KR102663908B1 - Method for providing meaning search service through semantic analysis - Google Patents

Abstract

An embodiment of the present invention relates to a method of providing a semantic search service through semantic analysis, which analyzes the meaning of a query in the form of a sentence, finds sentences that are the same or similar to the query in a text document, and performs semantic analysis along with page information. Services can be provided.

Description

{Method for providing meaning search service through semantic analysis}

The present invention relates to a method of providing a meaning search service through semantic analysis.

Semantic search identifies search intent through query analysis and provides search results that match the search intent.

Currently, portal sites and search sites perform a function of extracting documents that match the entered word from the search engine when a word is entered. However, for example, when a person asks about something after reading a book, a function that searches the sentence form by semantic analysis has not yet been implemented.

From text data, it is very difficult to extract sentences corresponding to sentence queries through natural language processing, and there is a problem in that the efficiency or performance of extracting sentences matching sentence queries is not high.

Even now, when a query is made in the form of a sentence, a semantic analysis service that finds books, papers, and other materials containing sentences with similar meanings through sentence semantic analysis is not active, and it still remains a subject that needs to be studied.

Korean Patent No. 10-2285232

In order to solve this problem, the present invention analyzes the meaning of a query in the form of a sentence, finds sentences that are the same or similar to the query in a text document, and analyzes the sentence meaning provided by a semantic analysis service along with page information. The purpose is to provide a semantic analysis providing system that searches for and provides semantic analysis.

A semantic analysis providing system that analyzes and finds sentence meaning according to the characteristics of the present invention to achieve the above purpose,

A sentence acquisition unit that receives text information in the form of a sentence;

a morpheme extraction unit that separates the sentences input from the sentence acquisition unit into morphemes, which are the smallest units of words with meaning;

An associated word database unit that vectorizes and stores substitute words, synonyms, inferred words, and semantic associations of emotional morphemes related to the emotion of the word corresponding to the input word (morpheme), and vectorizes one or more subject words matching the input word. A data storage unit consisting of a keyword database unit that analyzes the meaning of words and sentences composed of words, vectorizes and stores the semantic analysis sentences classified by meaning, and stores them in a data storage unit;

a word embedding processing unit that converts the words extracted from the morpheme extraction unit into word vectors;

A sentence embedding processing unit that processes sentence vector conversion for sentences that meet preset context connection criteria considering the context of words (forward and backward connection of word meaning) based on the word vector processed by the word embedding processing unit;

The sentences that have undergone the sentence vector conversion are compared and analyzed in the semantic analysis database unit, the semantic analysis sentences classified by meaning corresponding to the sentences are vectorized, and the subject word database unit uses the vectorized semantic analysis sentences as the basis. A control unit that extracts one or more subject words;

a target document database unit that vectorizes and stores a plurality of text data of books, papers, magazines, and publications in sentence units; and

A semantic analysis server comprising a vector set representing the extracted key words and a sentence similarity calculation unit that calculates similarity between sentences using a vector set of text data in the target document database unit, wherein the control unit determines that the sentence similarity calculation unit determines the similarity between sentences. or count similar sentences as similar sentences in the text data, extract text data with a preset higher order in the order of the number of similar sentences, and select a number of page numbers of the same or similar sentences for each text data with the extracted higher order. Page notification information is generated in the first paragraph.

In addition, the sentence similarity calculation unit calculates the Euclidean distance representing the similarity between the two sentences p and q according to Equation 1 below in the sentence embedding technique, and the calculated Euclidean distance is calculated from the preset first a first sentence similarity calculator that determines that the first sentence is the same or similar if the value is greater than or equal to the threshold; and

In the sentence embedding technique, cosine similarity is calculated according to Equation 2 below, and if the calculated cosine similarity is more than a preset second threshold, second sentence similarity is judged to be the same or similar. A semantic analysis providing system that analyzes and finds the meaning of sentences containing parts.

[Equation 1]

Here, sentence p represents a sentence input to the sentence acquisition unit, sentence q represents a plurality of text data stored in the target document database unit, and doc2vec of sentence p consisting of the keyword is , and the doc2vec of sentence q is lim.

[Equation 2]

Here, a represents a sentence input to the sentence acquisition unit, and b is a plurality of text data stored in the target document database unit.

In addition, the control unit compares one or more first sentences determined to be the same or similar sentences by the first sentence similarity calculation unit and one or more second sentences determined to be the same or similar sentences by the second sentence similarity calculation unit, and One sentence and a second sentence (a complement of the first sentences) that do not overlap with the first sentence are searched from the text data of the target document database unit, and a plurality of first sentences and a first sentence are searched in the target document database unit. Non-overlapping second sentences (the complement of the first sentences) are searched and counted by the number of similar sentences, and text data in the preset upper order is extracted in the order of the counted number of similar sentences.

Through the above-described configuration, the present invention has the effect of providing a semantic analysis service that finds books, papers, and other materials containing sentences with similar meanings through semantic analysis of sentences.

The present invention can search the meaning of a sentence by encompassing all semantic related words such as the same word, substitute word, synonym, inferred word, and emotional morpheme related to the emotion of the word, thereby improving the efficiency and performance of sentence search. It works.

Figure 1 is a block diagram showing the configuration of a semantic analysis providing system that analyzes and finds sentence meaning according to an embodiment of the present invention.
Figure 2 is a block diagram showing the internal configuration of a semantic analysis server according to an embodiment of the present invention.
Figure 3 is a diagram showing the configuration of a learning model and artificial neural processing network of a semantic analysis server according to an embodiment of the present invention.

Since the present invention can make various changes and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all changes, equivalents, and substitutes included in the spirit and technical scope of the present invention. While describing each drawing, similar reference numerals are used for similar components.

Terms such as first, second, A, B, etc. may be used to describe various components, but the components should not be limited by the terms. The above terms are used only for the purpose of distinguishing one component from another. For example, a first component may be named a second component without departing from the scope of the present invention, and similarly, the second component may also be named a first component. The term “and/or” includes any of a plurality of related stated items or a combination of a plurality of related stated items.

When a component is said to be "connected" or "connected" to another component, it is understood that it may be directly connected to or connected to the other component, but that other components may exist in between. It should be. On the other hand, when it is mentioned that a component is “directly connected” or “directly connected” to another component, it should be understood that there are no other components in between.

The terms used in this application are only used to describe specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this application, terms such as “comprise” or “have” are intended to designate the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification, but are not intended to indicate the presence of one or more other features. It should be understood that it does not exclude in advance the possibility of the existence or addition of elements, numbers, steps, operations, components, parts, or combinations thereof.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by a person of ordinary skill in the technical field to which the present invention pertains. Terms defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and unless explicitly defined in the present application, should not be interpreted in an ideal or excessively formal sense. No.

Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the attached drawings. In order to facilitate overall understanding when describing the present invention, the same reference numerals are used for the same components in the drawings, and duplicate descriptions for the same components are omitted.

Figure 1 is a block diagram showing the configuration of a semantic analysis providing system that analyzes and finds sentence meaning according to an embodiment of the present invention.

The semantic analysis providing system 100 that analyzes and finds sentence meaning according to an embodiment of the present invention includes one or more electronic devices 110 that are user terminals, a communication network 120, and a semantic analysis server 130.

The plurality of electronic devices 110 may be fixed terminals implemented as computer devices or mobile terminals.

The plurality of electronic devices 110 include, for example, smart phones, mobile phones, navigation devices, computers, laptops, digital broadcasting terminals, PDAs (Personal Digital Assistants), PMPs (Portable Multimedia Players), and tablet PCs. there is. For example, the electronic device 110 may communicate with the semantic analysis server 130 through the communication network 120 using a wireless or wired communication method.

The communication network 120 is not limited to communication methods, and may include not only communication methods using mobile communication networks, wired Internet, wireless Internet, and broadcasting networks, but also short-range wireless communication between devices. For example, the communication network 120 includes a personal area network (PAN), a local area network (LAN), a campus area network (CAN), a metropolitan area network (MAN), a wide area network (WAN), a broadband network (BBN), It may include one or more arbitrary networks such as the Internet. Additionally, the communication network 120 may include any one or more of network topologies including a bus network, star network, ring network, mesh network, star-bus network, tree or hierarchical network, etc. Not limited.

The semantic analysis server 130 may be implemented as a computer device or a plurality of computer devices that communicate with a plurality of electronic devices 110 and a communication network 120 to provide commands, codes, files, content, services, etc. .

The semantic analysis server 130 may provide files for installing an application to the electronic device 110 connected through the communication network 120. In this case, the electronic device 110 may install the application using a file provided from the semantic analysis server 130.

In addition, the electronic device 110 connects to the semantic analysis server 130 under the control of an operating system (OS) and at least one program (for example, a browser or the installed application) to establish the semantic analysis server 130. You can receive services or content provided by . For example, when the electronic device 110 transmits a service request message to the semantic analysis server 130 through the communication network 120 under the control of an application, the semantic analysis server 130 sends a code corresponding to the service request message. It can be transmitted to the electronic device 110, and the electronic device 110 can provide content to the user by configuring and displaying a screen according to the code under the control of the application.

The electronic device 110 receives a sentence consisting of words from the user, semantically analyzes the meaning of the sentence input from the semantic analysis server 130, and searches text data of books, papers, magazines, and publications to find similar sentences. Semantic analysis services are provided.

The semantic analysis server 130 receives information in the form of a sentence consisting of one or more words from the electronic device 110, analyzes the meaning of the input sentence, and searches for similar sentences in text data of books, papers, magazines, and publications. Perform analysis.

Figure 2 is a block diagram showing the internal configuration of a semantic analysis server according to an embodiment of the present invention.

The semantic analysis server 130 according to an embodiment of the present invention includes a sentence acquisition unit 131, a morpheme extraction unit 132, a data storage unit 133, a control unit 134, a word embedding processing unit 136, and a sentence embedding processing unit. (137), a target document database unit 135, a sentence similarity calculation unit 138, and a communication unit.

The sentence acquisition unit 131 receives information in the form of a sentence consisting of one or more words from the electronic device 110. For example, it represents information in the form of sentences, such as “I’m depressed today, what book should I read?” or “I want to know about 19th century Indian philosophy.”

The morpheme extraction unit 132 separates the sentence input from the sentence acquisition unit 131 into morphemes, which are the smallest units of words with meaning.

At this time, the morpheme extractor 132 performs filtering on unnecessary words, that is, stop words, in order to improve sentence discrimination.

Stop words are words that are included with high frequency in the majority of sentences and may consist of any one of a particle, ending, prefix, or suffix, and may include words specified according to user settings.

The data storage unit 133 is an associated word database unit that vectorizes and stores semantic related words such as substitute words, synonyms, inferred words, and emotional morphemes related to the emotion of the word corresponding to the input word (morpheme) through word2vec or doc2vec. (133a) and a subject word database unit 133b that vectorizes and stores one or more subject words matching the input word through word2vec or doc2vec, and meets the preset context connection standard considering the context of the words (connection before and after the word meaning) and a semantic analysis database unit 133c that analyzes the meaning of words and sentences composed of words, vectorizes and stores the semantic analysis sentences classified by meaning.

For example, if the input sentence is 19th century Indian philosophy, it is separated into morphemes of 19th century + India + philosophy, the meaning of the sentence is analyzed, and stored in the semantic analysis database unit 133c.

The target document database unit 135 stores a plurality of text data such as books, papers, magazines, publications, etc. by converting them into vectors in sentence units.

The target document database unit 135 includes a text conversion unit 135a, a page number identification unit 135b, and a storage unit 135c.

The text converter 135a converts a PDF file, which is input text data, into a text file (TXT file).

The page number identification unit 135b generates a page number every time the text file is converted in the text conversion unit 135a, inserts an identification tag at the start of the first page of the text file, and inserts an identification tag for each paragraph. You can identify pages by inserting them in order.

The page number identification unit 135b stores a text file into which page information consisting of an identification tag and page number is inserted in the storage unit 135c.

The word embedding processing unit 136 processes words extracted from the morpheme extraction unit 132 into word vectors.

The sentence embedding processing unit 137 processes sentence vector conversion for sentences that meet preset context connection criteria considering the context of words (connection before and after the word meaning) based on the word vector processed by the word embedding processing unit 136. .

The word embedding processing unit 136 and the sentence embedding processing unit 137 vectorize words and sentences using a neural network-based embedding algorithm. Here, the embedding algorithm can utilize well-known embedding technologies such as doc2vec, word2vec, and sense2vec.

The control unit 134 compares and analyzes sentences that have undergone sentence vector conversion in the semantic analysis database unit 133c, vectorizes the semantic analysis sentences classified by meaning corresponding to the sentences, and vectorizes them in the keyword database unit 133b. One or more topic words are extracted based on the semantic analysis sentence. Here, the keyword refers to a word or phrase that represents the central idea in a certain text or literary work.

If the input sentence is "I want to know about 19th century Indian philosophy," the control unit 134 classifies 19th century + India + philosophy by meaning in the semantic analysis database unit 133c to extract a semantic analysis sentence, and the keyword database unit ( Using the semantic analysis sentence extracted in 133b), Sankhya philosophy, Hindu philosophy, etc. related to the semantic analysis sentence are extracted as keywords.

If the input sentence is "I'm depressed today, which book should I read?", the control unit 134 classifies depression + book + read by meaning in the semantic analysis database unit 133c to extract a semantic analysis sentence, and the keyword database unit Using the semantic analysis sentence extracted from (133b), keywords such as discouragement, unhappiness, sad, misery, loneliness, despair, regret, etc. are extracted in relation to the semantic analysis sentence.

The sentence similarity calculation unit 138 calculates the similarity between sentences using a vector set representing the extracted subject word and a vector set of text data of the target document database unit 135, and calculates the similarity between the sentences using the first sentence similarity calculation unit 138a and the second sentence. It includes a similarity calculation unit 138b.

The first and second sentence similarity calculation units 138a and 138b may calculate similarity scores between sentences using a set of vectors representing key words.

The first sentence similarity calculation unit 138a calculates the doc2vec of the sentence p consisting of the keyword. , and the doc2vec of sentence q is When , the Euclidean distance representing the first degree of similarity between two sentences p and q can be defined as Equation 2 below.

Sentence p represents a sentence input to the sentence acquisition unit 131, and sentence q represents a plurality of text data stored in the target document database unit 135.

The first sentence similarity calculation unit 138a calculates the Euclidean Distance indicating the similarity between the two sentences p and q using the following equation 1 in the sentence embedding technique, and the calculated Euclidean distance is If it is more than the set first threshold, it is judged as the same or similar first sentence.

The second sentence similarity calculation unit 138b calculates cosine similarity according to the following equation 2 in the sentence embedding technique, and when the calculated cosine similarity is greater than or equal to a preset second threshold, the second sentence similarity is the same or similar. Judge by sentence.

Here, a represents a sentence input to the sentence acquisition unit 131, and b represents a plurality of text data stored in the target document database unit 135.

The control unit 134 counts the same or similar sentences determined by the sentence similarity calculation unit 138 as similar sentences in the text data, extracts text data with a preset higher order number in order of the higher number of similar sentences, and For each text data, page notification information is generated indicating the paragraph number of the page number of the same or similar sentences.

The control unit 134 combines one or more first sentences determined by the first sentence similarity calculation unit 138a to be the same or similar sentences and one or more second sentences determined by the second sentence similarity calculation unit 138b to be the same or similar sentences. After comparison, the first sentence and the second sentence (the complement of the first sentences) that do not overlap with the first sentence are searched from the text data of the target document database unit 135.

The control unit 134 searches for and counts a plurality of first sentences and second sentences (the complement of the first sentences) that do not overlap with the first sentences as the number of similar sentences in the target document database unit 135.

The control unit 134 extracts text data in a preset higher order in the order of the greater number of similar sentences counted. Here, text data refers to documents capable of natural language processing, such as books, papers, magazines, and publications, stored in the target document database unit 135.

For example, the top five text data with the highest number of similar sentences are extracted, and similar sentences are displayed in color for each text data.

The control unit 134 displays the first and second sentences, which are determined to be the same or similar to each extracted text data, in color on the block screen area, and uses page information for each displayed sentence to determine which paragraph of the page number it is. Create page notification information.

The control unit 134 generates result information including page notification information and the type of text data and transmits it to the electronic device 110 through the communication unit 139.

In another embodiment, the control unit 134 substitutes the Euclidean Distance calculated by Equation 1 and the Cosine Similarity calculated by Equation 2 into Equation 3 below to create a sentence Calculate the search index. The sentence search index represents a similarity value indicating the degree of similarity to the input sentence.

The control unit 134 calculates a sentence search index for each of the above-described first and second sentences, and displays the calculated sentence search index in a block screen area with different colors according to the preset search range.

For example, the range of the sentence search index is green for the first value to the second value, blue for the second value to the third value, yellow for the third value to the fourth value, and yellow for the fourth value to the fifth value. In this case, the sentence can be displayed in a different color, in red.

Here, W ₁ is a preset weight related to the Euclidean distance when searching for sentence similarity, W ₂ is a preset weight related to cosine similarity when searching for sentence similarity, ED is the Euclidean Distance, and CS is cosine similarity.

A method of extracting one or more subject words corresponding to morphemes from the subject word data storage unit 133 based on the separated morphemes of the input sentence will be described as follows.

The control unit 134 calculates the importance index of words included in text data according to the search frequency of the input word and the number of semantically related words corresponding to the input word (morpheme), as shown in Equation 4 below.

If the calculated importance index is greater than or equal to a preset third threshold, the control unit 134 determines that the word included in the text data corresponds to a key word matching the word in the input sentence and stores it in the key word database unit 133b.

The importance index is an indicator necessary to select a key word stored in the key word data storage unit 133.

Here, W ₃ is a weight related to the search frequency when extracting a keyword corresponding to a word in an input sentence from text data, and W ₄ is a weight related to the search frequency when extracting a keyword corresponding to a word in an input sentence from text data. It is a weight related to Semantic Relevance, FR is the number of search frequencies, and SR is Semantic Relevance indicating the number of semantic related words.

The importance index has a higher index value as the frequency with which words in the input sentence are searched in the text data increases, and the higher the number of semantic related words associated with the words in the input sentence in the text data, the higher the index value.

The control unit 134 selects text data with the largest number of similar sentences counted in conjunction with the target document database unit 135, generates result information including the type of selected text data and page notification information, and sends the communication unit 139 to the communication unit 139. It is transmitted to the electronic device 110 through.

The control unit 134 calculates the sentence search index of the first sentences and the second sentences (the complement of the first sentences) that do not overlap with the first sentence in each text data according to the above-described equation 3, and performs sentence search. It is determined whether the index is greater than or equal to a preset standard value, and the number of sentence search indices that are greater than or equal to the standard value is counted. The control unit 134 selects text data with the highest counted sentence search index, generates result information including the type of selected text data and page notification information, and transmits it to the electronic device 110 through the communication unit 139.

When a sentence is input, a method of generating topic words related to the input sentence using artificial intelligence, a method of generating a sentence vector composed of words and a vectorized semantic analysis sentence corresponding thereto, are detailed with reference to FIG. 3 below. Explain clearly.

Figure 3 is a diagram showing the configuration of a learning model and artificial neural processing network of a semantic analysis server according to an embodiment of the present invention.

The semantic analysis server 130 according to an embodiment of the present invention includes a control unit 134, a data collection unit 140, a display unit 150, a learning model unit 160, and an artificial neural processing network 170.

The data collection unit 140 receives and stores sentence vectors composed of words, vectorized semantic analysis sentences, and keywords of text data corresponding thereto.

The semantic analysis server 130 inputs a sentence vector and a vectorized semantic analysis sentence into the artificial neural processing network 170, and provides one or more subject words of text data corresponding to the semantic analysis sentence in response to the artificial neural processing network 170. outputs.

The data set stored in the data collection unit 140 is further divided into a training set and a test set. A training set is provided to a machine learning or deep learning model.

The learning model unit 160 includes a data processing unit 161, a learning unit 162, and a classification unit 163.

The artificial neural processing network 170 includes an input layer 171, a hidden layer 172 consisting of a convolution layer 173, a pooling layer 174, and a fully connected layer 175, and an output layer 176. do.

The data processing unit 161 receives a training set consisting of sentence vectors and corresponding vectorized semantic analysis sentences from the data set from the data collection unit 140 and transmits them to the artificial neural processing network 170. The training set represents learning data.

The data processing unit 161 receives a training set consisting of sentence vectors and corresponding vectorized semantic analysis sentences from the data set from the data collection unit 140 and transmits them to the artificial neural processing network 170.

The data processing unit 161 receives a test set consisting of sentence vectors and corresponding vectorized semantic analysis sentences from the data set from the data collection unit 140 and transmits them to the classification unit 163.

The data processing unit 161 may be formed as a database unit capable of distributed parallel processing.

The artificial neural processing network 170 inputs and applies a training set consisting of sentence vectors and corresponding vectorized semantic analysis sentences to the artificial neural processing network 170 to correct errors and use the corrected errors. It outputs whether the prediction result of key word generation of text data is correct.

The artificial neural processing network 170 inputs and applies a training set consisting of sentence vectors and corresponding vectorized semantic analysis sentences to the artificial neural processing network 170 to correct errors and use the corrected errors. The result of generating keywords from text data corresponding to the semantic analysis sentence is output.

At this time, the artificial neural processing network 170 uses deep convolutional neural networks (CNNs) and may include an input layer 171, a hidden layer 172, and an output layer 176.

The artificial neural processing network 170 uses a neural network-based model for predictive analysis.

The artificial neural processing network 170 includes an input layer 171 x, an output layer 176 y, and an arbitrary amount of hidden layers 172 containing four neurons.

Each layer consists of a set of biases and weights, denoted by band W, except the output layer 176. The sigmoid function is used as the activation function of each hidden layer. Bias and weight fine-tuning of the input data is performed to improve the model's prediction score. Each iteration in the training process includes the following steps:

The training process of the neural network model is performed, which consists of two steps: feed-forward, which includes calculation of the predicted output layer 176 y, and back-propagation, which updates weights and biases.

The artificial neural processing network 170 performs backpropagation to measure the error of the prediction loss and measures the prediction error (loss).

Differentiation of the loss function with respect to the bias and weights is used to adjust the weights and biases.

The input layer 171 acquires the training data stored in the data processing unit 161 and stores the acquired training data as a layer with a feature map. Here, the feature map can facilitate connection with the hidden layer 172, which will be described later, by having a structure in which multiple nodes are arranged in two dimensions.

The hidden layer 172 acquires the feature map of the layer located in the upper layer, and gradually extracts higher-level features from the obtained feature map. The hidden layer 172 may be formed of one or more layers and includes a convolutional layer 173, a pooling layer 174, and a fully connected layer 175.

The convolution layer unit 173 is a component that performs a convolution operation from learning data and includes a feature map connected to a plurality of input feature maps.

The pooling layer unit 174 is configured to receive the output of the convolution layer unit 173 as an input and perform a convolution operation, that is, a sub-sampling operation, and the convolution layer unit 173 located in the lower layer of the hidden layer 172 It contains the same number of feature maps as the number of input feature maps, and each feature map is connected one-to-one with the input feature map.

The fully connected layer unit 175 receives the output of the convolution layer unit 173 as an input and learns according to the output for each category output from the output layer 130, and synthesizes the learned local information, that is, features. to learn abstract content.

At this time, when the hidden layer 172 has a pooling layer unit 172, the polling connected layer unit 175 is connected to the polling layer unit 174, and features are synthesized from the output of the polling layer unit 174. Learn abstract content.

The output layer 176 maps the output for each category desired to be classified into a probability value using a function such as soft-max. At this time, the results output from the output layer 176 may be transmitted to the learning unit 162 or the classification unit 163 to perform error backpropagation or may be output as response data.

The learning unit 162 performs supervised learning. Supervised learning applies a machine learning algorithm to learning data to infer a function and finds an answer through the inferred function.

The learning unit 162 creates a linear model representing the learning data through supervised learning, and can predict future events through the linear model.

The learning unit 162 determines how the new data is classified with the previously learned data based on the previously learned data.

The learning unit 162 performs learning of the artificial neural processing network 170 on a training set consisting of sentence vectors and corresponding vectorized semantic analysis sentences from the data set from the data processing unit 161, and performs deep learning by type. Using learning feature values, we learn whether to generate topic words in text data corresponding to semantic analysis sentences.

The learning unit 162 learns from the artificial neural processing network 170 whether the subject word of the text data corresponding to the semantic analysis sentence is generated using deep learning feature values for each type.

In one embodiment of the present invention, learning of the artificial neural processing network 170 is performed through supervised learning.

Supervised learning is a method of inputting learning data and corresponding output data together into the artificial neural processing network 170, and updating the weights of connected edges so that output data corresponding to the learning data is output. For example, the artificial neural processing network 170 of the present invention can update connection weights between artificial neurons using delta rules and error back-propagation learning.

Error-back-propagation learning estimates the error using feed-forward for the given learning data, then starts from the output layer and estimates in the reverse direction toward the hidden layer 172 and the input layer 171. It propagates an error and updates the connection weights between artificial neurons in a way that reduces the error.

The artificial neural processing network 170 calculates the error from the results obtained through the input layer 171 - hidden layer 172 - polling connected layer unit 175 - output layer 176, and corrects the calculated error. Again, the connection weight can be updated by propagating the error in the following order: output layer 176 - polling connected layer unit 175 - hidden layer 172 - input layer 171.

The learning unit 162 uses a training set consisting of a sentence vector and the corresponding vectorized semantic analysis sentence as an input vector, and when it passes through the input layer 171, the hidden layer 172, and the output layer 176, It is learned through supervised learning to generate an output vector to determine whether the subject word of the text data corresponding to the semantic analysis sentence is generated.

The learning unit 162 uses the artificial neural processing network 170 to input sentence vectors and the corresponding vectorized semantic analysis sentence feature values as input vectors, and includes an input layer 171, a hidden layer 172, and an output layer. If (176) is passed, the subject word generation result of the text data corresponding to the semantic analysis sentence is learned through supervised learning to generate an output vector.

The learning unit 162 uses the subject word generation result of the text data corresponding to the semantic analysis sentence as learning data and learns it with artificial intelligence in conjunction with the artificial neural processing network 170.

The artificial neural processing network 170 knows in advance whether an output value (a subject word generation result of text data corresponding to the semantic analysis sentence) should be produced when an input value (a sentence vector and a corresponding vectorized semantic analysis sentence) is input.

The classification unit 163 may output output data of the artificial neural processing network 170 with connection weights updated through error backpropagation in the learning unit 162 as response data.

When training data, test data, or new data not used for learning are input to the artificial neural processing network 170 with updated connection weights, the classification unit 163 performs an input layer 171 - a hidden layer 172 - a polling connection. Tied layer unit 175 - The result output through the output layer 176 can be obtained and output as response data.

The artificial neural processing network 170 generates a deep learning-based classifier model through optimization based on the input sentence vector and the subject word generation result of the text data corresponding to the corresponding vectorized semantic analysis sentence.

The learning unit 162 may differently apply individual element weights regarding the layers in the artificial neural processing network and the strength of connections between layers according to the sentence vector and the corresponding vectorized semantic analysis sentence.

The learning unit 162 is trained through supervised learning to generate the key word generation results of text data as an output vector, calculates in the direction from the input layer 171 to the output layer 176, and conversely, from the output layer 176 to the input layer ( 171) The error is minimized by modifying the weight while repeating the calculation in the direction.

The classification unit 163 uses the input sentence vector, which is test data, and the corresponding vectorized semantic analysis sentence using the deep learning-based classifier model of the artificial neural processing network 170 to generate the result value of the response data (key words of text data). result).

The classification unit 163 determines whether the text data generates a subject word in response to the vectorized semantic analysis sentence.

The classification unit 163 uses the input sentence vector, which is test data, and the corresponding vectorized semantic analysis sentence using the deep learning-based classifier model of the artificial neural processing network 170 to determine whether the text data is a subject word generation result. .

The output unit 164 displays on the display unit 150 whether there is a result of generating a key word of text data corresponding to the semantic analysis sentence received from the classification unit 163.

The key word database unit 133b stores one or more key words of input sentences, semantic analysis sentences, and text data corresponding thereto.

Experts such as librarians and critics read text data such as books, papers, magazines, and publications and extract multiple keywords.

The data collection unit 140 receives and stores a plurality of key words extracted from each text data by experts through the electronic device 110.

When the control unit 134 outputs one or more first subject words of text data corresponding to the semantic analysis sentence in response to the artificial neural processing network 170, the output first subject word and the experts extracted from the data collection unit 140 Compare and analyze the second key word.

If the first key word and the second key word are different, the control unit 134 updates the text data by adding a different second key word to the key word of the text data in the key word database unit 133b.

Operations according to embodiments of the present specification can be implemented as a computer-readable program or code on a computer-readable recording medium. Computer-readable recording media include all types of recording devices that store data that can be read by a computer system. Additionally, computer-readable recording media can be distributed across networked computer systems so that computer-readable programs or codes can be stored and executed in a distributed manner.

When the embodiment is implemented in software, the above-described techniques may be implemented as modules (processes, functions, etc.) that perform the above-described functions. Modules are stored in memory and can be executed by a processor. Memory may be internal or external to the processor, and may be connected to the processor by a variety of well-known means.

Additionally, computer-readable recording media may include hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. Program instructions may include not only machine language code such as that created by a compiler, but also high-level language code that can be executed by a computer using an interpreter, etc.

Although some aspects of the invention have been described in the context of an apparatus, it may also refer to a corresponding method description, where a block or device corresponds to a method step or feature of a method step. Similarly, aspects described in the context of a method may also be represented by corresponding blocks or items or features of a corresponding device. Some or all of the method steps may be performed by (or using) a hardware device, such as a microprocessor, programmable computer, or electronic circuit, for example. In some embodiments, one or more of the most important method steps may be performed by such an apparatus.

In embodiments, a programmable logic device (e.g., a field programmable gate array) may be used to perform some or all of the functionality of the methods described herein. In embodiments, a field programmable gate array may operate in conjunction with a microprocessor to perform one of the methods described herein. In general, the methods are preferably performed by some hardware device.

Although the present invention has been described above with reference to preferred embodiments, those skilled in the art may make various modifications and changes to the present invention without departing from the spirit and scope of the present invention as set forth in the claims below. You will understand that you can do it.

100: Semantic analysis provision system 110: Electronic device
120: Communication network 130: Semantic analysis server
131: Sentence acquisition unit 132: Morpheme extraction unit
133: data storage unit 134: control unit
135: Target document database unit 136: Word embedding processing unit
137: Sentence embedding processing unit 138: Sentence similarity calculation unit
139: Communication Department 140: Data Collection Department
150: display unit 160: learning model unit
170: Artificial neural processing network

Claims (1)

A sentence acquisition unit that receives text information in the form of a sentence;
a morpheme extraction unit that separates the sentences input from the sentence acquisition unit into morphemes, which are the smallest units of words with meaning;
An associated word database unit that vectorizes and stores substitute words, synonyms, inferred words, and semantic associations of emotional morphemes related to the emotion of the word corresponding to the input word (morpheme), and vectorizes one or more subject words matching the input word. A data storage unit consisting of a keyword database unit that analyzes the meaning of words and sentences composed of words, vectorizes and stores the semantic analysis sentences classified by meaning, and stores them in a data storage unit;
a word embedding processing unit that converts the words extracted from the morpheme extraction unit into word vectors;
A sentence embedding processing unit that processes sentence vector conversion for sentences that meet preset context connection criteria considering the context of words (front and back connection of word meanings) based on the word vector processed by the word embedding processing unit;
The sentences that have undergone the sentence vector conversion are compared and analyzed in the semantic analysis database unit, the semantic analysis sentences classified by meaning corresponding to the sentences are vectorized, and the subject word database unit uses the vectorized semantic analysis sentences as the basis. A control unit that extracts one or more subject words;
a target document database unit that vectorizes and stores a plurality of text data of books, papers, magazines, and publications in sentence units; and
A semantic analysis server comprising a vector set representing the extracted key words and a sentence similarity calculation unit that calculates similarity between sentences using a vector set of text data in the target document database unit, wherein the control unit determines that the sentence similarity calculation unit determines the similarity between sentences. or count similar sentences as similar sentences in the text data, extract text data with a preset higher order in the order of the number of similar sentences, and select a number of page numbers of the same or similar sentences for each text data with the extracted higher order. Generate page notification information for the first paragraph,
The sentence similarity calculation unit calculates a Euclidean distance indicating the similarity between two sentences p and q according to Equation 1 below in the sentence embedding technique, and the calculated Euclidean distance is set to a preset first threshold. a first sentence similarity calculator that determines that the first sentence is the same or similar if the value is greater than or equal to the value; and
In the sentence embedding technique, cosine similarity is calculated using Equation 2 below, and if the calculated cosine similarity is greater than or equal to a preset second threshold, second sentence similarity is judged to be the same or similar to the second sentence. Contains wealth,
[Equation 1]

The sentence p represents a sentence input to the sentence acquisition unit, the sentence q represents a plurality of text data stored in the target document database unit, and the doc2vec of the sentence p consisting of the keyword is , and the doc2vec of sentence q is lim.
[Equation 2]

Wherein a represents a sentence input to the sentence acquisition unit, and b is a plurality of text data stored in the target document database unit,
The control unit compares one or more first sentences determined to be the same or similar sentences by the first sentence similarity calculation unit and one or more second sentences determined to be the same or similar sentences by the second sentence similarity calculation unit, and Search the text data of the target document database unit for a sentence, a second sentence (a complement of the first sentences) that does not overlap with the first sentence, and search for a plurality of first sentences and a plurality of first sentences that overlap with the first sentence in the target document database unit. Search for and count the second sentences (the complement of the first sentences) that are not used by the number of similar sentences, and extract text data in the preset upper order in the order of the counted number of similar sentences,
The target document database unit includes a text conversion unit that converts a PDF file, which is input text data, into a text file (TXT file); and
The text conversion unit generates a page number every time the text file is converted, inserts an identification tag at the start of the first page of the text file, and inserts the identification tags in order for each paragraph to identify the page. It further includes a page number identification unit that stores a text file containing page information consisting of an identification tag and a page number in a storage unit,
The semantic analysis server further includes a data collection unit that receives and stores sentence vectors consisting of words, vectorized semantic analysis sentences, and keywords of text data corresponding thereto, the sentence vectors, and the vectorized semantic analysis sentences. A semantic analysis providing system that analyzes and finds the meaning of a sentence by inputting it into an artificial neural processing network and outputting one or more keywords of text data corresponding to the semantic analysis sentence in response to the artificial neural processing network.

Images