KR102315181B1 - Method, apparauts and system for named entity linking and computer program thereof - Google Patents

Abstract

The present invention relates to a method, apparatus, system, and computer program for linking entity names, and more particularly, to a method for linking entity names with respect to a synonym included in a text. a word-vector conversion step of converting a plurality of words including a neural network input step of inputting the plurality of vectors into a deep neural network having a hidden layer; and an entity name connection step of performing entity name connection to the polynomial using the output value of the deep neural network.

Description

NAMED ENTITY LINKING AND COMPUTER PROGRAM THEREOF}

The present invention relates to a method, apparatus, system and computer program for linking entity names, and more particularly, to a method, apparatus, system and computer program for linking entity names to polyphosic words located in text.

Named entity linking means that when a word, etc. located in a given text can be interpreted in multiple ways, it is determined in what sense the word, etc. is used in consideration of usage in the given text, and Accordingly, it refers to the operation of connecting the words to a person, an object, a place, etc. (eg, determining whether the word "spider" used in the text means a singer or an arthropod). The entity name connection can be an important component in a search engine, a conversation system, etc., in which the result of determining the meaning of a polymorphic word affects the quality.

Conventionally, after calculating the features, such as surrounding words of the target word of the given text, mainly, support vector machine (SVM), logistic regression, similarity analysis, etc. are applied to name the entity A connection was made. However, in the case of the prior art, as the number of object names to be distinguished increases, the time required to connect the object names may increase, and furthermore, it may be difficult to obtain an accurate result within a given time. In particular, the object name connection is not limited to a small number of specific words, and in a search engine that can use a large number of multi-lingual words, the object name to be distinguished may increase from 100 units to 10,000 units or more, so the accuracy of entity name connection is difficult with the prior art. A big problem will follow.

Republic of Korea Patent Registration No. 10-1255957 (2013.4.24.)

The present invention was devised to solve the problems of the prior art as described above, and it is possible to perform exact entity name linking with respect to a synonym located in a text, and further efficiently use computer resources to perform entity name linking. An object of the present invention is to provide a method, device, system and computer program for linking object names that can be used.

A method for linking entity names according to an aspect of the present invention for solving the above problems is a method for linking entity names with respect to polyphosic words included in text. a word-vector conversion step of converting the words of a neural network input step of inputting the plurality of vectors into a deep neural network having a hidden layer; and an entity name connection step of performing entity name connection to the polynomial using the output value of the deep neural network.

A computer program according to another aspect of the present invention is characterized in that it is a computer program stored in a computer-readable medium for executing each step of the method for linking an entity name described above in a computer.

An object name linking system according to another aspect of the present invention is an entity name linking system for performing entity name linking with respect to polyphosic words included in text. a word-vector conversion unit that converts a vector into a vector; a neural network input unit for inputting the plurality of vectors into a deep neural network having a hidden layer; and an entity name connection unit for performing entity name connection to the polynomial using the output value of the deep neural network.

According to an embodiment of the present invention, an entity name linking method, apparatus, which can perform accurate entity name linking with respect to polyphosic words located in text, and further efficiently use computational resources to perform entity name linking, system and computer programs can be provided.

BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings, which are included as part of the detailed description to help the understanding of the present invention, provide embodiments of the present invention, and together with the detailed description, explain the technical spirit of the present invention.
1 is a view for explaining the configuration of an entity name connection system according to an embodiment of the present invention.
2 is a diagram for explaining the concept of entity name connection according to an embodiment of the present invention.
3 and 4 are exemplary views of a screen of a user terminal in the entity name connection system according to an embodiment of the present invention.
5 is an exemplary diagram of a web document to which an entity name is linked according to an embodiment of the present invention.
6 is a flowchart of a method for linking entity names according to an embodiment of the present invention.
7 is a diagram for explaining the operation of the deep neural network in the entity name connection system according to an embodiment of the present invention.
8 is a diagram for explaining the operation of an N-gram convolution filter in the entity name connection system according to an embodiment of the present invention.
9 is a diagram for explaining a dynamic data sampling technique in the entity name connection system according to an embodiment of the present invention.
10 is a diagram for explaining a static policy alignment loss function technique in the entity name association system according to an embodiment of the present invention.
11 is a diagram for explaining a negative sampling technique for a general entity in the entity name connection system according to an embodiment of the present invention.
12 is a diagram for explaining the operation of the entity name connection system according to an embodiment of the present invention.
13 is a diagram for explaining the operation of an automatic learning corpus building module according to an embodiment of the present invention.
14 is a diagram for explaining the configuration of each function of the entity name connection system according to an embodiment of the present invention.
15 is a diagram for explaining a distance-based word-vector model and post-processing of the entity name connection system according to an embodiment of the present invention.
16 is a block diagram of an entity name connection system according to an embodiment of the present invention.

The present invention can apply various transformations and can have various embodiments. Hereinafter, specific embodiments will be described in detail based on the accompanying drawings.

In describing the present invention, if it is determined that a detailed description of a related known technology may obscure the gist of the present invention, the detailed description thereof will be omitted.

Terms such as first, second, etc. may be used to describe various elements, but the elements are not limited by the terms, and the terms are used only for the purpose of distinguishing one element from other elements. used

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, exemplary embodiments of a method, apparatus, system and computer program for linking entity names according to the present invention will be described in detail with reference to the accompanying drawings.

First, FIG. 1 shows a diagram for explaining the operation of the entity name connection system 100 according to an embodiment of the present invention. As can be seen in FIG. 1 , the entity name connection system 100 according to an embodiment of the present invention is connected to a database 30 and the like, and documents (eg, blogs, With respect to various electronic documents including cafes, news, etc.), entity name associations are performed for polymorphic words located in texts including one or more sentences included in the document.

As a more specific example, when a specific electronic document stored in the database 30 includes the sentence "I went to the store with my daughter today and picked 'One Piece'" as can be seen in FIG. 2, the ' The word 'One Piece' may be 'One Piece', which is the title of a cartoon, or 'One Piece', a type of clothing, depending on its usage. On the other hand, in the entity name connection system 100 according to an embodiment of the present invention, it is determined whether the 'dress' in the sentence means a cartoon or clothes, and the The object name connection is performed. Accordingly, it is possible to connect an entity name such as 'one piece_clothing' to 'one piece' of the sentence and further store it.

In addition, in the entity name connection system 100 according to an embodiment of the present invention, a service can be provided to users by reflecting the entity name connection result. As a more specific example, as can be seen in FIG. 3 , when the user of the terminal 10a wants to search for the cartoon 'One Piece' (for example, polymorphic words according to the search word format supported by each search service) In the case of inputting a specific search word with a specific meaning (in the case of inputting in a form such as one-piece (comic) or one-piece (clothing) (310 in FIG. 3 )), the entity name connection system 100 according to an embodiment of the present invention ), by providing a search result in consideration of the meaning of a search term having multiple meanings ( 320 of FIG. 3 ), it is possible to help the user more conveniently obtain a desired search result.

Also, referring to FIG. 1 , referring to the entity name linking system 100 and related components according to an embodiment of the present invention in more detail, the entity name linking system 100 according to an embodiment of the present invention includes one or more It may be configured to include a server, and may be implemented in various ways, such as configured using dedicated hardware for performing entity name connection, or configured by interworking with one or more terminal devices.

In addition, as the terminals 10a and 10b, various portable terminals such as a smart phone, a tablet PC, a PDA, and a mobile phone may be used, and in addition, various terminals such as a personal computer (PC) and a notebook PC may be employed.

In addition, the communication network 30 connecting the terminals 10a and 10b and the entity name connection system 100 may include a wired network and a wireless network, and specifically, a local area network (LAN), an urban area It may include various communication networks such as a metropolitan area network (MAN) and a wide area network (WAN). In addition, the communication network 30 may include a well-known World Wide Web (WWW). However, the communication network 30 according to the present invention is not limited to the networks listed above, and may include various networks such as a well-known wireless data network, a well-known telephone network, or a well-known wired/wireless television network.

In addition, if the operation of the entity name connection system 100 according to an embodiment of the present invention will be described with a more specific example, if the text contains a synonym for "Jimin", "Jimin" included in the text Depending on its usage, the word may be “Jimin” of “Group AOA”, “Jimin” of “Group BTS”, or other “Jimin”. Also, even when the text contains a polymorphic word "one piece", the "one piece" may be 'one piece', which is the title of a cartoon, or 'one piece', which is a type of clothing, depending on its usage. In contrast, in the entity name connection system 100 according to an embodiment of the present invention, what kind of person "Jimin" used in a given text means, and whether "one piece" of the given text means a cartoon or clothing By judging, the entity name connection is performed with respect to the synonyms used in the text.

Accordingly, in the entity name connection system 100 according to an embodiment of the present invention, entity names such as “Jimin_singer AOA” and “One Piece_Manga” are added to the synonyms of “Jimin” and “One Piece” in the given text It can be connected and stored, and further, various services such as a search service can be provided to users by reflecting the object name connection result.

Furthermore, the entity name connection system 100 according to an embodiment of the present invention may provide a list of polyphrases so that the user can distinguishly input polyphrases, and provide an interface for the user to select. For example, as can be seen in FIG. 4 , when the user inputs the polyphoric "Jimin" or "One Piece", the user selects the It can be done ( 410 and 420 in FIG. 4 ).

As another example, when a user wants to perform a search for the cartoon 'One Piece', by allowing the user to specify and input a search word with a specific meaning among polyphosic words according to the search word format supported by each search service (eg, One Piece) (input in a format such as (comic) or one piece (clothing)), the entity name connection system 100 according to an embodiment of the present invention provides search results in consideration of the meaning of a search word having multiple meanings so that the user can more conveniently We can help you get the search results you want.

Accordingly, in the entity name connection system 100 according to an embodiment of the present invention, as can be seen in FIG. 5 , even if the same word ‘one piece’ is used in the text, what meaning is used in the text By grasping (eg, as a one-piece (clothing) in FIG. 5(a) and as a one-piece (comic) in FIG. 5(b), it is possible to more accurately provide information desired by the user.

In addition, in the present invention, a case in which one word such as "Jimin" and "One Piece" has a multiple meaning is described as an example, but the present invention is not necessarily limited thereto. The invention can be applied (eg, "Winter Sea (song title/Sea of a Specific Season)"). Accordingly, in the present invention, the expression “word” may include a phrase including two or more words.

6 is a flowchart of a method for linking entity names according to an embodiment of the present invention. As can be seen in FIG. 6 , in the method for concatenating entity names according to an embodiment of the present invention, in the method for linking entity names with respect to a synonym included in text, the entity name connection system 100 is A word-vector conversion step (S110) of transforming a plurality of words including a polymorphic word into a plurality of vectors corresponding to each other, a neural network input step (S120) of inputting the plurality of vectors into a deep neural network having a hidden layer (S120), and the deep neural network It may include an entity name connection step (S130) of performing entity name connection to the polynomial using the output value of .

Hereinafter, with reference to FIG. 6 , the method for linking entity names according to an embodiment of the present invention will be reviewed in detail by dividing each step.

First, in step S110, the object connection system 100 converts a plurality of words including a polymorphic word in the text into a plurality of vectors corresponding to each other. In this case, the object connection system 100 according to an embodiment of the present invention may select a plurality of words within a predetermined distance based on a polyphrase in the text, and then calculate each vector for the plurality of words. .

For a more specific example, in FIG. 7, the article about the drama <Chief Kim>, "In Manager Kim, Jeong Hae-seong (played by Ga-eun Hong) is lovely by matching a lace blouse and a tweed-colored dress" (710 in FIG. 7) In the case of linking the object name for the polymorphic word "one piece", after classifying the text into morpheme units, a plurality of words within a predetermined distance based on the "one piece" (for example, "blouse" which is four words before and after After calculating ", "and", "tweed", "color matching" and "a", "match", "sun", "lovely") ( 720 in FIG. 7 ), It can be converted to a vector. For example, as shown at 731 of FIG. 7 , a 128-dimensional vector (0.0,0.0, 0.1, 0.3, 0.8, .. .), a 128-dimensional vector for "and" (0.3, 0.2, 0.5, 0.7, 1.2, ...), etc., can be converted and input to the embedding layer. In this case, the dimension of the transformed vector may be determined in consideration of the accuracy of linking the entity name to the polynomial, the time required for linking the entity name, computing resources, and the like.

Next, in step S120, a plurality of vectors are input to a deep neural network having a hidden layer (733 in FIG. 7). In this case, the transformed vectors (M) are grouped into a plurality of units, and after a predetermined operation is performed to calculate N vectors (732 in FIG. 7 ) (here, M>N), they may be input to the deep neural network.

That is, the vector transformed in the present invention and input to the embedding layer is transferred to the hidden layer and processed (732 in FIG. 7). At this time, the input M vectors are convolved using an N-gram convolution filter 732, etc. By mixing through calculation and reducing the number to N, it is possible to significantly improve the learning and decoding speed and efficiency of the deep neural network.

That is, as can be seen in FIG. 8 , when the N-gram convolution filter 732 is used, in contrast to the case where the N-gram convolution filter 732 is not used ((a) of FIG. 8) ((b) of FIG. 8), the input dimension can be greatly reduced by half, such as when word 1 and word 2 are mixed, tag target polyphoric word and word 4 are mixed, etc. Its efficiency can be greatly improved.

As a more specific example, if 12 words are generated from the text, centered on the polymorphic word (multiple word + each 12 words before and after = 25 words), and each word is converted into a 128-dimensional vector and input, the deep neural network is 128 x 25 = 3200 are input and processed, whereas as shown in FIG. 8 , a plurality of words (M) is reduced to N words through convolution operation, etc., so the time required for learning and decoding and computation It is possible to effectively reduce the computing resource.

Finally, in step S130, an entity name connection is performed for a polymorphic word using the output value of the deep neural network.

In this case, the deep neural network may be a neural network that is learned by using a plurality of text groups in which polymorphic words are used with different meanings, and thus learned to distinguish and output the meanings of polyphosic words used in the text.

In addition, the information passing through the hidden layer can output the object name connection result through the output layer through a neuron function such as a Rectified Linear Unit (ReLU).

Accordingly, as shown in 734 of FIG. 7 , the output layer of the deep neural network may be distinguished for each polymorphic word used in a different meaning to form a structure constituting a plurality of nodes. For example, the output layer of the deep neural network divides the polymorphic word "one piece" according to its meaning and includes separate nodes corresponding to the case of being used as "one piece (clothing)" and the case of being used as "one piece (manga)". can be configured. Accordingly, when the text "Lovely by matching a lace blouse and a tweed color matching dress" of 710 of FIG. 7 is input, the output value (0.8) of the "One Piece (Clothing)" node is Since the output value (0.2) of "One Piece (Comic)" is higher than the output value (0.2), it is determined that the synonym "One Piece" used in the text is used in the meaning of "One Piece (Clothes)", and the object name is linked to the Synonym "One Piece (Clothes)" will be able to perform

Accordingly, the entity name connection system 100 according to an embodiment of the present invention 1) converts the words around the tagging target into distributed energy vector values and inputs them into the embedding layer, 2) at this time the input The vector value can be mixed and compressed through convolutional operation, etc. to reduce the amount of computation required for processing, 3) the compressed vector is multiplied by a hidden layer and non-linear activation is performed with the ReLU function 4) Then, it is multiplied by an output layer having a plurality of nodes that are distinguished for each polymorphic word with different meanings, and finally it is determined from the value of the output layer what meaning the tagging target polynomial represents.

Furthermore, additional techniques for optimizing the performance of the entity name connection system 100 may be further applied to the present invention.

First, in the object name connection system 100 according to an embodiment of the present invention, when the object name connection is performed using a neural network according to the prior art, it is biased toward a polynomial corresponding to a plurality of learning data due to the non-uniformity of the distribution of the learning data. This can improve the problem of linking object names.

That is, as can be seen in the table 910 of FIG. 9 , when the number of learning data for each meaning of a polymorphic word is significantly different (one piece_(clothing) 51,099 cases >> one piece_(comic) 5,313 cases, one piece_( (music group) 1,672 cases) and a small number of learning data (in the above case, one piece_(comic), one piece_(music group))) In the above case, there may be a problem that the individual names are connected by being biased towards One Piece_(Clothing)).

In contrast, in the entity name connection system 100 according to an embodiment of the present invention, as can be seen in Equation 920 of FIG. 9 , in sampling learning data for a specific polymorphic word, the accuracy of entity name connection, etc. By dynamically correcting to increase the sampling distribution for polyphosic words that show poor performance, and alleviating the non-uniformity of the distribution of learning data by meaning of polyphos, it is biased toward multiple objects due to the difference in frequency between each object in the learning data. problem can be effectively rectified.

In addition, in the entity name connection system 100 according to an embodiment of the present invention, as can be seen in FIG. 10 , by comparing only the connection candidate entities of the synonymous to the tagging target to calculate the ranking (Ranking), It can also effectively improve the problem of poor learning quality due to a large number of output object classes.

More specifically, as can be seen from the output layer 1010 of FIG. 10 , the output layer of the entity name connection system 100 according to an embodiment of the present invention includes a plurality of nodes corresponding to different meanings of each polynomial. (For example, One Piece (Comic), One Piece (Clothing), Jimin (AOA), Jimin (BTS), Butterfly (Singer), Butterfly (Animal), etc.), the learning process for the polymorphic word ‘One Piece’ By comparing only candidate objects (one piece (cartoon), one piece (clothing)) that can be connected to the polymorphic ‘one piece’ such as one piece (comic) and one piece (clothing) of the output layer in the output layer to judge the result, a large number of irrelevant output objects (Jimin (AOA), Jimin (BTS), butterfly (singer), butterfly (animal), etc.) can effectively suppress the deterioration of quality.

More specifically, as can be seen in Equation 1020 of FIG. 10, the f1_prior function, which is a non-differential function (convex function, etc.) that can reflect the F1 measure, is applied to predict the By adjusting the loss propagation to decrease or increase according to success or failure, the accuracy of entity name connection can be effectively improved. Accordingly, it was possible to achieve an improvement effect of 1% or more based on Macro F1.

In addition, in the entity name connection system 100 according to an embodiment of the present invention, when entity name connection is performed using a neural network according to the prior art, it will be used as a common sense entity such as a general noun, a verb, or an adjective. It is possible to improve the problem that the accuracy may be lowered when linking the entity name to a possible polynomial.

In contrast, in the entity name connection system 100 according to an embodiment of the present invention, as shown in FIG. 11 , general entities such as general nouns, verbs, and adjectives are defined as classes representing entities that do not require entity name connection. Thus, by applying the negative sampling technique, it is possible to effectively suppress the occurrence of an entity name connection error for a general entity.

At this time, in the present invention, a general entity does not have to include all of the general nouns, verbs, adjectives, etc., but may be defined as including one or more of the general nouns, verbs, and adjectives. In the entity name connection system 100 according to an embodiment, a case of learning a deep neural network using learning data for a general entity corresponding to a general noun, a verb, or an adjective is broadly included.

That is, as can be seen in 1110 of FIG. 11, a common sense entity such as a common noun (“share” skills and experience) is incorrectly recognized as a proper noun (“share”), an entity name connection error. In order to prevent this, by applying negative sampling to a common sense entity, it is possible to effectively prevent an entity name connection error with respect to a common entity such as a common noun.

For example, in order to prevent association of object names with common nouns, verbs, and adjectives that do not want to be tagged, these objects are defined and learned as a separate class, and for words classified in this class, do not connect object names. method can be used. In particular, it is possible to effectively prevent a tagging error by learning to classify words in which connection errors occur frequently into separate classes.

Furthermore, in the conventional deep neural network (DNN) model, there may be a problem in that the processing speed is decreased while the amount of calculation such as multiplication operation is greatly increased to connect the entity name to a plurality of multi-lingual words, but the entity name according to an embodiment of the present invention The connection system 100 may reduce the amount of computation by compressing an input embedding layer using an N-gram convolution filter or the like, thereby improving the processing speed of the object name connection task.

12 illustrates the operation of the entity name connection system 100 according to an embodiment of the present invention. As can be seen in FIG. 12 , the entity name connection system 100 according to an embodiment of the present invention may include an entity name connection engine 1220, and the entity name connection engine 1220 is a deep neural network. It may be configured to include a (Deep Neural Network, DNN) model decoder and a word-vector model decoder. In the present invention, by using a deep neural network (DNN) model decoder and at the same time converting words into a vector form and using them, it is possible to perform accurate object name connection while using computational resources more efficiently.

Furthermore, the entity name connection engine 1220 according to an embodiment of the present invention provides a manual tagging issue to compensate for entity name connection errors that may occur in a deep neural network (DNN) model decoder or a word-vector model decoder. It may further include a response module, and accordingly, it is possible to quickly and effectively respond to errors that may appear in the learning process.

Accordingly, the entity name connection engine 1220 can receive the text 1210 including one or more polyphrases, and effectively perform ( 1250 ) linking the entity names to the one or more synonyms.

More specifically, in the case of performing object name connection using a support vector machine (SVM), logistic regression, similarity analysis technique, etc. according to the prior art, as the number of object names to be distinguished increases, There may be a problem of a decrease in accuracy versus time with an increase in the required time. However, in the entity name connection system 100 according to an embodiment of the present invention, a deep neural network (DNN) model decoder is used and a word is converted into a vector form at the same time. By performing object name concatenation by (word to vector), it is possible to effectively suppress an increase in the required time and a decrease in accuracy even if the number of object names to be distinguished increases.

More specifically, in the prior art, in order to assign an entity ID to each word, a technique such as the existence of a word that describes each entity well around the context and a weight value thereof have been used, but in this case, the When there is no word (Unseen word), it becomes difficult to judge the object, and to compensate for this, it is difficult to collect words related to the object indefinitely. In addition, when the entity name connection is performed using the sum of the weights of simple words, there is a limit in expressing the correlation or dependency between words, and thus the coverage or quality of the entity name connection may be deteriorated. .

In contrast, the present invention introduces a deep neural network (DNN) model, learns a large amount of news, blogs, etc. in a word-vector conversion (word2vec) method before learning the deep neural network (DNN) model, and provides energy for a number of words By obtaining the vector expression value, it is possible to remarkably reduce cases in which it is difficult to judge, such as a case in which a word is not in the model dictionary (Unseen word).

In addition, the entity name connection system 100 according to an embodiment of the present invention can automatically generate a learning corpus required for learning of a deep neural network (DNN) model decoder using the automatic learning corpus building module 1230, Accordingly, it is possible to effectively construct a deep neural network (DNN) model decoder with higher accuracy.

More specifically, FIG. 13 describes the operation of the learning corpus building module 1230 according to an embodiment of the present invention. As can be seen in FIG. 13 , the learning corpus building module 1230 parses meta data such as keywords corresponding to different meanings from a pre-prepared database such as Wikipedia ( 1310 ).

For example, as can be seen in FIG. 13, through meta data parsing of the polynomial "summer", "Singer" as a "soccer player" of "Sangju Sangmu Professional Football Team" and "Singer" as a member of "Group Cosmic Girls" It becomes possible to obtain metadata about "in "summer" and the like.

Next, the learning corpus building module 1230 generates a search expression for each entity through rule-based query generation, etc. using the generated metadata ( 1320 ).

For example, as the above rule, appropriate rules such as name and occupation, name and organization can be applied, and accordingly, for "soccer player" and "summer", "summer, Gwangju FC, soccer player, NOT ( You can create a search expression that includes search terms such as "singer", be able to

Accordingly, the learning corpus building module 1230 searches for documents such as the web, news, and blogs using each generated search expression, and each polymorphic word such as “summer” as “soccer player” and “summer” as “singer” It becomes possible to generate training data for (1330).

Through this, the learning corpus building module 1230 according to the present invention can effectively construct a deep neural network (DNN) model decoder with high accuracy by effectively generating a large amount of learning data for a plurality of multilingual words.

Furthermore, the configuration of the entity name connection engine 1410 according to an embodiment of the present invention is described in more detail in FIG. 14 .

14 , the entity name connection engine 1410 according to an embodiment of the present invention includes a tokenizer 1411, a deep neural network model 1412, and a distance-based word-vector. It may be configured to include a distance based word vector model 1413 , a post-processing unit 1414 , and a resource unit 1415 .

The tokenizer 1411 analyzes the given text and classifies it into morphemes and the like.

Next, the deep neural network model 1412 receives the classified morphemes, etc., and connects the entity names to the synonyms used in the text.

In addition, in the distance-based word-vector model 1413, in converting the first word in the text into a vector, a vector for the first word is taken into consideration in consideration of the second word included in the text and the distance between the second word and the second word. will yield

In addition, the post-processing unit 1414 performs post-processing on the results calculated from the deep neural network model 1412 and the distance-based word-vector model 1413 .

In the entity name connection system 100 according to an embodiment of the present invention, a distance-based word-vector model 1413 using an established database 1415, etc. for entities that are difficult to classify as a deep neural network (DNN) model 1412 It is also possible to perform object name association using .

More specifically, in FIG. 15 , operations of the distance based word vector model 1413 and the post-processing unit 1414 are described in more detail.

As can be seen in FIG. 15 , in the distance-based word-vector model 1413 according to an embodiment of the present invention, the relationship with the surrounding words and the distance within the sentence with respect to the object name connection target word in the given text are calculated. can be used for analysis. Specifically, in the drama 'Chief Kim', when an actor named 'Hye-seong Jeong' appeared and took on the role of 'Hong Ga-eun', the words 'Jung Hye-sung' and 'Hong Ga-eun' appear around when analyzing 'Chief Kim'. Using the dots, it can be seen that it is the drama 'Chief Kim', not 'Chief Kim', which indicates the position in the company.

In this case, the distance between each word may be used as a weight. Similarly, when analyzing ‘Jeong Hye-sung’, ‘Chief Kim’ and ‘Hong Ga-eun’ in the front can be considered and ‘Jung Hye-sung’ can be determined as the actress Jeong Hye-seong. Using this method, even if the word ‘Hong Ga-eun’ is not registered or learned, ‘Hong Ga-eun’ can be tagged as a character.

As a more specific example, as can be seen in FIG. 15 , when performing entity name concatenation for the synonym of “determinism” in a given text 1510, the distance-based word-vector model is constructed in the database 1520. Keywords corresponding to the drama “Chief Kim” (“Chief Kim”, “Jeong Hae-Sung”, “Hong Ga-Eun”, etc. in FIG. 15 ) can be calculated using the It can be determined that the "Chief Kim" is the drama "Chief Kim" rather than "Chief Kim" indicating the rank in the company by giving weights considering the distances such as "Hae-Sung Jeong" and "Ga-Eun Hong" based on ".

Furthermore, in performing the entity name connection to "one piece," which is another polymorphic word included in the text 1510, the entity name connection can also be performed using the deep neural network model 1530, so the object name connection target word The distance-based word-vector model 1413 and the deep neural network model ( 1413 ) and the deep neural network model ( 1530) can be selectively applied, and if necessary, it is also possible to further improve the accuracy by performing entity name connection in consideration of both the results of the distance-based word-vector model 1413 and the deep neural network model 1530. do.

Subsequently, the post-processing unit 1414 considers one of the result values of the distance-based word-vector model 1413 and the neural network model 1530, or calculates and compares both of them (1540). A name connection is performed (1550).

Accordingly, in the entity name connection system 100 according to an embodiment of the present invention, performance of hundreds of TPS per core can be achieved even on the basis of “CPU”, and tagging of tens of millions of entities is also possible. In this case, the F1 result value can also have a performance of 88% or more.

In addition, the computer program according to another aspect of the present invention is characterized in that it is a computer program stored in a computer-readable medium in order to execute each step of the above salpin entity name connection method in a computer. The computer program may be not only a computer program including a machine code generated by a compiler, but also a computer program including a high-level language code that can be executed in a computer using an interpreter or the like. In this case, the computer is not limited to a personal computer (PC) or a notebook computer, and includes a central processing unit (CPU) such as a server, smart phone, tablet PC, PDA, mobile phone, etc. to process any information that can execute a computer program. includes the device.

In addition, the computer-readable medium may continuously store a computer-executable program, or may be temporarily stored for execution or download. In addition, the medium may be various recording means or storage means in the form of a single or several hardware combined, it is not limited to a medium directly connected to any computer system, and may exist distributed on a network. Examples of the medium include a hard disk, a magnetic medium such as a floppy disk and a magnetic tape, an optical recording medium such as CD-ROM and DVD, a magneto-optical medium such as a floppy disk, and those configured to store program instructions, including ROM, RAM, flash memory, and the like. In addition, examples of other media may include recording media or storage media managed by an app store that distributes applications, sites that supply or distribute various other software, and servers.

Also, FIG. 16 exemplifies the configuration of the entity name connection system 100 according to an embodiment of the present invention.

As can be seen in FIG. 16 , the entity name connection system 100 according to an embodiment of the present invention may be configured to include a word-vector conversion unit 110 , a neural network input unit 120 , and an entity name connection unit 130 . can

Hereinafter, the entity name connection system 100 according to an embodiment of the present invention will be divided for each component. Further details of the entity name connection system 100 according to an embodiment of the present invention can be inferred from the description of the entity name connection method according to an embodiment of the present invention described above, and will be described in more detail below. is omitted.

First, the word-vector conversion unit 110 converts a plurality of words including polyphrases in text into a plurality of vectors corresponding thereto.

Next, the neural network input unit 120 inputs the plurality of vectors to a deep neural network having a hidden layer.

Finally, the entity name connection unit 130 performs entity name connection to the polynomial using the output value of the deep neural network.

In this case, the deep neural network may be a neural network learned by using a plurality of text groups in which the polymorphic words are used in different meanings.

Furthermore, in the entity name connection system 100 according to an embodiment of the present invention, M transformed vectors are grouped into a plurality of units and calculated to be N vectors (here, M>N) and processed.

The above description is merely illustrative of the technical spirit of the present invention, and various modifications and variations will be possible without departing from the essential characteristics of the present invention by those skilled in the art to which the present invention pertains. Accordingly, the embodiments described in the present invention are not intended to limit the technical spirit of the present invention, but to explain, and are not limited to these embodiments. The protection scope of the present invention should be construed by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

10a, 10b: terminal
20: communication network
30: database
100: Entity name connection system
110: word-vector conversion unit
120: neural network input unit
130: entity name connection part

Claims (7)

delete delete In the method of performing a search for a search term including a polymorphic word,
A search word inputted by a user in a search service according to a predetermined search word format, a search word that receives a search word including a polyphoric word and a polyphrase whose meaning is specified including an entity name and an identifier for distinguishing it input step; and
and a search performing step of specifying the meaning of the polymorphic word using the entity name, and performing a search for the search word in consideration of the polymorphic word having the specified meaning. 4. The method of claim 3,
In the search word input step,
The method of performing a search, characterized in that the search term includes the polymorphic word and a word specifying the meaning of the polymorphic word which is input in a predetermined format subsequent to the polymorphic word. A computer program stored in a computer-readable medium for executing each step according to any one of claims 3 to 4 in a computer. delete In the search system for performing a search for a search word including a polymorphic word,
A search word inputted by a user in a search service according to a predetermined search word format, a search word that receives a search word including a polyphoric word and a polyphrase whose meaning is specified including an entity name and an identifier for distinguishing it input unit; and
and a search performing unit for specifying the meaning of the polymorphic word by using the entity name, and performing a search for the search word in consideration of the polymorphic word with the specified meaning.

Images