to an embodiment of the present invention can determine whether a news sentence is true or false, and is a candidate who has modified the grammar of a proposition sentence based on deep learning. You can create a sentence and use it to determine the authenticity of the news. Therefore, the fake news search apparatus 1 using the grammar modification on the neural network according to an embodiment of the present invention can present a fast and efficient fake news search model without human intervention.

The fake news search apparatus 1 using grammatical transformation on a neural network according to an embodiment of the present invention may be installed and executed with software (application) for searching for fake news, and a word embedding unit 10 and a context generating unit 30 ), the matching unit 50 and the inference unit 70 may be controlled by software for searching for fake news.

The structure of the word embedding unit 10, the context generating unit 30, the matching unit 50, and the inference unit 70 may be formed of an integrated module, or may be formed of one or more modules. However, on the contrary, each component may be composed of separate modules.

The fake news search apparatus 1 using a grammar modification on a neural network according to an embodiment of the present invention may have mobility or be fixed. The fake news search apparatus 1 using a grammar modification on a neural network according to an embodiment of the present invention may be in the form of a computer, a server, or an engine, and a device and an apparatus , Terminal (terminal), user equipment (UE), mobile station (MS), mobile terminal (MT), user terminal (UT), subscriber station (SS), wireless device (wireless device), personal digital assistant (PDA), It can also be called another term, such as a wireless modem or a handheld device.

Hereinafter, each configuration of the fake news search apparatus 1 using a grammar modification on a neural network according to an embodiment of the present invention shown in FIG. 1 will be described in detail.

The word embedding unit 10 may embed a propositional sentence in which true or false is determined from a news sentence as a word vector.

Here, the propositional sentence is a sentence for determining whether the news to be searched for in the present embodiment is authentic, for example, if the propositional sentence is determined to be false by the news sentence, the corresponding news may be determined as fake news. .

The word embedding unit 10 may generate a word vector from a propositional sentence through embedding. The word embedding unit 10 may convert each word constituting a propositional sentence into a vector by adopting a one-hot encoding method among embedding methods. This is because, in this embodiment, the neural network grammar transformation is used to generate candidate sentences, which have the same meanings as propositional sentences, but different grammars.

On the other hand, in the present embodiment, a sequence-to-sequence learning model based on a Long Short Term Memory (LSTM) neural network may be used when the neural network grammar is modified. The context generation unit 30, the matching unit 50, and the inference unit 70, which will be described later, configure each layer of the LSTM neural network-based sequence-to-sequence learning model to perform fake news search. have.

The LSTM neural network is a structure in which a cell state is added to a hidden state of a Recurrent Neural Network (RNN) neural network that is widely used to process flow between letters and words. The RNN neural network is a kind of artificial neural network in which a hidden node of a hidden state is connected to an edge having a direction to form a circular structure. The cell state added to the LSTM neural network can influence the next cell by using a formula that allows the gradient to propagate relatively well even after a long time.

The sequence-to-sequence learning model is a deep learning algorithm implemented using an LSTM neural network, and is easy to process data whose input dimension and output dimension are not fixed, such as text. The sequence-to-sequence learning model can receive a source sequence and a target sequence as inputs, and these inputs can be divided into two parts: an encoder and a decoder. The encoder converts the source sequence into a vector of fixed size, and the decoder converts the vector of the encoder into a target sequence. At this time, in the process of predicting the target sequence, the decoder may predict the result of the previous prediction as an input of the next step. In order to increase the accuracy of sequence prediction in a sequence-to-sequence learning model, an attention mechanism that impacts important words may be used.

In addition, the sequence-to-sequence learning model finally includes a softmax layer to decode a word to generate a sentence. At this time, a decoder is used to output sentences with high probability words. can do. Various types of such decoders have been proposed, but a sequence-to-sequence learning model applied to the present embodiment may be implemented by including a beam search decoder. The beam search decoder can generate the best combination of sentences by generating a given number of words with the highest probability according to the beam size.

The context generator 30 may generate a context vector by inputting a word vector into a predetermined natural language processing neural network. Here, the natural language processing neural network may be an LSTM neural network as described above.

The matching unit 50 may use the context vector to generate candidate sentences that have the same meaning as the propositional sentence but different grammar.

The matching unit 50 may apply the attention mechanism of the sequence-to-sequence learning model as described above to increase the accuracy of generating candidate sentences.

That is, the matching unit 50 may generate an attention vector from the context vector using the attention mechanism of the sequence-to-sequence learning model. The matching unit 50 may generate an attention vector that gives an impact to an important word by calculating a weighted sum of the context vector according to the attention mechanism.

The matching unit 50 may predict a matching vector by performing a matching operation comparing the hidden state vector generated when the context vector is generated by the attention vector and the context generating unit 30. As described above, the context generator 30 generates a context vector using the LSTM neural network, and a hidden state vector may be generated from the hidden state of the LSTM neural network. The matching unit 50 may predict a matching vector by performing a matching operation comparing the attention vector and the hidden state vector.

The matching unit 50 may generate a candidate sentence by inputting a matching vector into a predetermined natural language processing neural network. Here, the natural language processing neural network may be an LSTM neural network as described above.

The inference unit 70 may compare the candidate sentence and the news sentence to determine whether the propositional sentence is true or false, and search whether the news sentence corresponds to fake news.

The inference unit 70 compares the candidate sentence and the news sentence of the optimal combination to increase the accuracy of the fake news selection, and uses the softmax layer of the sequence-to-sequence learning model as described above. Can be used.

That is, the inference unit 70 may select and output a given number of words by inputting candidate sentences into a softmax function of a sequence-to-sequence learning model. Here, the candidate sentence may correspond to an aggregated matching vector that is the output of the matching unit 50, and the inference unit 70 may select a total matching vector corresponding to each word constituting the candidate sentence. By inputting to the softmax function, a given number of words can be selected and output. In this case, the output word may be a word having a high probability of appearing in a news sentence.

The inference unit 70 may generate the final candidate sentence by inputting the output of the softmax function to the beam search decoder. As described above, the beam search decoder may generate sentences by constructing words, which are outputs of the softmax function, in a complete combination.

The reasoning unit 70 may generate a candidate sentence with more completeness by constructing a softmax layer of the sequence-to-sequence learning model.

The reasoning unit 70 may compare the candidate sentence, which is a sentence in which the grammar of the proposition sentence is modified, to determine whether the proposition sentence is true or false.

To this end, the inference unit 70 may embed the candidate sentence, which is the output of the news sentence and the beam search decoder. For example, the inference unit 70 may embed a news sentence and a candidate sentence using the Doc2Vec model and convert it into a vector. At this time, there may be a plurality of candidate sentences that are outputs of the beam search decoder, and in this case, the inference unit 70 may form a paragraph group by collecting a plurality of candidate sentences.

The inference unit 70 may calculate the cosine similarity between the embedded news sentences and candidate sentences. The reasoning unit 70 may determine true or false of the propositional sentence by the news sentence using the cosine similarity. For example, the cosine similarity may have a value between 0 and 1, and when the cosine similarity is 0.5 or higher, the propositional sentence may be truly determined.

When the inference unit 70 constitutes a paragraph group, both the candidate sentences of the paragraph group and the cosine similarity of the news sentence may be calculated. The inference unit 70 may determine true or false of a propositional sentence by a news sentence by using each candidate sentence of the paragraph group and the highest cosine similarity among news sentences.

The inference unit 70 may select the corresponding news as true when the proposition sentence by the news sentence is true, and if the proposition sentence by the news sentence is false, the news is also fake Can be screened.

As described above, the fake news search apparatus 1 using the grammar transformation on the neural network according to an embodiment of the present invention constitutes an LSTM-based sequence-to-sequence learning model, so that the propositional sentence and the meaning are the same. However, it is possible to select candidate sentences, which are sentences with different grammar, and compare the candidate sentences with news sentences to select fake news.

The fake news search apparatus 1 using the grammar modification on the neural network according to an embodiment of the present invention does not require human judgment and can process fake news in real time.

FIG. 2 is a diagram illustrating an LSTM-based sequence-to-sequence learning model constructed by a fake news search apparatus using grammatical transformation on a neural network according to an embodiment of the present invention shown in FIG. 1.

Referring to FIG. 2, the LSTM-based sequence-to-sequence learning model includes a word embedding layer, a context generation layer, a matching layer, and a reasoning layer ( Inference Layer).

The word embedding unit 10 constitutes a word embedding layer, and may embed a propositional sentence in which true or false is determined from a news sentence as a word vector.

The context generation unit 30 constitutes a context generation layer and may generate a context vector by inputting a word vector into an LSTM neural network.

The matching unit 50 constitutes a matching layer, and may generate a matching vector by performing a matching operation on the context vector. At this time, the matching unit 50 inputs a context vector to the attention function to generate an attention vector, and performs a matching operation comparing the attention vector and the hidden state vector of the LSTM neural network included in the context generation layer Matching vectors can be generated. Then, the matching unit 50 may input a matching vector into the LSTM neural network to generate an associated sum matching vector corresponding to a candidate sentence having the same meaning as a propositional sentence but having a different grammar.

The inference unit 70 configures an inference layer, obtains a sum matching vector from the associated sum matching vector, and inputs it to a softmax function to select and output a given number of words. Then, the inference unit 70 may generate a plurality of candidate sentences by inputting the output of the softmax function to the beam search decoder. The reasoning unit 70 may compare a plurality of candidate sentences and news sentences to determine whether a propositional sentence is true or false based on the news sentence, and select fake news according to the result.

Hereinafter, a method of searching for fake news using a grammar modification on a neural network according to an embodiment of the present invention will be described.

3 is a flowchart of a fake news search method using grammatical transformation on a neural network according to an embodiment of the present invention.

The fake news search method using the grammar modification on the neural network according to an embodiment of the present invention may be executed in substantially the same configuration as the fake news search apparatus 1 using the grammar modification on the neural network shown in FIG. 1. Therefore, the same components as the fake news search apparatus 1 using the grammar transformation on the neural network shown in FIG. 1 are given the same reference numerals, and repeated descriptions are omitted.

Referring to FIG. 3, the word embedding unit 10 may embed a propositional sentence into a word vector (S100 ).

A propositional sentence corresponds to a sentence for which true or false will be determined by a news sentence. The word embedding unit 10 may convert each word constituting a propositional sentence into a vector by adopting a one-hot encoding method among embedding methods.

The context generator 30 may generate a context vector by inputting a word vector into the neural network (S200 ).

Here, the neural network is a natural language processing neural network, and may be an LSTM neural network.

The matching unit 50 may generate a candidate sentence by inputting a context vector into the neural network (S300).

Candidate sentences are sentences that have the same meaning as propositional sentences but different grammar.

The matching unit 50 may generate an attention vector from the context vector using the attention mechanism of the sequence-to-sequence learning model. The matching unit 50 may generate an attention vector that gives an impact to an important word by calculating a weighted sum of the context vector according to the attention mechanism.

The matching unit 50 may predict a matching vector by performing a matching operation that compares the hidden state vector generated when the attention vector and the context vector are generated.

The matching unit 50 may generate a candidate sentence by inputting a matching vector into the LSTM neural network.

The reasoning unit 70 may search for fake news by comparing the candidate sentence and the news sentence (S400).

The inference unit 70 may compare the candidate sentence and the news sentence to determine whether the propositional sentence is true or false, and search whether the news sentence corresponds to fake news.

The inference unit 70 may select and output a given number of words by inputting candidate sentences into a softmax function of a sequence-to-sequence learning model. Here, the candidate sentence may correspond to an aggregated matching vector that is the output of the matching unit 50, and the inference unit 70 may select a total matching vector corresponding to each word constituting the candidate sentence. By inputting to the softmax function, a given number of words can be selected and output. In this case, the output word may be a word having a high probability of appearing in a news sentence.

The inference unit 70 may generate the final candidate sentence by inputting the output of the softmax function to the beam search decoder. As described above, the beam search decoder may generate sentences by constructing words that are outputs of a softmax function in a complete combination.

The inference unit 70 may embed the candidate sentence, which is the output of the news sentence and the beam search decoder. For example, the inference unit 70 may embed news sentences and candidate sentences using the Doc2Vec model. At this time, there may be a plurality of candidate sentences that are outputs of the beam search decoder, and in this case, the reasoning unit 70 may form a paragraph group by collecting a plurality of candidate sentences.

The inference unit 70 may calculate the cosine similarity between the embedded news sentences and candidate sentences. The reasoning unit 70 may determine true or false of the propositional sentence by the news sentence using the cosine similarity. For example, the cosine similarity may have a value between 0 and 1, and when the cosine similarity is 0.5 or higher, the propositional sentence may be truly determined.

When the inference unit 70 constitutes a paragraph group, both the candidate sentences of the paragraph group and the cosine similarity of the news sentence may be calculated. The inference unit 70 may determine true or false of a propositional sentence by a news sentence by using each candidate sentence of the paragraph group and the highest cosine similarity among news sentences.

The fake news search method using the grammar modification on the neural network of the present invention is implemented in the form of program instructions that can be executed through various computer components and can be recorded in a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, or the like alone or in combination.

The program instructions recorded on the computer-readable recording medium are specially designed and configured for the present invention, and may be known and available to those skilled in the computer software field.

Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs, DVDs, and magneto-optical media such as floptical disks. media), and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like.

Examples of program instructions include not only machine language codes produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform processing according to the present invention, and vice versa.

Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art to which the present invention pertains may be implemented in other specific forms without changing the technical spirit or essential features of the present invention. You will understand. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive.

1: Fake news search device using grammar transformation on neural network
10: word embedding
30: context creation unit
50: matching unit
70: reasoning

Claims (13)

In the fake news search method using the grammar transformation on the neural network performed by the fake news search device,
Embedding a proposition sentence in which the true or false is determined by the news sentence as a word vector;
Generating a context vector by inputting the word vector into a predetermined natural language processing neural network;
Inputting the context vector into a predetermined natural language processing neural network to generate a candidate sentence having the same meaning as the propositional sentence but a sentence with a different grammar; And
Comparing the candidate sentence and the news sentence to determine the true or false of the propositional sentence and searching for whether the news sentence corresponds to the fake news,
Comparing the candidate sentence and the news sentence to determine the true or false of the propositional sentence and searching for whether the news sentence corresponds to fake news,
Selecting and outputting a given number of words by inputting a vector corresponding to each word constituting the candidate sentence into a softmax function of a sequence-to-sequence learning model;
Generating a final candidate sentence by combining output words of the softmax function using a beam search decoder; And
And comparing the final candidate sentence with the news sentence to determine whether the propositional sentence is true or false and searching whether the news sentence corresponds to fake news. According to claim 1,
The step of embedding a proposition sentence in which the true or false is determined by the news sentence as a word vector is:
A method of searching for fake news using grammar transformation on a neural network, comprising converting each word constituting the propositional sentence into a vector using a one-hot encoding method. According to claim 1,
Generating a context vector by inputting the word vector into a predetermined natural language processing neural network includes:
And generating the context vector by inputting the word vector into a Long Short Term Memory (LSTM) neural network. According to claim 1,
The step of generating a candidate sentence, which has the same meaning as the propositional sentence but has a different grammar, by inputting the context vector into a predetermined natural language processing neural network,
Generating an attention vector by calculating a weighted sum of the context vector according to the attention mechanism of the sequence-to-sequence learning model;
Predicting a matching vector by comparing the attention vector and a hidden state vector generated when the context vector is generated; And
And generating the candidate sentence by inputting the matching vector into a long short term memory (LSTM) neural network. delete According to claim 1,
Comparing the final candidate sentence and the news sentence to determine the true or false of the propositional sentence and searching for whether the news sentence corresponds to fake news,
Embedding the final candidate sentence and the news sentence;
Calculating cosine similarity between the embedded final candidate sentence and the news sentence; And
And determining whether the propositional sentence is true or false based on the news sentence using the cosine similarity. A computer-readable recording medium having a computer program recorded thereon for performing a fake news search method using a grammar modification on a neural network according to any one of claims 1 to 4. A word embedding unit for embedding a proposition sentence in which a true or false is determined by a news sentence as a word vector;
A context generating unit that inputs the word vector into a predetermined natural language processing neural network to generate a context vector;
A matching unit that inputs the context vector into a predetermined natural language processing neural network to generate a candidate sentence that has the same meaning as the propositional sentence but has a different grammar; And
And comparing the candidate sentence with the news sentence to determine whether the propositional sentence is true or false, and includes a reasoning unit to search whether the news sentence corresponds to fake news,
The reasoning unit,
A vector corresponding to each word constituting the candidate sentence is input to a softmax function of a sequence-to-sequence learning model to select and output a given number of words, and a beam search decoder is used. To generate a final candidate sentence combining the output words of the softmax function, comparing the final candidate sentence with the news sentence to determine true or false of the propositional sentence, and the news sentence corresponds to fake news A device for searching fake news using grammatical transformation on a neural network, searching for knowledge. The method of claim 8,
The word embedding unit,
A fake news search apparatus using grammatical transformation on a neural network that converts each word constituting the propositional sentence into a vector using a one-hot encoding method. The method of claim 8,
The context generating unit,
A fake news search apparatus using grammar transformation on a neural network that generates the context vector by inputting the word vector into a LSTM (Long Short Term Memory) neural network. The method of claim 8,
The matching unit,
Generates an attention vector by calculating the weighted sum of the context vector according to the attention mechanism of the sequence-to-sequence learning model, and occurs when the attention vector and the context vector are generated A fake news search apparatus using a grammar transformation on a neural network that generates a candidate sentence by comparing a hidden state vector to predict a matching vector and inputting the matching vector into a LSTM (Long Short Term Memory) neural network. delete The method of claim 8,
The reasoning unit,
Embedding the final candidate sentence and the news sentence, calculating the cosine similarity between the embedded final candidate sentence and the news sentence, and using the cosine similarity to determine true or false of the propositional sentence by the news sentence Fake news search device using grammar transformation on neural network.

Images