KR102053419B1 - Method, apparauts and system for named entity linking and computer program thereof - Google Patents

Abstract

The present invention relates to a method for associating an entity name, a device, a system, and a computer program, and more particularly, to a method for performing entity name association for a multilingual word included in a text, wherein the entity name association device is used in the text. A word-vector conversion step of converting a plurality of words including each into a plurality of corresponding vectors; A neural network input step of inputting the plurality of vectors into a deep neural network having a hidden layer; And an entity name concatenation step of performing an entity name concatenation to the multilingual word using an output value of the deep neural network.

Description

Object name connection method, device, system and computer program {METHOD, APPARAUTS AND SYSTEM FOR NAMED ENTITY LINKING AND COMPUTER PROGRAM THEREOF}

TECHNICAL FIELD The present invention relates to an entity name association method, an apparatus, a system, and a computer program, and more particularly, to a method, an apparatus, a system, and a computer program for performing entity name association for a multilingual term located in text.

Named entity linking means that when a word or the like located in a given text can be interpreted intentionally, the meaning of the word or the like is determined in consideration of its use in the given text. Accordingly, the term refers to an operation of connecting the word to a person, an object, a place, etc. (for example, determining whether the word "spider" used in the text means a singer or an arthropod). Entity linkage can be an important component in search engines, dialogue systems, etc., where the results of semantic judgments on multiwords affect quality.

Conventionally, after calculating a feature of surrounding words of a connection target word of a given text as a feature, and applying a support vector machine (SVM), logistic regression, similarity analysis, etc. The connection was performed. However, in the conventional technology, as the number of distinguished object names increases, the time required for concatenating the object names may increase, and furthermore, it may be difficult to obtain accurate results within a given time. In particular, in search engines where the name association is not limited to a few specific words and a large number of multiple words can be used, the distinguished entity name may increase from 100 units to 10,000 units or more. The problem of falling greatly follows.

Republic of Korea Patent No. 10-1255957 (April 24, 2013)

The present invention was devised to solve the above-mentioned problems of the prior art, and it is possible to perform exact entity name concatenation with respect to a multilingual word located in the text, and furthermore, to efficiently perform entity name concatenation using computational resources. It is an object of the present invention to provide a method, apparatus, system, and computer program for connecting entity names.

In accordance with an aspect of the present invention for solving the above problems, a method for performing entity name association with respect to a multilingual word included in a text, wherein the entity name association system includes a plurality of entities including the polynomial in the text. A word-vector conversion step of converting the word of to each corresponding plurality of vectors; A neural network input step of inputting the plurality of vectors into a deep neural network having a hidden layer; And an entity name concatenation step of performing entity name concatenation to the multilingual word using an output value of the deep neural network.

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

An object name linking system according to another aspect of the present invention is an entity name linking system that performs entity name linking for a multilingual word included in text, and includes a plurality of words each corresponding to a plurality of words including the multilingual word in the text. A word-vector conversion unit for converting to 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 that performs entity name connection to the multilingual word using the output value of the deep neural network.

According to an embodiment of the present invention, it is possible to perform the exact entity name association with respect to the multilingual words located in the text, and furthermore, the entity name association method, apparatus, which can perform entity name association using more efficient computational resources, It is possible to provide a system and a computer program.

BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings, included as part of the detailed description in order to provide a thorough understanding of the present invention, provide examples of the present invention and together with the description, illustrate the technical idea of the present invention.
1 is a view for explaining the configuration of the entity name connection system according to an embodiment of the present invention.
2 is a view for explaining the concept of entity name connection according to an embodiment of the present invention.
3 and 4 are exemplary diagrams of the screen of the user terminal in the entity name connection system according to an embodiment of the present invention.
5 is an exemplary view of a web document to which the entity name is linked according to an embodiment of the present invention.
6 is a flowchart illustrating a method for linking entity names according to an embodiment of the present invention.
7 is a view 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 view for explaining the operation of the N-gram convolution filter in the entity name connection system according to an embodiment of the present invention.
9 is a diagram for describing a dynamic data sampling scheme in a system for linking entity names according to an embodiment of the present invention.
10 is a view for explaining a static policy matching loss function scheme in the entity name linking system according to an embodiment of the present invention.
FIG. 11 is a diagram for describing a negative sampling scheme for a generic entity in an entity name linking system according to an embodiment of the present invention.
12 is a view for explaining the operation of the entity name connection system according to an embodiment of the present invention.
13 is a view for explaining the operation of the learning corpus automatic building module according to an embodiment of the present invention.
14 is a view for explaining the configuration of each function of the entity name connection system according to an embodiment of the present invention.
FIG. 15 is a diagram for explaining a distance-based word-vector model and post-processing of an entity connection system according to an embodiment of the present invention. FIG.
16 is a block diagram of an entity name connection system according to an embodiment of the present invention.

As the inventive concept allows for various changes and numerous embodiments, particular embodiments will be described in detail with reference to the accompanying drawings.

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

Terms such as "first" and "second" may be used to describe various components, but the components are not limited by the terms, and the terms are only used to distinguish one component from another component. Used.

DETAILED DESCRIPTION Hereinafter, exemplary embodiments of a method for connecting an entity name, an apparatus, a system, and a computer program according to the present invention will be described in detail with reference to the accompanying drawings.

First, Figure 1 shows a view for explaining the operation of the entity name connection system 100 according to an embodiment of the present invention. As can be seen in Figure 1, the entity name connection system 100 according to an embodiment of the present invention is connected to the database 30 and the like, the document (for example, blog, For various electronic documents including cafes, news, etc.), the entity name linking is performed for a multi-word word located in a text including one or more sentences included in the document.

More specifically, for example, when a specific electronic document stored in the database 30 includes the sentence “I went to the store with my daughter today and chose one piece” as shown in FIG. The word 'dress' may be 'one piece', the title of a cartoon, or 'one piece', a type of clothing, depending on its use. On the other hand, in the individual name connection system 100 according to an embodiment of the present invention, it is determined whether the 'one piece' means cartoon or clothing in the sentence, and thus, the multi-word ('one piece') of the sentence is determined. The object name connection will be performed. Accordingly, the object name may be connected to the one-piece of the sentence such as one-piece_clothing and the like.

In addition, the entity name connection system 100 according to an embodiment of the present invention may provide services to users by reflecting the entity name connection result. For example, as shown in FIG. 3, when the user of the terminal 10a wants to perform a search for a cartoon 'one piece' (for example, it is a multi-word depending on a search word format supported by each search service). In the case where a specific search word having a specific meaning is inputted (in case of inputting in a form such as one piece (cartoon) or one piece (clothing) (310 in FIG. 3)), the entity name connection system 100 according to an embodiment of the present invention. ) Provides a search result in consideration of the meaning of the search word having a multi-purpose meaning (320 of FIG. 3), so that the user can more conveniently obtain a desired search result.

In addition, referring to FIG. 1, in more detail with respect to the entity name connection system 100 and related configurations according to an embodiment of the present invention, the entity name connection system 100 according to an embodiment of the present invention is one or more; The server may be configured to include a server. In addition, the server may be configured using dedicated hardware for performing entity name connection, or may be implemented in various ways such that one or more terminal devices are linked to each other.

In addition, as the terminals 10a and 10b, various portable terminals such as a smartphone, a tablet PC, a PDA, a cellular phone, etc. may be used, and various terminals such as a personal computer (PC) and a notebook PC may be adopted.

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

In addition, when the operation of the entity name connection system 100 according to an embodiment of the present invention will be described in more detail as an example, when the text includes the term “jimin,” “jimin” included in the text. The word may be "Jimin" in "Group AOA", "Jimin" in "Group BTS", or other "Jimin" depending on its use. In addition, even if the text includes the term "one piece", the "one piece" may be a "one piece" which is a title of a cartoon, or a one piece of clothing, depending on its use. On the other hand, in the entity name connection system 100 according to an embodiment of the present invention, whether the "Jimin" used in a given text means a person, whether the "one piece" of the given text means cartoon or clothing. By judging, the entity name concatenation is performed for the multiword used in the text.

Accordingly, in the entity name connection system 100 according to an embodiment of the present invention, an entity name such as "Jimin_Singer AOA", "One Piece_Cartoon", or the like is used in a synonym such as "Jimin" or "One Piece" of a given text. It can be connected and stored, and further it is possible to provide various services such as a search service to users by reflecting the object name connection result.

Furthermore, the entity name connection system 100 according to an exemplary embodiment of the present invention may provide a multilingual list so that a user can input the multilingual word, and provide an interface for the user to select. For example, as shown in FIG. 4, when a user inputs the word "Jimin" or "one piece", the user selects the word "Jimin" or "one piece" by dividing the meaning. (410, 420 of FIG. 4).

As another example, when a user wants to perform a search for a cartoon 'one piece', the user may input a specific search word of a specific word according to a search word format supported by each search service (for example, one piece (Manga) or one-piece (clothing) type), the entity name connection system 100 according to an embodiment of the present invention provides a search result in consideration of the meaning of the search word having a multipurpose meaning more convenient to the user You can help them get the search results they want.

Accordingly, in the entity name connection system 100 according to an embodiment of the present invention, as shown in FIG. 5, even if the same word is used, what meaning is used in the text. By, for example, in Fig. 5 (a) as one piece (clothing), Fig. 5 (b) as a piece (cartoon), it is possible to more accurately provide the 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 plural meaning is described as an example. However, the present invention is not necessarily limited thereto, and the present invention is not necessarily limited thereto. The invention can be applied (e.g., "winter sea (song title / sea of specific seasons)"). Accordingly, the expression "word" in the present invention may also include a phrase (phrase) including two or more words.

6 is a flowchart illustrating a method for linking individual names according to an embodiment of the present invention. As can be seen in Figure 6, the entity name linking method according to an embodiment of the present invention, in the method for performing the entity name link for the multi-word included in the text, the entity name link system 100 in the text A word-vector conversion step (S110) for converting a plurality of words including multiwords into a plurality of corresponding vectors, a neural network input step (S120) of inputting the plurality of vectors into a deep neural network having a hidden layer, and the deep neural network The entity name concatenation step (S130) may be performed by using the output value of the entity name concatenation for the multiword.

Hereinafter, referring to FIG. 6, the method for linking individual names according to an embodiment of the present invention will be examined in detail.

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

For example, in FIG. 7, the article about the drama <Kim Governor> is included in the text, "Hae Sung-Jung (Hong Ga-eun Station) in Kim Goverment is a lovely match with a lace blouse and a tweed color dress" (710 of FIG. 7). In the case of concatenating individual names for the multilingual "one piece", a plurality of words within a predetermined distance based on the "one piece" after classifying the text in morpheme units (for example, four words before and after "blouse" &Quot;, " and " " " tweed " " colour " " " " match " " solution " and " lovely " Can be converted to a vector. For example, as can be seen at 731 of FIG. 7, a 128-dimensional vector (0.0,0.0, 0.1, 0.3, 0.8,... For the word "blouse" using a predetermined word-vector conversion (word2vec) function. .), And 128-dimensional vectors (0.3, 0.2, 0.5, 0.7, 1.2, ...) for "and" can be converted into the embedding layer. In this case, the dimension of the vector to be transformed may be determined in consideration of the accuracy of the concatenation of the entity names, the time required for concatenating the entity names, and the computing resources.

Next, in operation S120, a plurality of vectors are input to the deep neural network having the hidden layer (733 in FIG. 7). In this case, the converted vectors (M) may be grouped into a plurality of units and calculated as N vectors (732 of FIG. 7) through a predetermined operation (where M> N), and may be input to the deep neural network.

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

That is, as can be seen in Figure 8, when using the N-gram convolution filter 732 as compared to the case of not using the N-gram convolution filter 732 (Fig. 8 (a)) (B) of FIG. 8, the input dimension can be greatly reduced in half, such as word 1 and word 2 are mixed, tagging multiword and word 4 are mixed, and the learning and decoding speed is increased by two. The efficiency can be greatly improved.

For example, if 12 words are calculated based on the polynomial (multiword + 12 words before and after each word = 25 words) in the text, and each word is converted into a 128-dimensional vector and the deep neural network is 128 x While 25 = 3200 are received and processed, as shown in FIG. 8, the plurality of words (M) are reduced to N words through convolution operations, such as time and computation required for learning and decoding. Effectively reducing the computing resources.

Finally, in step S130, using the output value of the deep neural network is connected to the entity name for the word.

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

In addition, the information passed through the hidden layer may be output via the output layer through the neuron function, such as ReLU (Rectified Linear Unit).

Accordingly, as shown in 734 of FIG. 7, the output layer of the deep neural network may form a structure in which a plurality of nodes are distinguished by each lexical word used in different meanings. For example, the output layer of the deep neural network may include separate nodes corresponding to the case where the word "one piece" is divided according to its meaning and used as "one piece" and "one piece (cartoon)". Can be configured. Accordingly, when the text “Lovely matched lace blouse and tweed color dress” of 710 of FIG. 7 is inputted, as shown in 734 of FIG. 7, the output value (0.8) of the “One piece (clothing)” node is decreased. The output of "one piece (cartoon)" is higher than the output value (0.2), and it is determined that the term "one piece" used in the text is used to mean "one piece (clothing)", and the entity name connection for the multilingual "one piece" is determined. Will be able to perform

Accordingly, the entity name connection system 100 according to an embodiment of the present invention, 1) converts the tagging surrounding words into distributed energy vector values and inputs them into the embedding layer, and 2) Vector values can be mixed and compressed using convolution operations to reduce the amount of computation required, and 3) the compressed vector is multiplied by a hidden layer and non-linear activation with the ReLU function. 4) Subsequently, it is multiplied by an output layer having a plurality of nodes distinguished by different words of different meanings, and finally, it is determined from the value of the output layer which meaning of the tagging target word is represented.

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 entity name connection system 100 according to an embodiment of the present invention, when performing entity name connection using a neural network according to the prior art, the multilingualism is biased to correspond to a plurality of learning data by nonuniformity of learning data distribution. This can improve the problem of concatenating entity names.

That is, as shown in the table 910 of FIG. 9, when the number of learning data for each meaning of a multiword becomes largely different (one piece _ (clothing) 51,099 cases >> one piece _ (cartoon) 5,313 cases, one piece _ ( Music group) 1,672 cases), a small number of learning data (in the above case, one piece _ (cartoon), one piece _ (music group), the accuracy of the case corresponding to the multiple words corresponding to the plurality of learning data ( In the above case, one-piece _ (clothing)) may be biased to connect the individual name.

On the other hand, in the entity name linkage system 100 according to an embodiment of the present invention, as shown in Equation 920 of FIG. Dynamically corrected to increase sampling distribution for multilingual words that show poor performance, thereby reducing the nonuniformity of the distribution of training data for each semantic word. Effectively ameliorate the problem.

In addition, in the entity name connection system 100 according to an embodiment of the present invention, as shown in FIG. 10, by comparing only the candidate candidates of the multilingual targets to be tagged, the rank is calculated (Ranking). Many output object classes can effectively improve the problem of poor learning quality.

More specifically, as can be seen in the output layer 1010 of Figure 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 multilingual (For example, dress (cartoon), dress (clothing), Jimin (AOA), Jimin (BTS), butterfly (singer), butterfly (animal), etc.) In the process, only candidate objects (one piece, one piece, and one piece) that can be connected to the same language, such as one piece (cartoon) and one piece (clothing) of the output layer, are compared with each other and judge the result. It is possible to effectively suppress the deterioration caused by output entities (JIA (AOA), JTS (BTS), butterflies (singers), butterflies (animals), etc.).

More specifically, as shown in equation 1020 of FIG. 10, the f1_prior function, which is a non-differential function (such as a convex function) that can reflect the F1 measure, is applied to the same entity prediction for the multilingual. By adjusting the loss propagation to decrease or increase according to success or failure, the accuracy of the entity name association can be effectively improved. Accordingly, the improvement effect of 1% or more of the Macro F1 standard could be achieved.

In addition, in the entity name connection system 100 according to an embodiment of the present invention, when performing entity name connection using a neural network according to the prior art, it is used as a common sense entity such as a common noun, a verb, an adjective, etc. It is possible to improve the problem that the accuracy may be reduced when concatenating entity names for multiwords.

In contrast, in the entity name linkage system 100 according to an embodiment of the present invention, as shown in FIG. 11, a general entity such as a normal noun, a verb, an adjective, and the like is defined as a class representing an entity for which entity name linkage is unnecessary Therefore, by applying the negative sampling technique, it is possible to effectively suppress the occurrence of an entity name connection error for a general entity.

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

That is, as shown in 1110 of FIG. 11, an entity name connection error that incorrectly recognizes a common sense entity such as a common noun (a "share" of technology and experience) as a proper noun (an actor "share"). In order to prevent this, by applying negative sampling on a common sense entity, it is possible to effectively prevent an object name connection error on a general entity such as a common noun.

For example, to avoid linking entity names for common nouns, verbs, adjectives, etc. that you do not want to tag, you define these classes as separate classes, and do not link them to words classified in this class. Can be used. In particular, tagging errors can be efficiently prevented by learning to classify words that frequently cause connection errors into separate classes.

Furthermore, in the conventional deep neural network (DNN) model, a problem may occur in which the processing speed decreases with a large increase in arithmetic operations such as multiplication operations for concatenating individual names for a plurality of multilingual terms, but according to an embodiment of the present invention. The concatenation system 100 may reduce the amount of computation by compressing an input embedding layer by using an N-gram convolution filter or the like, thereby improving the processing speed of the concatenation of entity names.

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

Furthermore, the entity name connection engine 1220 according to an embodiment of the present invention is 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, thereby enabling a quick and effective response to errors that may appear in the learning process.

Accordingly, the entity name connection engine 1220 may receive the text 1210 including one or more synonyms, and effectively perform 1250 the entity name association for the one or more synonyms.

More specifically, when performing object name concatenation using a support vector machine (SVM), logistic regression, or similarity analysis technique according to the related art, as the number of distinguished object names increases, Although the accuracy of time may decrease with increasing time, the entity name connection system 100 according to an embodiment of the present invention converts a word into a vector form while using a deep neural network (DNN) model decoder. By performing the object name concatenation by (word to vector), it is possible to effectively suppress the increase in the time required and the deterioration of accuracy even if the number of distinguished object names increases.

More specifically, conventionally, in order to give an individual ID to each word, a technique such as whether a word that describes each object well and a weight value thereof is used around the context is used. When there is no word (Unseen word), it is 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 linkage is performed by using the sum of the weights of simple words, there is a limit in expressing the correlation or dependency relationship between words, which may cause the coverage or quality of the entity name linkage to be deteriorated. .

In contrast, the present invention introduces a deep neural network (DNN) model, a large amount of news, blogs, etc. by learning a word-vector transformation (word2vec) before learning the deep neural network (DNN) model, energy for a number of words By obtaining vector representations, it is possible to significantly reduce cases that are difficult to judge, such as when words are not in the model dictionary (Unseen word).

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

More specifically, FIG. 13 illustrates the operation of the learning corpus building module 1230 according to an embodiment of the present invention. As shown in FIG. 13, the learning corpus building module 1230 parses meta data, such as keywords corresponding to a multiword used in different meanings, from a prepared database such as Wikipedia.

For example, as can be seen in FIG. 13, as a member of "Summer" and "Group Space Girl" who are "Soccer Players" of the "Resident Business Professional Football Team" through parsing metadata for the term "Summer". Meta data about "summer", etc. can be obtained.

Subsequently, the learning corpus building module 1230 generates a search expression for each entity through rule generation, etc. using the generated metadata, in operation 1320.

For example, as the rule, appropriate rules such as name and occupation, name and affiliated organization can be applied, and accordingly, "Soccer player" and "summer" can be applied to "summer, Gwangju FC, soccer player, NOT ( Singer) ", and the like, and search terms including" singer "and" summer "include search terms such as" summer, space girl, starship entertainment, and NOT. " It becomes possible.

Accordingly, the learning corpus building module 1230 searches for documents such as web, news, blogs, and the like by using the generated search expressions, and the respective words such as "summer", "soccer player", and "summer", "singer". It is possible to generate training data for (1330).

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

14 illustrates a configuration of the entity name connection engine 1410 according to an embodiment of the present invention in more detail.

As can be seen in Figure 14, the entity name connection engine 1410 according to an embodiment of the present invention is a tokenizer (14), a deep neural network model (DNN model) 1412, distance-based word-vector A model may include a distance based word vector model 1413, a post processor 1414, and a resource unit 1415.

The tokenizer 1411 analyzes the given text and classifies it into a morpheme.

Subsequently, the deep neural network model 1412 receives the classified morphemes and the like and performs entity name concatenation with respect to the multilingual term used in the text.

In addition, in the distance-based word-vector model 1413, when converting the first word into a vector in the text, the vector for the first word is considered in consideration of the second word included in the text and the distance from the second word. Will yield.

In addition, the post processor 1414 performs the post processing on the result calculated by the deep neural network model 1412 and the distance-based word-vector model 1413.

In the entity name connection system 100 according to an exemplary embodiment of the present invention, a distance-based word-vector model 1413 may be classified by using an instrumented database 1415 or the like that are difficult to classify into a deep neural network (DNN) model 1412. You can also perform entity name concatenation using

More specifically, FIG. 15 illustrates the operations of the distance based word vector model 1413 and the post processor 1414 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, a relationship between neighboring words and a distance in a sentence with respect to an object name connection target word in a given text is shown. It can be considered and used for analysis. Specifically, when the actor 'Hye-sung' appeared in the drama 'Kim, Chief' and played the role of 'Hong-ga-eun', the words 'jeong-hye-sung' and 'hong-ga-eun' appeared when analyzing 'Kim-jang-jang'. By using the dots, we can see that it is the drama 'Kim Manager' rather than 'Kim Manager' who indicates the position in the company.

In this case, the distance between each word may be used as a weight. Similarly, if you analyze Chung Hye-sung, you can decide Jeong Hye-sung as the actor Jung Hye-sung, taking into account Kim and Jang Hong-eun. Using this method, even if the word 'honggaeun' is not registered or learned, it is possible to tag 'honggaeun' as a character.

For a more specific example, as can be seen in FIG. 15, when performing entity name concatenation for a synonym of "deterministic" in a given text 1510, the distance-based word-vector model is an instrumented database 1520. Keywords (eg, "Kim", "Jeong Hae-sung", "Hong-ga-eun", etc. in FIG. 15) corresponding to the drama "Kim, Chief" can be calculated, and "Kim, Chief," which is the object name concatenated word in the text 1510. The weighting may be given in consideration of distances such as "Jeong Hae Sung" and "Hong Ga Eun" based on "based on" to determine that the "Kim Kim" is a drama "Kim Manager" rather than "Kim Manager" indicating the position in the company.

Furthermore, in performing the entity name link to another multi-word “one piece” included in the text 1510, the entity name link may be performed using the deep neural network model 1530. In consideration of the accuracy of the distance-based word-vector model 1413 and the deep neural network model 1530, the time required for the connection of individual names, and the computational resources, the distance-based word-vector model 1413 and the deep neural network model ( 1530 may be selectively applied, and if necessary, the accuracy of the name connection may be further improved by taking into account 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 results of the distance-based word-vector model 1413 and the neural network model 1530, calculates and compares both, and the like (1540). The connection is performed (1550).

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

In addition, the computer program according to another aspect of the present invention is characterized in that the computer program stored in a computer readable medium for executing each step of the method of linking the salping individual name in the computer. The computer program may be a computer program including not only a computer program including machine code generated by a compiler, but also a high-level language code executable on 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, but is provided with a central processing unit (CPU) such as a server, a smart phone, a tablet PC, a PDA, a mobile phone, and any information processing capable of executing a computer program. Device.

In addition, the computer-readable medium may be to continuously store a computer-executable program, or to temporarily store for execution or download. In addition, the medium may be a variety of recording means or storage means in the form of a single or several hardware combined, not limited to a medium directly connected to any computer system, it may be distributed on the network. Examples of the medium include magnetic media such as hard disks, floppy disks and magnetic tape, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, And ROM, RAM, flash memory, and the like, configured to store program instructions. In addition, examples of another medium may include a recording medium or a storage medium managed by an app store that distributes an application, a site that supplies or distributes various software, a server, or the like.

16 illustrates a configuration diagram 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 includes a word-vector converter 110, a neural network input unit 120, and an entity name connection unit 130. Can be.

In the following, the entity name connection system 100 according to an embodiment of the present invention is divided and searched for each component. For more information on 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, it will be described in more detail below Is omitted.

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

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

Finally, the entity name connection unit 130 performs the entity name connection to the multilingual 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 multiword is used in a different meaning.

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

The above description is merely illustrative of the technical idea of the present invention, and those skilled in the art to which the present invention pertains may make various modifications and changes without departing from the essential characteristics of the present invention. Therefore, the embodiments described in the present invention are not intended to limit the technical spirit of the present invention but to describe the present invention, and are not limited to these embodiments. The protection scope of the present invention should be interpreted 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: object name connection system
110: word-vector converter
120: neural network input unit
130: object name connection

Claims (12)

In the method of performing entity name concatenation for a multiword contained in text,
A word-vector conversion step of converting, by the entity name concatenation system, a plurality of words in the text including the multiword into respective corresponding plurality of vectors;
A neural network input step of inputting the plurality of vectors into a deep neural network having a hidden layer; And
And an entity name concatenation step of performing entity name concatenation with respect to the multilingual word using an output value of the deep neural network.
In the word-vector conversion step,
And selecting a plurality of words within a predetermined distance based on the multiword within the text, and calculating each vector for the plurality of words. delete In the method of performing entity name concatenation for a multiword contained in text,
A word-vector conversion step of converting, by the entity name concatenation system, a plurality of words in the text including the multiword into respective corresponding plurality of vectors;
A neural network input step of inputting the plurality of vectors into a deep neural network having a hidden layer; And
And an entity name concatenation step of performing entity name concatenation with respect to the multilingual word using an output value of the deep neural network.
And converting the transformed M vectors into a plurality of units, calculating them as N vectors (where M > N), and processing them. In the method of performing entity name concatenation for a multiword contained in text,
A word-vector conversion step of converting, by the entity name concatenation system, a plurality of words in the text including the multiword into respective corresponding plurality of vectors;
A neural network input step of inputting the plurality of vectors into a deep neural network having a hidden layer; And
And an entity name concatenation step of performing entity name concatenation with respect to the multilingual word using an output value of the deep neural network.
And a deep neural network learning step of learning the deep neural network using a plurality of text groups in which the multilingual meaning is different from each other. The method of claim 4, wherein
The deep neural network learning step,
Calculating each keyword corresponding to a plurality of meanings of the multiword by using an instrumented database;
Detecting texts in which the multilingual term is used for each of the plurality of meanings using the keyword; And
Learning the deep neural network using the detected texts. The method of claim 4, wherein
In selecting the text to learn the deep neural network from the plurality of text groups,
And a ratio of selecting a text to be learned from each text group in consideration of a result of the entity name linking in each text group. The method of claim 4, wherein
In learning the deep neural network,
And a loss propagation value is adjusted by comparing output values between a plurality of nodes of an output layer corresponding to different meanings of the multiword. The method of claim 4, wherein
The deep neural network learning step,
And training the deep neural network using training data on a common sense entity corresponding to a common noun, a verb, or an adjective. In the method of performing entity name concatenation for a multiword contained in text,
A word-vector conversion step of converting, by the entity name concatenation system, a plurality of words in the text including the multiword into respective corresponding plurality of vectors;
A neural network input step of inputting the plurality of vectors into a deep neural network having a hidden layer; And
And an entity name concatenation step of performing entity name concatenation with respect to the multilingual word using an output value of the deep neural network.
For a first word located within the text,
And the entity name linkage is performed on the first word in consideration of a correlation between the second word included in the text and a distance between the first word and the second word. The method of claim 1,
In the method of performing entity name concatenation for a multiword contained in text,
A word-vector conversion step of converting, by the entity name concatenation system, a plurality of words in the text including the multiword into respective corresponding plurality of vectors;
A neural network input step of inputting the plurality of vectors into a deep neural network having a hidden layer; And
And an entity name concatenation step of performing entity name concatenation with respect to the multilingual word using an output value of the deep neural network.
The output layer of the deep neural network,
A method of concatenating entity names, characterized in that it has a structure including a plurality of nodes distinguished by each synonym used in different meanings. A computer program stored in a computer readable medium for executing each step according to any one of claims 1 to 10 in a computer. In the entity name concatenation system for performing entity name concatenation for a multiword contained in text,
A word-vector converting unit converting a plurality of words including the multiword in the text into a plurality of corresponding vectors;
A neural network input unit for inputting the plurality of vectors into a deep neural network having a hidden layer; And
And an entity name connection unit configured to perform entity name connection to the multilingual word using an output value of the deep neural network.
The output layer of the deep neural network,
An entity name connection system, characterized in that it has a structure including a plurality of nodes which are distinguished by each lexical term used in different meanings.

Images