KR102370729B1 - Sentence writing system - Google Patents

Abstract

Disclosed is a sentence writing system. The system comprises: a user terminal; and a sentence writing apparatus constructing a database storing a document of which copyright is expired, acquiring an input sentence, which is a sentence input by a user, from the user terminal, and extracting a sentence similar with the input sentence as an exemplary sentence from the database to output the sentence from the user terminal.

Description

Sentence writing system {SENTENCE WRITING SYSTEM}

The present invention relates to a sentence writing system, and more particularly, to a sentence writing system for presenting sentences that can improve a user's writing ability using data whose copyright has expired.

Most of the serviced writing systems are related to translation, and for example, when a person who has acquired one language such as Hangul wants to learn another language, the system is mainly useful.

However, even if you know the word you want to express, you need an auxiliary means of presenting various sentences and words so that you can more easily understand the form that changes depending on the usage, and also a method of presenting example sentences suitable for the user's level.

Meanwhile, various methods for identifying similar sentences have been proposed.

According to (Patent Document 1), a method for identifying similar sentences using a machine translation technique is presented, and according to (Patent Document 2), a method for analyzing a sentence using artificial intelligence is presented.

There is a need to develop a writing system capable of presenting a sentence similar to a sentence input by a user through such similar sentence identification.

US 7,412,385 B2 US 2021/0042586 A1

One aspect of the present invention discloses a sentence writing system that constructs a database storing data whose copyright has expired, extracts a sentence similar to a user input sentence from the database, and outputs it to a user terminal.

The technical problems of the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description.

Sentence writing system according to an embodiment of the present invention is a user terminal; and constructing a database storing data whose copyright has expired, obtaining an input sentence that is a sentence input by the user from the user terminal, extracting a sentence similar to the input sentence from the database as an example sentence, and outputting it from the user terminal It includes a sentence writing device.

On the other hand, the sentence writing apparatus learns the sentences stored in the database according to the Word2Vec algorithm, builds a neural network that extracts a vector value representing context information for the input sentence, and sends the input sentence obtained from the user terminal to the neural network. input, calculating an input sentence vector value indicating context information for the input sentence, calculating a similarity between example sentence vector values indicating context information of each sentence stored in the database and the input sentence vector value, and the database Extracting a sentence having an example sentence vector value having the highest similarity with the input sentence vector value among the sentences stored in the as the example sentence and transmitting it to the user terminal,

an auxiliary output device connected to the user terminal through a wireless communication method to receive and output the example sentence received from the sentence writing device in the user terminal; and an installation module for installing the auxiliary output device on a table on which the user terminal is installed;

The installation module, the installation bar provided in the form of a rectangular parallelepiped; a seating part formed in the form of a plate of a predetermined thickness at one end of the installation bar, and forming a seating groove in the lower part of the auxiliary output device to be inserted and installed; and an installation bracket installed at the other end of the installation bar to enable rotation of the installation bar;

The installation bracket, the bottom surface; a pair of wall surfaces formed upwardly from both ends of the bottom surface and having an upper end formed in a semicircular shape; and a plurality of insertion grooves concave inwardly on the edge of each of the pair of wall surfaces and formed in a spiral shape with a locking groove formed therein;

The installation bar, forming a long hole at the other end, is rotatably installed between the pair of wall surfaces through a rotation pin penetrating the long hole from the outer surface of the pair of wall surfaces, the insertion groove at the other end It is possible to form an insertion protrusion that is provided to be inserted into the.

In addition, the user terminal, a sentence writing application operated by the sentence writing device may be loaded and executed.

In addition, the user terminal may be loaded with a sentence writing application having an interface in which an input region for receiving a sentence from a user and an output region for outputting the example sentence are arranged side by side and may be executed.

Also, the sentence writing apparatus may include an input unit configured to receive the input sentence from the user terminal.

In addition, the sentence writing apparatus may include a DB construction unit for accessing a web site provided to use the copyrighted material for which the copyright has expired and collecting the copyrighted material.

Also, the sentence writing apparatus may include a DB construction unit that collects the data whose copyright has expired, extracts the sentences, and stores the extracted sentences in the database.

In addition, the sentence writing apparatus includes a DB construction unit that collects the data whose copyright has expired, and the DB construction unit, a similar sentence processing system for identifying a similar sentence relationship between the copyrighted materials; and, the similar sentence processing system comprising: accessing a plurality of documents; identifying, from the plurality of documents, a cluster of related texts written by different authors on a common subject; receiving the cluster of related texts; selecting a set of text segments from the cluster; and identifying sentence-like relationships between texts in text segments included in the set of related text segments using textual alignment.

In addition, the sentence writing apparatus includes a DB construction unit that collects the data whose copyright has expired; , wherein the mapping prediction system interprets m sentences (m is an arbitrary integer greater than or equal to 2) and extracts n words (n is an arbitrary integer greater than or equal to 2) from the m sentences. wealth; a sentence vector calculation unit for calculating m sentence vectors comprising q axis components by vectorizing the m sentences in q dimensions (q is an arbitrary integer greater than or equal to 2) according to a predetermined rule; a word vector calculating unit for calculating n word vectors comprising q axis components by vectorizing the n words in q dimensions according to a predetermined rule; an index value calculating unit calculating m × n similarity index values reflecting the relationship between the m sentences and the n words by taking the dot product of the m sentence vectors and the n word vectors, respectively; Classifying the m sentences into a plurality of maps based on a sentence indicator value group including n similarity indicator values for one sentence using the m×n similarity indicator values calculated by the indicator value calculating unit. a classification model generation unit generating a classification model for a prediction data input unit for inputting one or more sentences as prediction data as prediction data; and a similarity index value obtained by executing the processing of the word extraction unit, the sentence vector calculation unit, the word vector calculation unit, and the index value calculation unit on the prediction data input by the prediction data input unit, the classification and an event prediction unit that predicts any one of the plurality of events from the prediction target data by applying to the classification model generated by the model generation unit.

In addition, the sentence writing apparatus further includes a sentence display unit that extracts a sentence similar to the input sentence from the database as an example sentence, wherein the sentence display unit includes a vector value representing the sentences stored in the database and the input sentence The example sentence may be extracted by calculating a Euclidean distance between vector values representing .

In addition, the sentence writing apparatus further includes a sentence display unit that extracts a sentence similar to the input sentence from the database as an example sentence, wherein the sentence display unit includes a vector value representing the sentences stored in the database and the input sentence The example sentence may be extracted by calculating cosine similarity between vector values representing .

In addition, the sentence writing apparatus further includes a sentence display unit that extracts a sentence similar to the input sentence from the database as an example sentence, wherein the sentence display unit includes a vector value representing the sentences stored in the database and the input sentence The example sentence may be extracted by calculating a Tanimoto coeffieient between vector values representing .

In addition, the network provided as a wired communication network for connecting the user terminal and the sentence writing device; may further include.

In addition, the network provided as a mobile communication network connecting the user terminal and the sentence writing device; may further include.

In addition, a network provided as a Wibro (Wireless Broadband) network for connecting the user terminal and the sentence writing apparatus; may further include.

In addition, a network provided as a High Speed Downlink Packet Access (HSDPA) network for connecting the user terminal and the sentence writing apparatus may be further included.

In addition, a network provided as a satellite communication network connecting the user terminal and the sentence writing device; may further include.

In addition, a network provided as a Wi-Fi (Wireless Fidelity) network connecting the user terminal and the sentence writing device; may further include.

The apparatus may further include a voice output device connected to the user terminal in a wireless communication method, and receiving and outputting a voice file for the example sentence from the user terminal.

On the other hand, according to another embodiment of the present invention, there is provided a sentence writing method in a sentence writing device connected to a user terminal and a network, the method comprising: collecting data whose copyright protection period has expired; storing the collected data in a database; receiving a sentence from a user through the user terminal; extracting a sentence similar to a sentence input from a user from the database using a neural network that outputs a vector value representing context information of the input sentence; and outputting the extracted sentence to the user terminal.

According to the present invention described above, by outputting an example sentence similar to the user input sentence while the user is inputting a sentence, it is possible to help the user to write an improved sentence with reference to this, and also to improve the sentence power of the user .

1 is a conceptual diagram of a sentence writing system according to an embodiment of the present invention.
FIG. 2 is a conceptual diagram of the sentence writing apparatus shown in FIG. 1 .
3 is a flowchart of a sentence writing method according to an embodiment of the present invention.
4 is a diagram showing an example of a similar sentence processing system according to an embodiment of the present invention.
FIG. 5 is a flowchart illustrating the operation of the system shown in FIG. 4 .
6 is a diagram for explaining alignment between pairs of two sentences according to an embodiment of the present invention.
7 is a view showing another example of a similar sentence processing system according to an embodiment of the present invention.
8 is a block diagram of a mapping prediction system according to an embodiment of the present invention.
9 and 10 are flowcharts illustrating an operation example of a mapping prediction system according to an embodiment of the present invention.
11 is a block diagram of a mapping prediction system according to another embodiment of the present invention.
12 is a diagram illustrating an audio output device according to an embodiment of the present invention.
13 is a view showing an installation module according to an embodiment of the present invention.
14 to 17 are schematic diagrams of a neural network used in the present invention.
18 to 20 are schematic diagrams of the Word2Vec algorithm used in the present invention.

Advantages and features of the present invention and methods of achieving them will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in a variety of different forms, and only these embodiments allow the disclosure of the present invention to be complete, and common knowledge in the technical field to which the present invention belongs It is provided to fully inform the possessor of the scope of the invention, and the present invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout.

The terminology used herein is for the purpose of describing the embodiments and is not intended to limit the present invention. In this specification, the singular also includes the plural unless specifically stated otherwise in the phrase. As used herein, "comprises" and/or "comprising" refers to the stated elements, steps, and acts do not exclude the presence or addition of one or more other elements, steps and acts.

1 is a conceptual diagram of a sentence writing system according to an embodiment of the present invention.

Referring to FIG. 1 , a sentence writing system 1000 according to an embodiment of the present invention may include a user terminal 10 , a network 20 , a sentence writing apparatus 30 , and a database 40 .

The user terminal 10 and the sentence writing device 30 may be connected to the network 20 , and the database 40 may store data required by the sentence writing device 30 .

The user terminal 10 may be provided as a general computing terminal such as a PC, a notebook computer, a netbook, a mobile device, a tablet, etc. used by the user.

The user terminal 10 may be loaded with a sentence writing application according to an embodiment of the present invention and executed. A user may write a predetermined sentence by executing a sentence writing application through the user terminal 10 .

The network 20 may be one of a wired communication network, a mobile communication network, a Wibro (Wireless Broadband) network, a High Speed Downlink Packet Access (HSDPA) network, a satellite communication network, and a Wi-Fi (Wireless Fidelity) network.

The sentence writing apparatus 30 may be provided as a normal server, and may be an operating server of the sentence writing application according to an embodiment of the present invention.

The sentence writing apparatus 30 may be connected to the user terminal 10 through the network 20 , and may receive a sentence input through the sentence writing application from the user terminal 10 . In addition, the sentence writing device 30 builds a database 40 in which copyright has expired (eg, novel, etc.) is stored, and extracts a sentence similar to the user input sentence from the database 40 to the user terminal 10 can be sent to In this case, the sentence writing application executed in the user terminal 10 outputs the example sentence received from the sentence writing device 30 so that the user can write an improved sentence by referring to it.

In the following description, a sentence input by the user through the user terminal 10 will be referred to as a user input sentence, and a sentence extracted from the database 40 will be referred to as an example sentence.

As such, the sentence writing system 1000 according to an embodiment of the present invention outputs an example sentence similar to the user input sentence while the user inputs the sentence, thereby helping the user to write an improved sentence with reference to this. In addition, it promotes the improvement of the user's sentence power.

FIG. 2 is a conceptual diagram of the sentence writing apparatus shown in FIG. 1 .

Referring to FIG. 2 , the sentence writing apparatus 30 according to an embodiment of the present invention may include an input unit 31 , a DB construction unit 33 , and a sentence display unit 35 .

The input unit 31 may receive a user input sentence that is a sentence input by the user through the sentence writing application from the user terminal 10 .

The DB building unit 33 may access an external server to collect data whose copyright has expired, and store the collected data in the database 40 .

Here, data whose copyright has expired is a work whose copyright protection period has expired, which is a work that has passed a certain period after the author's death, and anyone can freely use it without a separate license or approval procedure.

The DB construction unit 33 may collect data whose copyright has expired by accessing a web site provided so that the copyrighted material can be used. For example, the website may be a shared yard (http://gongu.dopyright.or.kr/), and provides not only expired works, but also shared works such as privately held works and public contents with high social preservation value, and in addition, Expired works abroad can also be found.

The DB construction unit 33 may extract sentences from the collected data and store the sentences in the database 40 .

The sentence display unit 35 may extract a sentence similar to the user input sentence collected by the input unit 31 from the database 40 and transmit it to the user terminal 10 .

For example, the sentence display unit 35 may build a neural network that learns sentences stored in the database 40 and outputs a vector value indicating context information of each sentence. Here, the sentence display unit 35 may extract context information of the sentence for the learning data by using the Word2Vec algorithm. In this case, the Word2Vec algorithm may include a Neural Network Language Model (NNLM). The neural network language model is basically a Neural Network consisting of an Input Layer, a Projection Layer, a Hidden Layer, and an Output Layer. The neural network language model is what is used to vectorize words. Since the neural network language model is a known technology, a detailed description thereof will be omitted. The Word2vec algorithm, for text mining, is an algorithm that determines the proximity by looking at the front and back relationships between each word. The Word2vec algorithm is an unsupervised learning algorithm. The Word2vec algorithm may be a metric that expresses the meaning of a word in a vector form, as the name indicates. The Word2vec algorithm can represent each word as a vector in a space of about 200 dimensions. Using the Word2vec algorithm, a vector corresponding to a word can be obtained for each word. The Word2vec algorithm may enable a dramatic improvement in precision in the field of natural language processing compared to other conventional algorithms. Word2vec can learn the meaning of words by using the relationship between words in the sentences of the input corpus and adjacent words. The Word2vec algorithm is based on an artificial neural network, and it starts from the premise that words with the same context have close meanings. The Word2vec algorithm learns through text documents, and trains the artificial neural network to learn other words that appear nearby (about 5 to 10 words before and after) for one word as related words. Since words with related meanings are more likely to appear close to each other in the document, two words can have vectors that are closer to each other in the process of repeating learning. The learning method of the Word2vec algorithm is divided into a CBOW (Continuous Bag Of Words) method and a skip-gram method. The CBOW method predicts the target word using the context created by the surrounding words. The skip-gram method predicts words that can come around based on one word. The skip-gram method is known to be more accurate in large datasets. Therefore, in the embodiment of the present invention, the Word2vec algorithm using the skip-gram method is used. For example, if learning is successfully completed through the Word2vec algorithm, similar words may be located nearby in a high-dimensional space. According to the Word2vec algorithm as described above, the closer the distribution of the surrounding words in the learning document is, the more similar the calculated vector values may be, and the similarly the calculated vector values may be regarded as similar. Since the Word2vec algorithm is a well-known technique, a more detailed description related to vector value calculation will be omitted.

The sentence display unit 35 extracts the sentences stored in the database 40 as learning data, and learns the learning data through the Word2vec algorithm to construct a neural network that extracts context information about the content of the event.

The sentence display unit 35 may input the user input sentence obtained through the input unit 31 into the neural network to calculate an input sentence vector value representing context information on the user input sentence.

The sentence display unit 35 may calculate a similarity between example sentence vector values indicating context information of each sentence stored in the database 40 and an input sentence vector value. In this case, the sentence display unit 35 may use at least one of a Euclidean distance, a cosine similarity, and a Tanimoto coeffieient as a similarity calculation method.

The sentence display unit 35 may extract a sentence having an example sentence vector value having the highest similarity to the input sentence vector value among the sentences stored in the database 40 as an example sentence and transmit it to the user terminal 10 .

Here, the application executed in the user terminal 10 may have an interface in which an input area for receiving a sentence from the user and an output area for outputting an example sentence received from the sentence display unit 35 are arranged side by side. Accordingly, the user can check the example sentence while inputting the sentence in the input area.

As such, the sentence display unit 35 may extract a copyright expired sentence having a context most similar to the sentence input through the user terminal 10 using an artificial intelligence neural network and provide it to the user.

3 is a flowchart of a sentence writing method according to an embodiment of the present invention.

The sentence writing method according to an embodiment of the present invention may be performed under substantially the same configuration as that of the sentence writing apparatus 30 illustrated in FIGS. 1 and 2 . Accordingly, the same components as those shown in FIGS. 1 and 2 are given the same reference numerals, and repeated descriptions are omitted.

Referring to FIG. 3 , the method for writing a sentence according to an embodiment of the present invention includes the steps of collecting data whose copyright protection period has expired (S10), storing the collected data in the database 40 (S20), and a user It may include the step of receiving a sentence from the input (S30), the step of extracting a sentence similar to the sentence input from the user in the database 40 (S40), and the step of outputting the extracted sentence to the user terminal 10 (S50). there is.

In the step (S10) of the DB construction unit 33 collecting data whose copyright protection period has expired, the DB construction unit 33 accesses a website provided so that the copyright has expired data and the copyright has expired. may be a step in collecting

The step of storing the data collected by the DB building unit 33 in the database 40 ( S20 ) may be a step of extracting sentences from the collected data and storing the sentences in the database 40 .

The step ( S30 ) of the input unit 31 receiving a sentence from the user may be a step of the input unit 31 receiving the sentence through an application executed in the user terminal 10 .

The step (S40) of the sentence display unit 35 extracting a sentence similar to the sentence input from the user from the database 40 is performed using an artificial intelligence neural network in which the sentence display unit 35 extracts context information for the input sentence. It may be a step of extracting an example sentence having context information similar to the sentence input by the user.

The step (S50) of the sentence display unit 35 outputting the extracted sentence to the user terminal 10 may be a step of the sentence display unit 35 outputting an example sentence extracted through an application executed in the user terminal 10. there is.

Meanwhile, the DB construction unit 33 may identify similar sentences from among the collected sentences, extract representative sentences representing the identified similar sentences, and store the extracted sentences in the database 40 . For example, among the collected sentences, sentences composed of a combination of different words may exist even if the context is the same. The DB construction unit 33 identifies similar sentences in order to reduce the load by preventing these similar sentences from being overlapped and stored in the database 40, and extracts representative sentences that can be expressed by integrating the identified similar sentences. .

The DB construction unit 33 may include a similar sentence processing system 200 for identifying similar sentences and extracting representative sentences. In this regard, an example will be described with reference to FIG. 4 or less.

4 is a diagram showing an example of a similar sentence processing system according to an embodiment of the present invention.

Referring to FIG. 4 , the system 200 accesses a document database 202 , a document clustering system 204 , a text segment selection system 206 , a word/phrase sorting system 210 , and an identification system input text 211 . ), and generation system input text 212 .

FIG. 5 is a flowchart illustrating the operation of the system shown in FIG. 4 .

As an example, the document database 202 includes copyright expired data as sentences collected by the DB building unit 33 . It also includes, by way of example, various other news articles written by various other news agencies. As an example, each article includes a timestamp that briefly indicates when the article was written. Also, as an example, a plurality of articles from different news agencies may be written for a variety of different events.

Hereinafter, a news article will be described as an example.

The document clustering system 204 accesses the document database 202 as shown by block 214 in FIG. 5 . Although a single database 202 is shown in FIG. 4 , multiple databases may be accessed instead.

The clustering system 204 identifies articles in the document database 202 created for the same event.

In one embodiment, articles are also identified as being written at approximately the same time (within a predetermined temporal range of each other, eg, within a month, a week, a day, several hours, etc.). Articles identified as being written about the same event (or at the same time) form a document cluster 218 . This is illustrated by block 216 of FIG. 5 .

Once articles from relevant sources are identified by cluster 218, text segments (sentences, phrases, headlines, paragraphs, etc.) within these articles are extracted, as desired. For example, the journalistic convention of news articles recommends that the first two sentences of an article represent a summary of the rest of the article. Thus, in accordance with one embodiment of the present invention, articles (eg, written by different news agencies) are clustered into clusters 218 and provided to a text segment selection system 206 , within each cluster 218 . The first two sentences of each article are extracted. Although the description of the present invention has been directed to sentences, this is exemplary only and other text segments may be readily used. The sentences in each cluster 218 of articles are output as a set of sentences 222 corresponding to the clustered articles. The sentence sets 222 are output by the text segment selection system 206 to the word/phrase sorting system 210 . This is indicated by block 220 in FIG. 5 .

In the specific example where sentences are used, many sentences collected in this way appear as versions of several single-source sentences, which have been slightly rewritten by editors of different news agencies for stylistic reasons. Often, these sets of sentences appear to differ only slightly in method, such as the order of clauses appearing in the sentence.

The text segment selection system 206 generates sets 222 of sentences in each cluster. Note that word/phrase alignment system 210 can operate on large sets of sentences by extracting mappings between words or phrases based on a holistic examination of sentences in the set. However, this discussion only describes the creation of sentence pairs and the performance of sorting on those pairs, as an example. Thus, in one embodiment, the identified sets of sentences are formed into pairs of sentences. Accordingly, the text segment selection system 206 performs pairing of each sentence in the set with respect to each other sentence (hereinafter, pairing). In one embodiment, pairs of sentences are selectively filtered, and in another embodiment, they are output directly to the word/phrase sorting system 210 . Although filtering for this embodiment will be described, it is noted that the steps involved in filtering are optional.

In one embodiment, the text segment selection system 206 implements a heuristic that filters sentence pairs based on words of shared content. For example, in one embodiment, system 206 filters sentence pairs to remove sentence pairs that do not share at least three words, each of at least four letters. Of course, filtering is optional, but if used, the filtering algorithm implemented can vary widely. Filtering on past results (this requires a feedback loop to feed back word/phrase sorting system 210 to text segment selection system 206), filtering on different numbers of content words, semantic and syntactic Any of a variety of other filtering techniques may be used, such as filtering for undesirable information. In some cases, sets of sentences may be paired, filtered, and provided to word/phrase sorting system 210 .

In one embodiment, word/phrase sorting system 210 implements a conventional word/phrase sorting algorithm, derived from statistical machine translation literature, to learn lexical correspondences between sentences in sets 222 . . For example, suppose the following two sentences are input to the machine translation system 210 as sentence pairs: "Storms and tornadoes killed at least 14 people as they ripped through the central US States of Kansas and Missouri" "A swarm of tornadoes crashed through the Midwest, killing at least 19 people in Kansas and Missouri".

These sentences, with some differences, may have a common editorial source. In any case, as an example, these sentences were written by two different news agencies at about the same time for the same incident. Within sentences, "ripped through" corresponds to "crashed through", "Central US states" corresponds to "Midwest", there is a difference in word order, and "killed" corresponds to "killing". Thus, there are morphological differences between the words, including differences in the number of reported victims.

6 is a diagram for explaining alignment between pairs of two sentences according to an embodiment of the present invention.

6 shows the correspondence between words in sentences and multiple word phrases after words and phrases are sorted according to the conventional sorting system 210 . The statistical sorting algorithm established connections between different but parallel pieces of information, for most correspondences, as indicated by the lines connecting the words. For example, the noun phrases "Storms and tornadoes" and "a swarm of tornadoes" are not directly compared. Thus, as more data is required, the link between "storms" and "swarm" and the link between "storms" and "a" will disappear. A difference in the order of words may be revealed by a crossing pattern of the connection relationship between two sentences.

In one embodiment, "P.F. Using techniques known in Brown et al., "The Mathematics of Statistical Machinie Translation: Parameter Estimation", Computational Linguistics, 19:263-312, (June 1993), a word/phrase alignment system 210 is implemented. . Of course, other machine translation or word/phrase alignment techniques may be used to identify relationships between words and input text. Developing a sorting model using sorting system 210 on sentence sets and performing statistical word and/or phrase sorting is represented by block 230 of FIG. 5 .

The word/phrase sorting system 210 then outputs the sorted word and phrases 232 according to the sorting models 234 that have been generated based on the input data. Basically, in the sorting system cited above, models are trained to identify word correspondences. In the alignment technique, as shown in FIG. 6 , first, a word alignment is found between words in text segments. The system then assigns a probability to each alignment and optimizes the probability based on subsequent training data to produce a more accurate model. Outputting the aligned words and phrases 232 with the alignment models 234 is shown in block 236 of FIG. 5 .

As an example, the alignment model 234 may calculate a translation probability assigned to a word alignment, a movement probability representing the probability that a word or phrase will move within a sentence, and a probability that a single word may correspond to two different words within different text segments. Including the conventional translation model parameters, such as the diversity (fertility) probability to represent.

Blocks 237 , 238 , and 239 represent optional processing steps used to bootstrap the system with respect to training itself. These are described in more detail below with respect to FIG. 7 .

In embodiments where bootstrapping is not used, system 211 receives the output of system 210 to identify words, phrases, or sentences that are paraphrased from one another. The identified similar sentences 213 are output by the system 211 . This is indicated by block 242 of FIG. 5 .

Also, aligned phrases and models may be provided as generation system input text 212 . As an example, system 212 is a conventional decoder that receives as input words and/or phrases that produce pseudosentence 238 for that input.

Accordingly, the system 212 may be used to generate pseudosentences of the input text using the aligned words and phrases 232 and the alignment models 234 generated by the alignment system 210 . Generating similar sentences of the input text based on the aligned words and phrases and the alignment models is illustrated by block 240 of FIG. 5 . As a generation system according to an embodiment, "Y. Wang and A. Waibel, "Deoding Algorithm in Statistical Machine Translation", Proceedings of 35th Annual Meeting of the Association of Computational Linguistics (1997) are known.

7 is a view showing another example of a similar sentence processing system according to an embodiment of the present invention.

Figure 7 is similar to Figure 4, except that the identification system 211 is also used for bootstrap training. This is illustrated in more detail in blocks 237-239 of FIG. 5 . For example, with respect to FIGS. 7 and 5 , assume that the word/phrase sorting system 210 has the output sorting model 234 and the sorted words and phrases 232 as described above. However, now the entire text of each document cluster 218 is fed to an identification system 211 that identifies a supplemental sentence set 300 for use in further training the system (also, sentences are used as examples, etc.) text segments of may be used). Together with the alignment model 234 and the aligned words and phrases 232 , the identification system 211 can process the text of the clustered documents 218 to reselect the sentence sets 300 from the respective clusters. there is. This is illustrated by block 237 . The reselected sentence set 300 is then provided to the word/phrase sorting system 210 , and based on the reselected sentence sets 300 , alignment models 234 and sorted words and phrases 232 . ), and their associated probability matrices are generated or recalculated. The performance of word and phrase alignment and generation of aligned words and phrases and alignment models for reselected sentence sets is illustrated by blocks 239 and 239 of FIG. 5 .

Hereinafter, the recalculated alignment model 234 and the newly sorted words and phrases 232 are input back into the identification system 211, which in turn processes the text in the document clusters 218 to identify a new set of sentences. can be used by system 211 . In addition, the new sentence sets are fed back to the word/phrase sorting system 210 , and the process may proceed to further refine the training of the system.

For pseudo-sentences processed using the present invention, various applications exist. For example, potential applications of the pseudo-sentence processing system include question-and-answer systems, more general information processing systems, as known in the prior art. Such a system may generate similar sentence scores to determine the degree of similarity between two text segments in returning a set of documents based on a query. Likewise, such a system can perform query expansion (generating a single original query of multiple forms) using its pseudo-sentence generation capabilities to find better matches or improve calls.

Another application to the recognition and generation of similar sentences includes a summary of a plurality of documents. Utilizing the similar sentence recognition technique, in order to generate a summary, the automatic document summary system finds similar sentences in other documents and determines the most prominent information in the document set.

Another application for similar sentence recognition and generation is a dialog system. Such a system can generate a response that echoes the input, but is expressed differently, preventing repeating the exact same input. This allows the dialog system to express a more natural and interactive sound.

In addition, similar sentence recognition and generation techniques may be used in a word processing system. The word processing system can be used to automatically generate the rewritten style and suggest the rewritten style to the user. For example, this is effective when a user repeats a phrase many times even in one paragraph while writing a document. Likewise, a word processing system may include a feature to flag repeated (but otherwise expressed) information spread within a document. Similarly, such a word processing system may include a feature of rewriting a piece of prose as a similar sentence.

As such, the DB construction unit 33 is a method of training a similar sentence processing system for a plurality of copyright expired materials, the method comprising: accessing a plurality of documents; identifying, from the plurality of documents, a cluster of related texts written by different authors on a common subject; receiving the cluster of related texts; selecting a set of text segments from the cluster, wherein the selecting comprises grouping required ones of the related texts into a set of related text segments; and using textual alignment to identify syntactic relationships between texts within text segments included within the set of related text segments, wherein the text alignment achieves statistical textual alignment. aligning words within text segments within the set of related text segments using the

The DB construction unit 33 may further include calculating an alignment model based on the identified similar sentence relationship, and may generate representative sentences representing similar sentences using the alignment model.

Meanwhile, the DB construction unit 33 may predict events with respect to a plurality of copyright expired data, match the predicted event keywords, and store them together in the database 40 . In this case, the sentence display unit 35 may receive an input of a mapping keyword preferred by the user in advance from the user, and may extract an example sentence to be presented to the user from among the sentences to which the corresponding mapping keyword is assigned.

The DB construction unit 33 may include a mapping prediction system 300 for prediction of mapping keywords for each collected sentence. In this regard, an example will be described below with reference to FIG. 8 .

8 is a block diagram illustrating a mapping prediction system according to an embodiment of the present invention.

Referring to FIG. 8 , the mapping prediction system of the present embodiment has, as its functional configuration, a learning data input unit 310 , a word extraction unit 311 , a vector calculation unit 312 , an index value calculation unit 313 , and a classification A model generation unit 314 , a prediction data input unit 320 , and an event prediction unit 321 are provided and configured. The vector calculating unit 312 includes a sentence vector calculating unit 312a and a word vector calculating unit 312b as a more specific functional configuration. Moreover, the mapping prediction system of this embodiment is equipped with the classification model storage part 330 as a storage medium.

In addition, for the convenience of the following description, the part composed of the word extraction unit 311 , the vector calculation unit 312 , and the index value calculation unit 313 is called the similarity index value calculation unit 300 . The similarity index value calculating unit 300 inputs sentence data about a sentence, calculates and outputs a similarity indicator value reflecting the relationship between the sentence and the word included therein. In addition, the event prediction system of the present embodiment predicts a specific event from the text content (predicting which of a plurality of events it corresponds to) using the similarity index value calculated by the similarity index value calculating unit 300 . ) will be. In addition, the predictive model generating apparatus of the present invention is configured by the learning data input unit 310 , the similarity index value calculating unit 300 , and the classification model generating unit 314 .

Each of the functional blocks 310 to 314 and 320 to 321 can be configured by hardware, a digital signal processor (DSP), or software. For example, when configured by software, each of the functional blocks 310 to 314 and 320 to 321 is actually configured with a computer CPU, RAM, ROM, etc., and includes RAM, ROM, hard disk, or semiconductor. This is realized by operating a program stored in a recording medium such as a memory.

The learning data input unit 310 inputs, as learning data, sentence data related to m sentences (m is an arbitrary integer greater than or equal to 2) for which one of the plurality of maps is known. Here, the number of the plurality of events may be two or three or more. For example, it is possible to set it as two events indicating the presence or absence of occurrence possibility related to one matter, such as the possibility of occurrence of a specific disorder or symptom. Alternatively, it is also possible to set it as a combination of two or more ideas different from each other, such as a person's personality type, hobbies, and the like. In addition, the thought exemplified here is only an example, and is not limited thereto.

The input sentence data is preferably data in which sentences related to a plurality of events to be predicted are described. For example, when data for learning is input for the purpose of constructing a predictive model for predicting the presence or absence of the possibility of occurrence of a system failure, sentence data for a report describing the results of monitoring or inspection of the system is input. am.

However, in the case where the purpose is to predict a person's personality type, hobbies, etc., the relationship between the sentence and the thought can be determined by the interpretation described below, even if the sentence is thought to have no relation at all to the plural ideas to be predicted. It is possible that this could be found. Therefore, it is not essential to use only sentences determined by a human to be related to a plurality of events to be predicted as learning data. That is, depending on the content of the plurality of events to be predicted, data for learning includes not only data in which sentences clearly related to the plurality of events are described, but also data in which sentences that are thought to be unrelated to the plurality of events are described. It may be entered as

In addition, the sentence input by the data input unit 310 for learning, that is, the sentence to be analyzed to be described later, may consist of one sense (a unit divided by a sphere), or may have a plurality of senses. may be done. A sentence composed of a plurality of senses may be a part or all of a sentence included in one document. When a part of sentences included in one document are used as learning data, the learning data input unit 310 inputs sentence data in a state in which it is set which part of the sentences in the document to be used as learning data (strictly, speaking, input the document data, and use the setting part in the document as the sentence data). For example, among documents in which a plurality of description items exist, it is conceivable to set a sentence relating to a specific description item to be used as learning data. The number of description items to be set may be one, or a plurality may be sufficient as them.

The word extraction unit 311 interprets the m sentences input by the learning data input unit 310 and extracts n words (n is an arbitrary integer greater than or equal to 2) from the m sentences. As an analysis method of a sentence, it is possible to use a well-known morpheme analysis, for example. Here, the word extraction unit 311 may extract all morphemes of the part-of-speech divided by the morpheme analysis as words, or may extract only the morphemes of a specific part-of-speech as words.

In addition, a plurality of the same word may be included in m sentences. In this case, the word extraction unit 311 does not extract a plurality of the same word, but only one. That is, the n words extracted by the word extraction unit 311 mean n types of words. Here, the word extracting unit 311 measures the frequency at which the same word is extracted from m sentences, and n words (n types) from the one with the largest appearance frequency, or n words ( n types of words) may be extracted.

The vector calculating unit 312 calculates m sentence vectors and n word vectors from the m sentences and n words. Here, the sentence vector calculating unit 312a vectorizes the m sentences to be analyzed by the word extracting unit 311 in q dimensions according to a predetermined rule, so that q sentences (q is an arbitrary integer of 2 or more) ) computes m sentence vectors composed of the axis components. In addition, the word vector calculating unit 312b calculates n word vectors composed of q axis components by vectorizing the n words extracted by the word extracting unit 311 in q dimensions according to a predetermined rule, respectively. .

In this embodiment, as an example, sentence vectors and word vectors are calculated as follows. Now, consider a set S=<d∈D, w∈W> consisting of m sentences and n words. Here, for each sentence di(i=1, 2, …, m) and each word wj (j=1, 2, …, n), respectively, the sentence vector di and the word vector wj (hereinafter, the symbol “→” indicates that it is a vector). Then, with respect to the arbitrary word wj and the arbitrary sentence di, the probability P(wj|di) shown in the following formula (1) is calculated.

And, this probability P(wj|di) is, for example, in the article ""Distributed Representations of Sentences and Documents" by Quoc Le and Tomas Mikolov, Google Inc, Proceedings of the 31st International Conference on Machine Learning Held in Bejing, China on 22-24 June 2014” is a value that can be calculated by imitating the probability p. In this paper, for example, when there are three words "the", "cat", and "sat", "on" is predicted as the fourth word, and the calculation formula for the prediction probability p is published. . The probability p(wt|wt-k, ..., wt+k) described in the paper is a plurality of words wtk, ... , is a positive probability when another word wt is predicted from wt+k.

On the other hand, the probability P(wj|di) expressed in Equation (1) used in the present embodiment indicates a positive probability that one word wj out of n words is expected from one sentence di out of m sentences. are doing Predicting one word wj from one sentence di means predicting the possibility that the word wj will be included in a certain sentence di, specifically, when a certain sentence di appears.

In Equation (1), an exponential function value is used with e as the base and the dot product of the word vector w→ and the sentence vector d→ as the exponent. Then, an exponential function value calculated from the combination of the sentence di and the word wj to be predicted, and n exponential function values calculated from each combination of the sentence di and n words wk (k=1, 2, …, n) The ratio of the sum values of is calculated as the positive probability that one word wj is expected from one sentence di.

Here, the dot product of the word vector wj→ and the sentence vector di→ is a scalar value when the word vector wj→ is projected in the direction of the sentence vector di→, that is, the direction of the sentence vector di→ possessed by the word vector wj→ It can also be said to be the component value of This can be considered to indicate the degree to which the word wj contributes to the sentence di. Therefore, by using the exponential function value calculated using such a dot product, for the sum of the exponential function values calculated for n words wk (k=1, 2, ..., n), for one word wj Calculating the ratio of the calculated exponential function values corresponds to finding the positive probability that one word wj out of n words is expected from one sentence di.

In addition, since Equation (1) is symmetric with respect to di and wj, the probability P(di|wj) at which one sentence di out of m sentences is expected may be calculated from one word wj out of n words. Predicting one sentence di from one word wj means predicting the probability that a certain word wj will be included in the sentence di when it appears. In this case, the dot product of the sentence vector di→ and the word vector wj→ is a scalar value when the sentence vector di→ is projected in the direction of the word vector wj→, that is, the direction of the word vector wj→ possessed by the sentence vector di→ It can also be said to be the component value of This can be considered to indicate the degree to which the sentence di contributes to the word wj.

Incidentally, here, a calculation example using an exponential function value in which the dot product of the word vector w ? and the sentence vector d ? A calculation formula using the dot product of the word vector w→ and the sentence vector d→ may be sufficient, for example, the probability may be calculated|required by the ratio of the dot product itself.

Next, the vector calculating unit 312 maximizes the value L obtained by summing the probability P(wj|di) calculated by the above equation (1) for all sets S, as shown in the following equation (2). A sentence vector di→ and a word vector wj→ are calculated. That is, the sentence vector calculating unit 312a and the word vector calculating unit 312b calculate the probability P(wj|di) calculated by Equation (1) above for all combinations of m sentences and n words. Then, a sentence vector di→ and a word vector wj→ maximizing the target variable L are calculated using the sum of these values as the target variable L.

Maximizing the sum value L of the probability P(wj|di) calculated for all combinations of m sentences and n words means from a certain sentence di(i=1, 2, ..., m) to a certain word wj(j =1, 2, …, n) is to maximize the expected probability of being positive. That is, it can be said that the vector calculation unit 312 calculates the sentence vector di→ and the word vector wj→ such that this determination probability is maximized.

Here, in the present embodiment, as described above, the vector calculating unit 312 calculates m sentence vectors di ? composed of q axial components by vectorizing the m sentences di in q dimensions, and By vectorizing each of the n words in q dimensions, n word vectors wj→ composed of q axis components are calculated. This corresponds to calculating a sentence vector di→ and a word vector wj→ as maximized by the above-described target variable L by making q axial directions variable.

The index value calculation unit 313 takes the dot product of the m sentence vectors di→ and the n word vectors wj→ calculated by the vector calculation section 312, respectively, so that the relationship between the m sentences di and the n words wj m × n similarity index values reflecting gender are calculated. In the present embodiment, the index value calculation unit 313 has a sentence matrix D and a sentence matrix D having q axial components d11 to dmq of the m sentence vectors di→ , calculates an index value matrix DW having m × n similarity index values as each element by multiplying the word matrix W with each q axis components (w11 to wnq) of the n word vectors wj→ do. Here, Wt is the transpose matrix of the word matrix.

Each element of the index value matrix DW calculated in this way can be said to indicate which word contributes to which sentence to what extent. For example, the element dw12 in row 1 and column 2 is a value indicating how much the word w2 contributes to the sentence d1. Thereby, each row of the index value matrix DW can be used for evaluating the similarity of sentences, and each column can be used for evaluating the similarity of words.

The classification model generation unit 314 uses the m × n similarity index values calculated by the index value calculation unit 313 to generate n sentences for one sentence di (i = 1, 2, ..., m). A classification model for classifying the m sentences di into a plurality of mappings is generated based on the sentence indicator value group including the similarity indicator values dwj (j=1, 2, ..., n). For example, when generating a classification model classifying into three first to third events, the classification model generating unit 314 is a sentence index value group calculated based on a sentence in which the first event is already known. is classified as a “first event”, and a sentence index value group calculated based on a sentence in which the second event is known is classified as a “second event”, and the sentence corresponding to the third event is already known. For the sentence index value group calculated on the basis of , a classification model as classified as a "third event" is generated. Then, the classification model generation unit 314 stores the generated classification model in the classification model storage unit 330 .

Here, for the sentence index value group, for example, in the case of the first sentence d1, n similarity index values dw11 to dw1n included in the first line of the index value matrix DW correspond to this. Similarly, in the case of the second sentence d2, n similarity index values dw21 to dw2n included in the second row of the index value matrix DW correspond to this. Hereinafter, the same applies to sentence index value groups (n similarity index values dwm1 to dwmn) related to the m-th sentence dm.

The classification model generating unit 314, for example, calculates a feature amount for each sentence index value group of each sentence di, and according to the value of the calculated feature amount, a Markov chain Monte Carlo method A classification model for classifying each sentence di into a plurality of events is generated by optimizing the plural group separation by . Here, the classification model generated by the classification model generating unit 314 is a learning model that takes a sentence index value group as an input and outputs any one of a plurality of events to be predicted as a solution. Alternatively, a learning model that outputs a probability corresponding to each of a plurality of events to be predicted may be used as a probability. The form of the learning model is arbitrary.

For example, the form of the classification model generated by the classification model generation unit 314 is a regression model (a learning model based on linear regression, logistic regression, support vector machine, etc.), a tree model (decision tree, regression tree, Learning models based on random forests, gradient boosting trees, etc.), neural network models (perceptrons, convolutional neural networks, reconstructed neural networks, residual networks, RBF networks, stochastic neural networks, spiking) Learning models based on neural networks, complex neural networks, etc.), Bayesian models (learning models based on Bayesian inference, etc.), clustering models (k-root method, hierarchical clustering, non-hierarchical clustering, topic models, etc.) It is possible to do it in any one of the learning models based on the In addition, the classification model cited here is only an example, and is not limited to this.

The prediction data input unit 320 inputs sentence data related to one or more sentences to be predicted as prediction data. The sentence data input by the prediction data input unit 320 is sentence data regarding a sentence whose corresponding one of a plurality of maps is unknown. The sentence data input by the prediction data input unit 320 may be data in which sentences related to a plurality of events to be predicted are described, similar to the sentence data input by the data input unit 310 for learning, or a plurality of events to be predicted. It may be data in which a sentence which is thought to be unrelated at first glance is described.

The number of sentence data (number of sentences) m' input by the prediction data input unit 320 may not necessarily be the same as the number (m sentences) input by the learning data input unit 310 . The number of sentence data input by the data input unit 320 for prediction may be one or plural. However, a similarity index value is also calculated for a sentence input by the prediction data input unit 320 . Since the similarity index value indicates which word contributes to which sentence to what extent and to what extent which sentence contributes to which word, even for the sentence input by the prediction data input unit 320 . It is preferable to set it as plural.

The mapping prediction unit 321 includes the word extraction unit 311, the vector calculation unit 312 and the index value of the similarity index value calculation unit 300 for the prediction data input by the prediction data input unit 320 . Prediction by applying the similarity index value obtained by executing the processing of the calculation unit 313 to the classification model generated by the classification model generation unit 314 (the classification model stored in the classification model storage unit 330 ) Any one of a plurality of events is predicted from the target data.

For example, when m' sentence data are inputted as prediction data by the prediction data input unit 320, the mapping prediction unit 321 performs the similarity index value calculation unit 300 for the m' sentence data. By executing the processing of , m' sentence index value groups are obtained. The mapping prediction unit 321 applies the m' sentence index value groups calculated by the similarity index value calculation unit 300 one by one to the classification model as input data, so that a plurality of mappings for each of the m' sentences are given. predict where the

Here, it is preferable that the word extraction unit 311 extracts the same words as n words extracted from the m pieces of data for learning from the m' pieces of data for prediction. Since the sentence index value group consisting of n words extracted from the prediction data has the same word as the sentence index value group consisting of n words extracted from the learning data as elements, the classification stored in the classification model storage unit 330 This is because the fit of the model increases. However, it is not essential to extract the same n words as when learning, even when predicting. When a sentence index value group for prediction is generated by a combination of words that are different from those during learning, the fit to the classification model is lowered, but the low fit itself is also a factor in the evaluation, and the possibility that corresponds to the mapping is reduced. Prediction itself is possible.

9 and 10 are flowcharts illustrating an operation example of a mapping prediction system according to an embodiment of the present invention.

Referring to FIG. 9 , first, the learning data input unit 310 inputs, as learning data, sentence data related to m sentences for which a plurality of mappings are already known (step S1 ). The word extraction unit 311 interprets the m sentences input by the learning data input unit 310 and extracts n words from the m sentences (step S2).

Next, the vector calculating unit 312, from the m sentences input by the learning data input unit 310 and the n words extracted by the word extraction unit 311, m sentence vectors di→ and n word vectors wj→ is calculated (step S3). Then, the index value calculation unit 313 takes the dot product of the m sentence vectors di → and the n word vectors wj →, respectively, so as to reflect the relationship between the m sentences di and the n words wj, m × n similarities An index value matrix DW having an index value (m × n similarity index values as each element) is calculated (step S4).

In addition, the classification model generation unit 314 uses the m × n similarity index values calculated by the index value calculation unit 313 to form a sentence index value group consisting of n similarity index values dwj for one sentence di. Based on , a classification model for classifying the m sentences di into a plurality of events is generated, and the generated classification model is stored in the classification model storage unit 330 (step S5).

As a result, the operation at the time of learning is finished.

Referring to FIG. 10 , the prediction data input unit 320 inputs sentence data related to one or more sentences whose corresponding ones of the plurality of events are unknown (step S11). The mapping prediction unit 321 supplies the prediction data input by the prediction data input unit 320 to the similarity index value calculation unit 300, and instructs the calculation of the similarity index value.

According to this instruction, the word extraction unit 311 interprets the m' sentences input by the prediction data input unit 320, and n words (the same word as extracted from the learning data) from the m' sentences. ) is extracted (step S12).

In addition, it is not limited that all n words are included among the m' sentences. A null value is set for a word that does not exist among m' sentences.

Next, the vector calculating unit 312 generates m' sentence vectors di→ and n from the m' sentences input by the prediction data input unit 320 and the n words extracted by the word extracting unit 311 . n word vectors wj→ are calculated (step S13).

Then, the index value calculation unit 313 takes the dot product of the m' sentence vectors di → and the n word vectors wj →, respectively, so that the relationship between the m' sentences di and the n words wj is reflected in m'× n similarity index values (index value matrix DW having m'xn similarity index values as elements) are calculated (step S14). The index value calculation unit 313 supplies the calculated m'×n similarity index values to the mapping prediction unit 321 .

The mapping prediction unit 321 stores m' sentence index value groups in the classification model storage unit 330 based on the m'×n similarity index values supplied from the similarity index value calculation unit 300 , respectively. By applying it to the model, it is predicted for each of the m' sentences where it corresponds to a plurality of events (step S15). Thereby, the operation at the time of prediction ends.

11 is a block diagram of a mapping prediction system according to another embodiment of the present invention.

Referring to FIG. 11 , the event prediction system according to another embodiment of the present invention further includes a reward determining unit 322 in addition to the configuration shown in FIG. 8 . In addition, a classification model generating unit 314' is provided in place of the classification model generating unit 314 shown in FIG.

The reward determining unit 322 determines the reward given to the classification model generating unit 314' according to the actual mapping to the event predicted by the event prediction unit 321 . For example, the reward determining unit 322 determines to give a positive reward when the event predicted by the event predicting unit 321 matches the actual event, and when it does not match, no reward or Decide to give a negative payoff. Determination of whether the predicted event coincides with the actual event can be performed by various methods.

For example, when a user's taste preference is predicted as a plurality of events, information tailored to the predicted taste preference is presented to the user, and when the user takes an action on the information, the predicted event is an actual event. It is possible to judge that it is consistent with As a specific example, advertisement information of a product or service tailored to the predicted hobby or taste is displayed on the web page the user is browsing, and the user clicks the advertisement information to view the detailed information, or the product or service posted in the advertisement information. When the user causes an action such as purchasing a service, it is determined that the predicted event coincides with the actual event.

In addition, when predicting the possibility of occurrence of a specific failure with respect to a certain system, based on the historical data in which the monitoring history of the system is recorded, whether or not a specific failure has actually occurred is monitored, and the fact that the predicted failure has actually occurred is historical data When detected from , it is possible to determine that the predicted event coincides with the actual event. Similarly, when predicting the possibility of occurrence of a specific symptom for multiple users, based on historical data such as the user's medical examination history, it is monitored whether or not the specific symptom actually occurs, and it is determined that the predicted symptom actually occurs. In the case of detection from the historical data, it is possible to determine that the predicted event coincides with the actual event.

The classification model generation unit 314' generates a classification model based on the training data input by the training data input unit 310, similarly to the classification model generation unit 314 shown in FIG. 8, and a classification model storage unit. (330) is memorized. In addition to this, the classification model generation unit 314' changes the classification model stored in the classification model storage unit 330 in accordance with the reward determined by the reward determination unit 322 .

In this way, the DB construction unit 33 interprets m sentences (m is an arbitrary integer greater than or equal to 2) and extracts n words (n is an arbitrary integer greater than or equal to 2) from the m sentences. wealth; a sentence vector calculation unit for calculating m sentence vectors comprising q axis components by vectorizing the m sentences in q dimensions (q is an arbitrary integer greater than or equal to 2) according to a predetermined rule; a word vector calculating unit for calculating n word vectors comprising q axis components by vectorizing the n words in q dimensions according to a predetermined rule; an index value calculating unit calculating m × n similarity index values reflecting the relationship between the m sentences and the n words by taking the dot product of the m sentence vectors and the n word vectors, respectively; Classifying the m sentences into a plurality of maps based on a sentence indicator value group including n similarity indicator values for one sentence using the m×n similarity indicator values calculated by the indicator value calculating unit. a classification model generation unit generating a classification model for a prediction data input unit for inputting one or more sentences as prediction data as prediction data; and a similarity index value obtained by executing the processing of the word extraction unit, the sentence vector calculation unit, the word vector calculation unit, and the index value calculation unit on the prediction data input by the prediction data input unit, the classification and an event prediction unit that predicts any one of the plurality of events from the prediction target data by applying to the classification model generated by the model generation unit.

Meanwhile, the sentence writing system 1000 according to another embodiment of the present invention may further include a voice output device 50 in addition to the components shown in FIG. 1 .

The voice output device 50 may be applied as a speaker equipped with a wireless communication function, and is connected to the user terminal 10 in a wireless communication method to provide an example sentence received from the sentence display unit 35 in the user terminal 10 by voice. can be output as

12 is a diagram illustrating an audio output device according to an embodiment of the present invention.

Referring to FIG. 12 , the audio output device 50 according to an embodiment of the present invention is attached to the front part of the main body and exposed to the outside, and speakers having a plurality of apertures are formed in the vertical direction, and sound is generated from each speaker. It may include a speaker unit 51 through which a signal is output.

On the other hand, the voice output device 50 can be configured by modularizing a Wi-Fi module, a Bluetooth module, etc. that support wireless communication inside the main body, and reproduces a sound source provided from the user terminal 10 in a Wi-Fi or Bluetooth environment. Allow users to listen. In this case, the user terminal 10 may convert the sentence received from the sentence display unit 35 into voice data, and transmit the voice data to the voice output device 50 so that the user can listen.

Meanwhile, the sentence writing system 1000 according to another embodiment of the present invention may further include an auxiliary output device in addition to the components shown in FIG. 1 .

The auxiliary output device may be applied as a display device equipped with a wireless communication function, and is connected to the user terminal 10 in a wireless communication method to output an example sentence received from the sentence display unit 35 in the user terminal 10 there is.

Such an auxiliary output device may be installed through an installation module on a table on which the user terminal 10 is installed. In this regard, it will be described with reference to FIG. 13 .

13 is a view showing an installation module according to an embodiment of the present invention.

Referring to FIG. 13 , the installation module 60 according to an embodiment of the present invention may include an installation bar 62 , a seating part 63 , and an installation bracket 61 .

The installation bar 62 may be provided in the form of a column of a rectangular parallelepiped, and a seating portion 63 may be provided at one end.

The seating part 63 is formed in the form of a plate having a predetermined thickness, and a seating groove may be formed in the lower portion for seating the auxiliary output device. That is, the lower portion of the auxiliary output device may be inserted and installed in the seating groove.

The installation bracket 61 may be provided at the other end of the installation bar 62 so that the rotation of the installation bar 62 is possible.

The installation bracket 61 may be composed of a bottom surface and a pair of wall surfaces formed upward from both ends of the bottom surface, and in this case, the pair of wall surfaces may be formed in a semicircular shape with an upper end portion. The other end of the installation bar 62 may be rotatably installed between a pair of wall surfaces of the installation bracket 61 .

For example, the installation bar 62 may form a long hole 611 at the other end thereof, and the installation bracket 61 through a rotation pin 612 penetrating the long hole 611 from the outer circumferential surface of the installation bracket 61 . can be installed on At this time, the installation bar 62 may be moved in the front-rear direction with respect to the installation bracket 61 by the long hole 611 .

The installation bracket 61 may have a bottom surface attached to and fixed to the table on which the user terminal 10 is installed.

A plurality of insertion grooves 611b concave to the inside may be formed in a plurality of edges of a pair of wall surfaces of the installation bracket 61 . Here, an insertion protrusion 62a provided to be inserted into the insertion groove 611b may be formed on the outer peripheral surface of the other end of the installation bar 62 . In this case, the insertion groove 611b may be provided in a spiral shape. That is, since the insertion groove 611b is provided in a spiral shape having a locking groove 611a formed therein, it is possible to prevent the insertion protrusion 62a inserted therein from being separated.

Briefly describing the rotation method of the installation bar 62 as described above, the user moves the installation bar 62 forward and then rotates it to a desired angle, and in that state, the insertion groove 611b by moving the installation bar 62 backward. ) can be fixed by inserting the insertion protrusion (62a). This installation module 60 has the advantage of being easy to operate by a user.

On the other hand, the seating portion 63 may be entirely made of silver or copper, or may have a configuration in which an antibacterial material of any one or more of silver, titanium dioxide, and copper is integrally coated on the upper side of the metal plate.

In addition, the antibacterial material may be integrally coated on the upper side of the metal plate using a method such as ion deposition, plating, or spraying.

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

1000: sentence writing system
10: user terminal
20: network
30: sentence writing device
40: database

Claims (20)

user terminal; and
Constructing a database storing data whose copyright has expired, obtaining an input sentence that is a sentence input by the user from the user terminal, extracting a sentence similar to the input sentence from the database as an example sentence, and outputting it from the user terminal a sentence writing device; including;
The sentence writing device,
By learning the sentences stored in the database according to the Word2Vec algorithm, a neural network that extracts a vector value representing context information for an input sentence is constructed, and an input sentence obtained from the user terminal is input to the neural network to provide a context for the input sentence. Calculating an input sentence vector value representing information, calculating a similarity between example sentence vector values representing context information of each of the sentences stored in the database and the input sentence vector value, and calculating the input sentence among sentences stored in the database Extracting a sentence having an example sentence vector value with the highest degree of similarity to the vector value as the example sentence and transmitting it to the user terminal,
an auxiliary output device connected to the user terminal through a wireless communication method to receive and output the example sentence received from the sentence writing device in the user terminal; and
Further comprising; an installation module for installing the auxiliary output device on the table on which the user terminal is installed,
The installation module is
Installation bar provided in the form of a rectangular parallelepiped;
a seating part formed in the form of a plate of a predetermined thickness at one end of the installation bar, and forming a seating groove in the lower part of the auxiliary output device to be inserted and installed; and
Including; and an installation bracket installed at the other end of the installation bar so that the rotation of the installation bar is possible;
The installation bracket is
bottom surface;
a pair of wall surfaces formed upwardly from both ends of the bottom surface and having an upper end formed in a semicircular shape; and
A plurality of insertion grooves that are recessed inwardly on the edge of each of the pair of wall surfaces and formed in a spiral shape with a locking groove formed on the inside;
The installation bar,
A long hole is formed at the other end, and is rotatably installed between the pair of wall surfaces from the outer surface of the pair of wall surfaces through a rotation pin penetrating the long hole, and is inserted into the insertion groove at the other end. Forming an insertion protrusion provided,
The seating part,
A sentence writing system consisting of an integral coating of any one or more antibacterial substances of silver, titanium dioxide, and copper on the upper side of the metal plate.

delete According to claim 1,
The user terminal is
A sentence writing system that is loaded and executed with a sentence writing application operated by the sentence writing device.
According to claim 1,
The user terminal is
A sentence writing system, wherein a sentence writing application having an interface in which an input region for receiving a sentence from a user and an output region for outputting the example sentence are arranged side by side is mounted and executed.
According to claim 1,
The sentence writing device,
Including; an input unit for receiving the input sentence from the user terminal;
According to claim 1,
The sentence writing device,
A system for writing sentences, including; a DB construction unit for accessing a web site provided so that the copyright has expired material can be used and collecting the copyright expired material.
According to claim 1,
The sentence writing device,
A DB construction unit for collecting the copyrighted materials, extracting sentences, and storing the extracted sentences in the database; including, a sentence writing system.
According to claim 1,
The sentence writing device,
A DB construction unit for collecting the copyright-expired materials and identifying a similar sentence relationship between the copyright-expired materials; including, a text writing system.
According to claim 1,
The sentence writing device,
Including; DB construction unit to collect the copyright expired data,
The DB construction unit,
Including a; event prediction system for identifying the event of each of the copyright-expired materials;
The event prediction system is
a word extraction unit that interprets m sentences (m is an arbitrary integer greater than or equal to 2) and extracts n words (n is an arbitrary integer greater than or equal to 2) from the m sentences;
a sentence vector calculation unit for calculating m sentence vectors comprising q axis components by vectorizing the m sentences in q dimensions (q is an arbitrary integer greater than or equal to 2) according to a predetermined rule;
a word vector calculating unit for calculating n word vectors comprising q axis components by vectorizing the n words in q dimensions according to a predetermined rule;
an index value calculation unit calculating m × n similarity index values reflecting the relationship between the m sentences and the n words by taking dot products of the m sentence vectors and the n word vectors, respectively;
Classifying the m sentences into a plurality of maps based on a sentence indicator value group including n similarity indicator values for one sentence using the m × n similarity indicator values calculated by the indicator value calculating unit. a classification model generation unit generating a classification model for
a prediction data input unit for inputting one or more sentences as prediction data as prediction data; and
A similarity index value obtained by executing the processing of the word extraction unit, the sentence vector calculation unit, the word vector calculation unit, and the index value calculation unit on the prediction data input by the prediction data input unit is set to the classification model and an event prediction unit that predicts any one of the plurality of events from the prediction data by applying to the classification model generated by the generation unit.
According to claim 1,
The sentence writing device,
A sentence display unit for extracting a sentence similar to the input sentence from the database as an example sentence; further comprising,
The sentence display part,
and extracting the example sentence by calculating a Euclidean distance between a vector value representing the sentences stored in the database and a vector value representing the input sentence.
According to claim 1,
The sentence writing device,
A sentence display unit for extracting a sentence similar to the input sentence from the database as an example sentence; further comprising,
The sentence display part,
and extracting the example sentence by calculating a cosine similarity between a vector value representing the sentences stored in the database and a vector value representing the input sentence.
According to claim 1,
The sentence writing device,
A sentence display unit for extracting a sentence similar to the input sentence from the database as an example sentence; further comprising,
The sentence display part,
and extracting the example sentences by calculating a Tanimoto coeffieient between a vector value representing the sentences stored in the database and a vector value representing the input sentence.
According to claim 1,
and a network provided as a wired communication network connecting the user terminal and the sentence writing device.
According to claim 1,
The system further comprising; a network provided as a mobile communication network connecting the user terminal and the sentence writing device.
According to claim 1,
The text writing system further comprising a; a network provided as a Wibro (Wireless Broadband) network for connecting the user terminal and the sentence writing device.
According to claim 1,
The system further comprising; a network provided with a High Speed Downlink Packet Access (HSDPA) network connecting the user terminal and the sentence writing device.
According to claim 1,
The system further comprising; a network provided as a satellite communication network connecting the user terminal and the sentence writing device.
According to claim 1,
A network provided as a Wi-Fi (Wireless Fidelity) network connecting the user terminal and the sentence writing device; further comprising, a sentence writing system.
According to claim 1,
and a voice output device connected to the user terminal in a wireless communication manner, and receiving and outputting a voice file for the example sentence from the user terminal.
In the sentence writing method in the sentence writing device connected through the user terminal and the network,
collecting data whose copyright protection period has expired;
storing the collected data in a database;
receiving a sentence from a user through the user terminal;
extracting a sentence similar to a sentence input from a user from the database using a neural network that outputs a vector value representing context information of the input sentence; and
Including; outputting the extracted sentence to the user terminal;
The sentence writing device,
By learning the sentences stored in the database according to the Word2Vec algorithm, a neural network that extracts a vector value representing context information for an input sentence is constructed, and an input sentence obtained from the user terminal is input to the neural network to provide a context for the input sentence. Calculating an input sentence vector value representing information, calculating a similarity between example sentence vector values representing context information of each of the sentences stored in the database and the input sentence vector value, and calculating the input sentence among sentences stored in the database Extracting a sentence having an example sentence vector value with the highest degree of similarity to the vector value as the example sentence and transmitting it to the user terminal,
an auxiliary output device connected to the user terminal through a wireless communication method to receive and output the example sentence received from the sentence writing device in the user terminal; and
Further comprising; an installation module for installing the auxiliary output device on the table on which the user terminal is installed,
The installation module is
Installation bar provided in the form of a rectangular parallelepiped;
a seating part formed in the form of a plate of a predetermined thickness at one end of the installation bar, and forming a seating groove in the lower part of the auxiliary output device to be inserted and installed; and
Including; and an installation bracket installed at the other end of the installation bar so that the rotation of the installation bar is possible;
The installation bracket is
bottom surface;
a pair of wall surfaces formed upwardly from both ends of the bottom surface and having an upper end formed in a semicircular shape; and
A plurality of insertion grooves that are recessed inwardly on the edge of each of the pair of wall surfaces and formed in a spiral shape with a locking groove formed on the inside;
The installation bar,
A long hole is formed at the other end, and is rotatably installed between the pair of wall surfaces from the outer surface of the pair of wall surfaces through a rotation pin penetrating the long hole, and is inserted into the insertion groove at the other end. Forming an insertion protrusion provided,
The seating part,
A method of writing a sentence consisting of an integral coating of any one or more antibacterial substances of silver, titanium dioxide, and copper on the upper side of the metal plate.

Images