KR20200064943A - Fake news detection server and method based on korean grammar transformation - Google Patents

Abstract

According to an embodiment of the present invention, provided is a fake news detection method based on Korean grammar conversion performed by a fake news detection server, which includes the steps of: receiving a query sentence and a news article to be detected; extracting an article sentence related to a query sentence as a key sentence in the news article to be detected; checking whether a word in the query sentence corresponds to a word in the key sentence; checking whether the query sentence and the key sentence are semantically matched according to whether the word matches; determining whether a grammatical correspondence between a query sentence and a core sentence is made according to whether or not there is a semantic correspondence; and determining whether a key sentence is true or false based on the confirmed semantic response result and grammatical response result.

Description

Fake NEWS DETECTION SERVER AND METHOD BASED ON KOREAN GRAMMAR TRANSFORMATION}

The fake news detection system based on Korean grammar conversion according to the present invention relates to a server and a method for selecting fake news by matching query sentences related to news articles and news article sentences based on Korean grammar.

In general, the large amount of content shared by people creates different opinions. Sometimes fake news created with false commercial and political intentions can adversely affect public opinion. With the development of various media and communication technologies, fake news detection has become an essential and challenging problem in determining the truth of news.

On the other hand, sentence matching is a core technique of natural language processing, and it is possible to check whether two sentences to be compared are semantically similar. Deep learning research has been recently activated with the development of hardware such as GPUs. The natural language processing model based on deep learning has been developed in various attempts for sentence matching. Some of them used a recurrent neural network (RNN) to understand the meaning of contexts of varying lengths. Since RNN can process a large amount of data sequentially, it is suitable for semantic analysis of several sentences, but has a problem with vanishing (exploding gradient) of information. This problem has been improved with a long short-term memory (LSTM) method that adds a Forget Gate to RNN.

Prior literature related to the present invention includes Bilateral Multi-Perspective Matching for Natural Language Sentences (BiMPM, Zhiguo Wang, 2017). The BiMPM model achieved the latest performance in the test results using the English dataset. Despite these achievements, there are some limitations in applying the Korean news dataset to the BiMPM model.

First, because the morphological characteristics of English and Korean are different, the application of Korean sentence matching is limited based on English sentence matching technology. The second is that it is difficult to capture important information in articles with many sentences, such as news articles.

The problem to be solved of the present invention is to find a sentence related to a query sentence in a news article, and provide a result of determining semantic matching using a deep learning model based on the result of analysis at the word level, and analysis considering Korean grammar By providing a sentence match result through, semantically sentence matching results and grammatically sentence matching results are aggregated to detect fake news.

A fake news detection method based on Korean grammar conversion performed by a fake news detection server according to an embodiment of the present invention, comprising: receiving a query sentence and a news article to be detected; Extracting an article sentence related to a query sentence as a key sentence from the news article to be detected; Checking whether a word in the query sentence corresponds to a word in the key sentence; Checking whether the query sentence and the key sentence are semanticly matched according to whether the word matches; Determining whether a grammatical correspondence between a query sentence and a core sentence is made according to whether or not there is a semantic correspondence; And determining whether the core sentence is true or false based on the confirmed semantic correspondence result and the grammatical correspondence result.

In the present embodiment, the core sentence extraction step includes: decomposing a query sentence and an article sentence into word configuration units including a root and a suffix; Calculating a cosine similarity by vectorizing the word construction units of the query sentence and the word construction units of the article sentence and comparing them with each other; And extracting an article sentence including a word construction unit of a query sentence and a word construction unit having the highest cosine similarity as a core sentence, and providing a fake news detection method based on Korean grammar conversion.

In the present embodiment, the step of determining whether a word in a query sentence corresponds to a word in a key sentence includes: decomposing the query sentence into words and storing the words in a batch order to generate a word matching set; Generating a word matching set by decomposing a key sentence into words and storing each word in the arrangement order of words; Comparing the word matching set of the query sentence and the word matching set of the core sentence with each other; And classifying and outputting the comparison result into full word matching and partial word matching, and providing a fake news detection method based on Korean grammar conversion.

In the present exemplary embodiment, when the comparison result is output as a complete word match, it is determined that the query sentence and the core sentence are semantically matched, and the grammatical correspondence between the query sentence and the core sentence is checked. It can provide news detection methods.

In the present embodiment, the step of determining whether the query sentence and the key sentence are semantic includes: extracting a word vector of the query sentence and a word vector of the key sentence; Matching the word vector of the extracted query sentence with the word vector of the core sentence with the word embedding vector of pre-trained Korean news articles; Applying deep learning to the word vector of the matched query sentence and the word vector of the core sentence to extract each as a context-related embedding vector; Extracting the query sentence matching vector and the core sentence matching vector by applying deep learning to the context-related embedding vector and the core sentence context-related embedding vector of the extracted query sentence, respectively; Step of applying deep learning to the extracted query sentence matching vector and core sentence matching vector to extract into the context aggregated query matching vector and contextual aggregated key sentencing set matching vector ; Aggregated Question Matching Vector and Aggregated Key Sentence Set by matching the Last Time Step Matching Vector to the context aggregate query sentence matching vector and the context aggregate core sentence matching vector Matching Vector); And calculating the semantic similarity between the query sentence and the core sentence using artificial neural networks and normalization based on the extracted aggregate query sentence matching vector and the aggregate core sentence matching vector. Can provide.

In the present embodiment, the step of determining whether the query sentence and the key sentence are grammatical includes: setting the order of words in the query sentence and the order of words in the key sentence, respectively; Morphologically analyzing the words of the query sentence in which the order is set and the words of the core sentence and classifying them into each morpheme; Classifying the morphemes of the classified query sentence words and the morphemes of the key sentence words based on the word; Generating a pattern of a query sentence classified based on a word phrase and a pattern of a core sentence; And comparing the pattern of the generated query sentence with the pattern of the key sentence to check whether the query sentence and the key sentence are mutually deformable sentences.

In the fake news detection server according to an embodiment of the present invention, a memory in which a fake news detection method program based on Korean grammar conversion is recorded; And a processor that executes a program recorded in the memory, and the processor receives a query sentence and a news article to be detected according to execution of the program, and extracts an article sentence related to the query sentence from the news article to be detected as a core sentence. , And checks whether the word of the query sentence and the key sentence correspond, and whether the word matches or not, checks the semantic correspondence of the query sentence and the key sentence, and according to the semantic response, of the query sentence and the key sentence. It is intended to provide a fake news detection server based on Korean grammar conversion, which checks for grammatical correspondence and determines whether a key sentence is true or false based on the confirmed semantic correspondence results and grammatical correspondence results.

The effect of the present invention is to find a sentence related to a query sentence in a news article and provide a result of determining semantic matching using a deep learning model based on the result of analysis at the word level, and sentence through analysis considering Korean grammar By providing a matched result, it is possible to detect fake news by counting the semantically sentence-matched result and the grammatically sentence-matched result.

In addition, according to the present invention, it is possible to determine the authenticity of a news article by targeting news data generated from mass media such as Internet newspapers and social network services. In addition, information erroneously delivered by the mass media can be determined in consideration of the authenticity of the news article. Furthermore, public opinion formed by fake news can be prevented.

1 is a schematic diagram illustrating a fake news detection method according to the present invention.
2 is a block diagram illustrating a fake news detection method according to the present invention.
3 is a flowchart illustrating a fake news detection method according to the present invention.
4 is a flow chart showing a core sentence extraction method according to the present invention.
5 is a schematic diagram showing a core sentence extraction model according to the present invention.
6 is a flowchart illustrating a method for confirming whether a word corresponds according to the present invention.
7 is a schematic diagram showing a word matching module according to the present invention.
8 is a flowchart illustrating a method for confirming whether or not a semantic response is made according to the present invention.
9 is a schematic diagram showing a semantic correspondence checking module according to the present invention.
10 is a schematic diagram showing a part of a module for determining whether a grammatical correspondence exists according to the present invention.
11 is a diagram showing sentence structure for checking whether grammatical correspondence is in accordance with the present invention.
12 is a diagram showing an example of an algorithm of a method for checking whether a semantic correspondence is performed according to the present invention.
13 is a diagram showing an example of an algorithm of a method for determining whether a grammatical correspondence is performed according to the present invention.
14 is a diagram showing input sentence and morpheme analysis results of a method for determining whether a grammatical correspondence is in accordance with the present invention.
15 is a diagram illustrating sentence-by-word sentence analysis of a method for determining whether grammatical correspondence is in accordance with the present invention.
16 is a diagram showing the result of generating a sentence pattern according to the present invention.

Hereinafter, with reference to the accompanying drawings, embodiments of the present invention will be described in detail so that those skilled in the art to which the present invention pertains can easily practice. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. In addition, in order to clearly describe the present invention in the drawings, parts irrelevant to the description are omitted, and like reference numerals are assigned to similar parts throughout the specification.

Throughout the specification, when a part is "connected" to another part, this includes not only "directly connected" but also "electrically connected" with another element in between. . Also, when a part is said to "include" a certain component, this means that other components may be further included rather than excluding other components, unless otherwise stated.

Hereinafter, a fake news detection server based on Korean grammar conversion according to an embodiment of the present invention will be described.

1 is a schematic diagram illustrating a fake news detection method according to the present invention.

Referring to FIG. 1, a fake news detection server based on Korean grammar conversion according to an embodiment of the present invention receives a query sentence and a news article from a user terminal, and calculates the similarity between the query sentence and the truth sentence in the news article. Based on this, the authenticity of the query sentence can be determined and the result value can be provided to the user terminal.

Meanwhile, the fake news detection server may include a communication module, a memory, and a processor.

The communication module provides a communication interface to a fake news detection server by interworking with a communication network, and may serve to transmit and receive data to and from a user terminal. Here, the communication module may be a device including hardware and software necessary for transmitting and receiving a signal such as a control signal or a data signal through a wired or wireless connection with another network device.

The memory may be a record of a fake news detection program. In addition, the memory may function to temporarily or permanently store data processed by the processor. Here, the memory may include volatile storage media or non-volatile storage media, but the scope of the present invention is not limited thereto.

The processor may control the entire process performed by the fake news detection program in the fake news detection server. Each step of the process performed by the processor will be described later with reference to FIGS. 2 to 10.

Here, the processor may include any kind of device capable of processing data, such as a processor. Here, a'processor' may mean a data processing device embedded in hardware having physically structured circuits, for example, to perform functions represented by codes or instructions included in a program. As an example of such a data processing device embedded in hardware, a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, and an application-specific integrated ASIC circuit, a field programmable gate array (FPGA), or the like, but the scope of the present invention is not limited thereto.

Hereinafter, a method of detecting fake news based on Korean grammar conversion according to an embodiment of the present invention will be described.

2 is a block diagram illustrating a fake news detection method according to the present invention. 3 is a flowchart illustrating a fake news detection method according to the present invention.

2 and 3, the fake news detection program according to an embodiment of the present invention is a core sentence extraction model for finding an article sentence related to a query sentence in a news article, whether a word in the query sentence corresponds to a word in the core sentence We provide a fake news detection program based on Korean grammar conversion that integrates a model to check, a model to check whether a query sentence and a key sentence have a semantic response, and a model to check whether a query sentence and a key sentence have a grammatical response.

The program roughly divides query sentences related to news articles into word units, and extracts key sentences with high relevance from news articles using word units. Then, it checks whether the word of the news article sentence related to the query sentence is matched, checks the semantic correspondence by matching the query sentence and the core sentence, checks the grammatical correspondence, and then uses Bidirectional Long Short-term Memory (BiLSTM). By semantically matching the result of sentence matching and the result of grammatically matching sentence, fake news can be read. Here, the query sentence is a sentence that requires confirmation of authenticity, and the news article to be detected may be a news article composed of truth information without false information.

In the method of detecting fake news based on Korean grammar conversion performed by the fake news detection server according to an embodiment of the present invention, first, a step S310 of receiving a query sentence and a news article to be detected may be performed.

Next, a step (S320) of extracting an article sentence related to a query sentence as a key sentence in the news article to be detected may be performed.

4 is a flowchart showing a method of extracting a key sentence according to the present invention. 5 is a schematic diagram showing a core sentence extraction model according to the present invention.

Referring to FIGS. 4 and 5, in the core sentence extraction step, first, a step (S410) of decomposing the query sentence and the article sentence into word configuration units including a root and a suffix may be performed.

Next, a step (S420) of calculating a cosine similarity by vectorizing and comparing each of the word unit of the query sentence and the word unit of the article sentence may be performed. This is for retrieving the sentence most similar to the query sentence in terms of the frequency of the word using the decomposed word unit in the news article.

Next, a step (S430) of extracting the article sentence including the word unit of the query sentence and the word unit of the highest cosine similarity as a key sentence may be performed. This is to extract the core sentence by checking the frequency of appearance of the query sentence word composition unit in each sentence of the news article.

6 is a flowchart illustrating a method for confirming whether a word corresponds according to the present invention. 7 is a schematic diagram showing a word correspondence checking module according to the present invention.

Next, referring to FIGS. 6 and 7, a step of determining whether a word in a query sentence corresponds to a word in a key sentence may be performed.

In the step of determining whether a word in the query sentence corresponds to a word in the key sentence, first, a step (S610) of generating a word matching set by decomposing the query sentence into words and storing the words in the arrangement order of each word may be performed.

Next, a step of generating a word matching set (WMS: Word Matching Sentence Set) may be performed by decomposing the key sentences for each word and storing them in the order of arrangement of each word (S620 ). The word matching set is for splitting the input sentences and storing them in the list in the order of division.

Next, a step of comparing the word matching set of the query sentence with the word matching set of the core sentence (S630) may be performed.

Next, a step of classifying and outputting the comparison result into complete word matching and partial word matching may be performed. Here, the complete word matching is that the words of the query sentence and the key sentence are all matched, and the partial word matching is partially matched among words and/or sequences, or the word non-word matching is the word and This means that the order was not matched at all.

On the other hand, when the comparison result is output as a complete word match, it is determined that the query sentence and the core sentence are semantically matched, and it is possible to check whether the query sentence and the core sentence are grammatically matched. The method of checking whether or not the grammatical correspondence will be described later.

8 is a flowchart illustrating a method for confirming whether a semantic response is made according to the present invention. 9 is a schematic diagram showing a semantic correspondence checking module according to the present invention.

Next, referring to FIGS. 8 and 9, a step of determining whether a query sentence and a key sentence are semanticly matched may be performed according to word matching.

In the step of determining whether the query sentence and the key sentence are semanticly matched, a step S810 of extracting the word vector of the query sentence and the word vector of the key sentence may be performed.

Specifically, the words of the query sentence and the words of the key sentence may be output as a query word vector and a key sentence set word vector through a word representation layer.

Next, a step (S820) of matching the word vector of the extracted query sentence with the word vector of the core sentence with the word embedding vector of a pre-trained Korean news article may be performed.

Specifically, the query sentence word vector and the core sentence word vector may match Korean newspaper articles with words of word embedding learned using the word2vec algorithm.

Next, a step S830 of extracting each context-related embedding vector by applying deep learning to the word vector of the matched query sentence and the word vector of the core sentence may be performed.

Specifically, the output word vector is to be output as a context related query sentence embedding vector and a key sentence set contextual embedding vector through BiLSTM in a context representation layer. Can be.

Here, word embedding refers to a process of converting a character of a word into a vector. The process of converting a word vector into a context-related embedding vector first inputs a query sentence word vector and a key sentence word vector into BiLSTM. At this time, BiLSTM encodes each input word vector information and integrates the information according to word order and surrounding word distribution. And each encoded information is output as the context-related embedding vector of the query sentence and the context-related embedding vector of the core sentence. Accordingly, the information can be expressed as an embedding vector in a wider range than the information expressed in the word vector.

Next, step S840 of extracting the query sentence matching vector and the core sentence matching vector by applying deep learning to the context-related embedding vector and the core sentence context-related embedding vector of the extracted query sentence may be performed, respectively.

Specifically, the embedding vector related to the context of the output query sentence and the core sentence is a matching vector by applying deep matching and full matching in the matching representation layer. , key sentence set matching vector).

In order to generate a matching vector, first, a context-related embedding vector of a query sentence and a context-related embedding vector of a core sentence are intersected and input into a layer performing a matching operation. At this time, the layer that performs the matching operation performs a full matching (Full matching) and detailed matching (Attentive matching) operation process. Here, full matching and detailed matching are computational tools for generating a matching vector. In each operation process, first, full matching sequentially matches all vector information expressed in the query sentence context-related embedding vector and the core sentence context-related embedding vector to generate vectors.

In the process of generating a matching vector, for example, sequentially matched vectors, the current vector information and the next order of vector information are respectively calculated through a cosine similarity operation. Here, the vector that is continuously calculated is an embedding vector related to the context of the core sentence. The current vector information and the next order of vector information are compared, and the weights are updated until the last vector appears to output the vectors.

Overall reorganization allows context-related vectors to be inputted crossing each other, and does not perform the same operation as each other, but one side sequentially calculates context vectors of key sentences based on the query statement, and the other side executes the core statements. Based on, the context vectors of the query statement are sequentially calculated.

Next, the detailed matching is similar to the perfect matching operation technique, but instead of continuously updating the weights generated in the weight calculation process, the results are multiplied by the weighted sum operation, and the results are multiplied. It is to sum them up again.

Lastly, the results calculated through full matching and detailed matching are combined to output matching vectors. Accordingly, by comparing the information of the query sentence and the key sentence, similar information can be highlighted more similarly, and unsimilar parts can be highlighted as more dissimilar information.

Next, deep learning is applied to the extracted query sentence matching vector and the core sentence matching vector, resulting in a context aggregated query matching vector and a contextual aggregated key matching set. Extracting step (S850) may be performed.

Specifically, each output matching vector may be output as a context aggregation query sentence matching vector and a context aggregation core sentence matching vector through another BiLSTM.

Here, vector information may be implied through a process of encoding the output matching vectors using another BiLSTM. Accordingly, it is possible to reduce computational complexity in the process of calculating semantic similarity by implicating complex matching vectors with aggregated vectors.

Next, an aggregate query sentence matching vector and an aggregate core sentence matching vector are matched by matching a last time step matching vector to the context aggregate query sentence matching vector and the context aggregate core sentence matching vector ( Step S860 of extracting the Aggregated Key Sentence Set Matching Vector may be performed.

Next, using the artificial neural network and normalization based on the extracted aggregate query sentence matching vector and the aggregate core sentence matching vector, step S870 of calculating the semantic similarity between the query sentence and the core sentence may be performed. .

Specifically, the output aggregated query sentence matching vector and aggregated core sentence matching vector are used to determine the query sentence and the query statement using a feed forward neural network composed of two layers in a decision layer and a softmax activation function. Semantic similarity between key sentences can be calculated.

10 is a schematic diagram showing a part of a module for determining whether a grammatical correspondence exists according to the present invention. 11 is a diagram showing sentence structure for checking whether grammatical correspondence is in accordance with the present invention.

Next, referring to FIGS. 10 and 11, a step of determining whether a grammatical correspondence between a query sentence and a core sentence may be performed according to semantic correspondence. The grammatical correspondence checking step is a step of checking grammatical variations by matching the grammatical characteristics of the query sentence and the extracted news article core sentence.

The step of determining whether a grammatical correspondence may be performed may include first setting the order of words in the query sentence and the order of words in the core sentence (S1010 ).

Next, a step (S1020) of sorting the words of the query sentence in which the order is set and the words of the core sentence by morpheme analysis may be performed. In order to specify the grammatical characteristics of the query sentence and the core sentence, the sentence may be structured in the form of a parsing tree as shown in FIG. 11. Here, terminal is the minimum unit of words expressed in a sentence. Non-terminal is a terminal and other non-terminal degradable units.

Next, a step (S1030) of classifying the morphemes of the classified query sentence words and the morphemes of the key sentence words based on the word may be performed.

Next, a step (S1040) of generating a pattern of a query sentence classified based on a word and a pattern of the core sentence may be performed. Each sentence classified based on a word is classified into the components of a subject, an object, a complement, a verb, an adverb, and an independent component to generate a pattern. Can be.

Next, a step (S1040) of determining whether the query sentence and the core sentence are mutually deformable may be performed by comparing the pattern of the generated query sentence with the pattern of the core sentence. If the pattern of the query sentence and the pattern of the core sentence are identified as mutually deformable sentences, the query sentence can output a true value in authenticity judgment.

Finally, a step (S1050) of determining whether a key sentence is true or false based on the confirmed semantic correspondence result and the grammatical correspondence result and outputting the result value may be performed. For example, as a result of semantic correspondence, if the semantic similarity is calculated as a predetermined value or more, and if the query sentence and the core sentence are identified as mutually deformable sentences as a result of the grammatical correspondence, the query sentence may output a true value in authenticity determination. .

12 is a diagram showing an example of an algorithm of a method for determining whether a semantic correspondence is performed according to the present invention. 13 is a diagram showing an algorithm of a method for determining whether a grammatical correspondence is performed according to the present invention. 14 is a diagram showing input sentence and morpheme analysis results of a method for determining whether a grammatical correspondence is in accordance with the present invention. 15 is a diagram illustrating sentence-by-word analysis of a method for determining whether a grammatical correspondence is performed according to the present invention. 16 is a diagram showing the result of generating a sentence pattern according to the present invention.

12 to 16, algorithms and performance results of a method for determining semantic correspondence performed by a computer of the present invention, and algorithms and performance results for a method for determining grammatical correspondence can be confirmed.

When describing grammatical correspondence, it is assumed that the following sentence disclosed in FIG. 14 is given as input.

Question sentence: Pohang Citizen Bowling Center, which is used by 300 people on average per day, operates in 24 lanes.

Key sentence: Pohang Citizen Bowling Alley operates 24 lanes from 10 am to 12 pm and is used by an average of 300 people per day.

First, patterns of input sentences are created with the following six main and sub components.

These may be Sbj (main component), Obj (target component), Cmp (bore component), Vp (verb component), Adverb (adverb component), and independent components, respectively.

In addition, there may be conditions in which a pattern cannot be generated, so a pattern for exception handling is also generated, and the generated one may be Etc (other components).

And follow the part-of-speech tag criteria as shown in [Table 1] below. (It may be different for each morpheme analyzer to be used, but here we will follow the mecab morpheme analyzer standards.)

NNG Common noun NNP Proper noun NNB Dependent nouns NNBC Noun NR Investigation NP pronoun VV verb VA adjective VX Auxiliary sayings VCP A positive designator VCN Fraud MM Adjective MAG General adverb MAJ Adverb IC interjection JKS Subjective investigation JKC An interrogation investigation JKG Pipe type investigation JKO Objective investigation JKB Secondary fire investigation JKV Census JKQ Citation investigation JX Assistant JC Connection investigation EP A fresh mother EF The ending mother EC Connecting mother ETN A noun form ETM Tubular malleable mother XPN Body prefix XSN Noun derived suffix XSV Verb-derived suffix XSA Adjective derivative suffix XR radix SF Period, question mark, exclamation mark SE Ellipsis? SSO Opening parenthesis (, [ SSC Closing parenthesis ),] SC Separator , · / : SY Other symbols SL Foreign language SH Chinese character SN number

Here, the pattern of the sentence may be formed on the basis of the results (see FIG. 14) output from the morpheme analysis (JK ~ component) or auxiliary components of the assistant's sentence. For example, look at the subject and bore component pattern classification, which has a slightly more complicated rule than other components. If "Average 300 people per day" is analyzed, "this" is a subject investigation (JKS), so it can be a subject. And the words "silver, silver, tooth, and a", which can be universally subject, can be a subject. But there are exceptions. If it is judged that the subject is simply included in the subject investigation (JKS), the bore component cannot be found. Here, by applying the Bohr rule additionally, even if the subject is found, if another subject candidate such as'+' EMT' (tubular pronoun mother) appears in the next word, it can be judged as a bore and separated from the first subject candidate word. have. For example, it can be performed in the order of "average of 300 people per day" -> subject ingredient "used" -> bore ingredient (supplement the subject of the sentence). It is merged with words marked as found (that is, the sentence component is not yet distinguishable), and expressed as a sentence pattern. Therefore, it can be output in the same form as the sentence pattern generation result disclosed in FIG. 16.

The above-described effect of the present invention provides a result of determining a semantic match using a deep learning model based on a result of finding a sentence related to a query sentence in a news article and analyzing it at a word level, and an analysis considering Korean grammar Through this, it is possible to detect fake news by providing sentence matching results and counting semantically matching sentences and grammatically matching sentences.

In addition, according to the present invention, it is possible to determine the authenticity of a news article by targeting news data generated from mass media such as Internet newspapers and social network services. In addition, information erroneously delivered by the mass media can be determined in consideration of the authenticity of the news article. Furthermore, public opinion formed by fake news can be prevented.

Meanwhile, the method for detecting fake news based on Korean grammar conversion according to an embodiment of the present invention may be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. Computer readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. In addition, computer readable media may include computer storage media. 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. Although the methods and systems of the present invention have been described in connection with specific embodiments, some or all of their components or operations may be implemented using a computer system having a general purpose hardware architecture.

The above description is merely illustrative of the technical idea of the present invention, and those skilled in the art to which the present invention pertains may make various modifications and variations without departing from the essential characteristics of the present invention. Therefore, the embodiments disclosed in the present invention are not intended to limit the technical spirit of the present invention, but to explain them, and the scope of the technical spirit of the present invention is not limited by these embodiments. The scope of protection of the present invention should be interpreted by the claims below, and all technical spirits within the scope equivalent thereto should be interpreted as being included in the scope of the present invention.

Claims (13)

In the fake news detection method based on Korean grammar conversion performed by the fake news detection server,
Receiving a query sentence and a news article to be detected;
Extracting an article sentence related to the query sentence as a key sentence in the news article to be detected;
Checking whether a word in the query sentence corresponds to a word in the core sentence;
Checking whether the query sentence and the core sentence are semanticly matched according to whether the word matches;
Checking whether a grammatical correspondence exists between the query sentence and the core sentence according to the semantic correspondence; And
And determining whether the core sentence is true or false based on the identified semantic correspondence result and the grammatical correspondence result. According to claim 1,
The core sentence extraction step,
Decomposing the query sentence and the article sentence into word composition units including a root and a suffix;
Calculating a cosine similarity by vectorizing and comparing the word construction units of the query sentence with the word construction units of the article sentence; And
And extracting the article sentence including the word construction unit of the query sentence and the word construction unit having the highest cosine similarity as the core sentence. According to claim 1,
The step of checking whether a word in the query sentence corresponds to a word in the key sentence is
Decomposing the query sentence for each word and storing each word in the arrangement order of the words to generate a word matching set;
Decomposing the key sentence into words and storing the words in the arrangement order of each word to generate a word matching set;
Comparing the word matching set of the query sentence and the word matching set of the core sentence with each other; And
A method for detecting fake news based on Korean grammar conversion, comprising the step of classifying and outputting the comparison result into full word matching and partial word matching. According to claim 3,
When the comparison result is output as a complete word match,
A fake news detection method based on Korean grammar conversion by determining whether the query sentence and the core sentence are semantically matched and checking whether the query sentence and the core sentence are grammatically matched. According to claim 1,
The step of checking whether the query sentence and the key sentence are semanticly corresponded is:
Extracting the word vector of the query sentence and the word vector of the core sentence;
Matching the word vector of the extracted query sentence and the word vector of the core sentence with a word embedding vector of pre-trained Korean news articles;
Applying deep learning to the word vector of the matched query sentence and the word vector of the core sentence to extract each as a context-related embedding vector;
Extracting a query sentence matching vector and a core sentence matching vector by applying deep learning to the context-related embedding vector of the extracted query sentence and the core sentence context-related embedding vector;
Deep learning is applied to the extracted query sentence matching vector and the core sentence matching vector to extract them into a context aggregated question matching vector and a contextual aggregated key sentencing set matching vector To do;
Aggregated Question Matching Vector and Aggregated Key Sentence Matching Vector by matching the Last Time Step Matching Vector to the context aggregate query sentence matching vector and the context aggregate core sentence matching vector Sentence Set Matching Vector); And
And calculating a semantic similarity between the query sentence and the core sentence using artificial neural networks and normalization based on the extracted aggregate query sentence matching vector and the aggregate core sentence matching vector. Detection method. According to claim 1,
The step of checking whether the query sentence and the core sentence are grammatically mapped is:
Setting the order of words in the query sentence and the order of words in the core sentence, respectively;
Morphologically analyzing words of the query sentence in which the order is set and words of the core sentence and classifying each morpheme;
Classifying the morpheme of the classified query sentence word and the morpheme of the core sentence word based on a word;
Generating a pattern of the query sentence classified based on the word and the pattern of the core sentence; And
Comprising the steps of comparing the pattern of the generated query sentence and the pattern of the core sentence to determine whether the query sentence and the core sentence are mutually deformable sentences, Korean grammar conversion-based fake news detection method. For fake news detection servers,
A memory containing a program for detecting fake news based on Korean grammar conversion; And
It includes a processor for executing the program recorded in the memory,
The processor according to the execution of the program,
Receive a query sentence and a news article to be detected, extract an article sentence related to the query sentence from the news article to be detected as a key sentence, check whether a word in the query sentence corresponds to a word in the key sentence, and According to whether the word matches, the semantic correspondence between the query sentence and the core sentence is checked, and according to the semantic correspondence, the grammatical correspondence between the query sentence and the core sentence is checked, and the identified semantic A fake news detection server based on Korean grammar conversion that determines whether the core sentence is true or false based on a result of the correspondence and the result of the grammatical correspondence. The method of claim 7,
Extracting the key sentence,
The query sentence and the article sentence are decomposed into a word configuration unit including a root and a suffix, the word composition unit of the query sentence and the word construction unit of the article sentence are vectorized and compared with each other to calculate cosine similarity, and the query A fake news detection server based on Korean grammar conversion, comprising extracting the article sentence including the word structure unit of the sentence and the word structure unit having the highest cosine similarity as the core sentence. The method of claim 7,
Checking whether the word in the query sentence corresponds to the word in the key sentence is
The query sentence is decomposed by word and stored in the order of placement of each word to generate a word matching set, and the core sentence is decomposed by word and stored in the order of placement of each word to generate a word matching set. A fake news detection server based on Korean grammar conversion, which includes comparing the word matching set and the word matching set of the core sentence and classifying the result of the comparison into full word matching and partial word matching. The method of claim 9,
When the comparison result is output as a complete word match,
A fake news detection server based on Korean grammar conversion by determining whether the query sentence and the core sentence are semantically matched and checking whether the query sentence and the core sentence are grammatically corresponding. The method of claim 7,
The semantic correspondence between the query sentence and the key sentence is checked.
The word vector of the query sentence and the word vector of the core sentence are extracted, the word vector of the extracted query sentence and the word vector of the core sentence are matched with the word embedding vector of a pre-learned Korean news article, and the matched Deep learning is applied to the word vector of the query sentence and the word vector of the core sentence to extract each as a context-related embedding vector, and deep learning is applied to the context-related embedding vector and the core sentence context-related embedding vector of the extracted query sentence. To extract the query sentence matching vector and the core sentence matching vector respectively, and apply deep learning to the extracted query sentence matching vector and the core sentence matching vector to context aggregate aggregated query matching vector and context aggregate Extracted as a core sentence matching vector (Contextual Aggregated Key Sentence Set Matching Vector), and aggregated query sentences by matching the last time step matching vector to the context aggregate query sentence matching vector and the context aggregate key sentence matching vector Extracting the Aggregated Question Matching Vector and Aggregated Key Sentence Set Matching Vector, and using artificial neural networks and normalization based on the extracted aggregate query sentence matching vector and the aggregated core sentence matching vector And calculating a semantic similarity between the query sentence and the core sentence. The method of claim 7,
Checking the grammatical correspondence between the query sentence and the key sentence,
The order of words in the query sentence and the order of words in the key sentence are respectively set, the words of the query sentence in which the order is set and the words in the key sentence are morphologically analyzed, classified into each morpheme, and the classified query sentence The morpheme of the word and the morpheme of the key sentence word are classified based on a word, and a pattern of the query sentence and the pattern of the key sentence classified based on the word are generated, and the pattern and the key sentence of the generated query sentence are A fake news detection server based on Korean grammar conversion, comprising comparing the patterns with each other and checking whether the query sentence and the core sentence are mutually deformable sentences. A computer-readable recording medium on which a program for implementing the method according to claims 1 to 6 is recorded.

Images