KR20220132679A - Clinical information search system and method using structure information of natural language - Google Patents

Abstract

The present invention relates to a clinical information search system using natural language structure information and a method thereof. According to the present invention, the clinical information search system using natural language structure information comprises: a natural language processing device preprocessing received natural language data by a rule pattern, then tagging morphemic parts of speech by separating the natural language data into clauses in minimal units, labeling words or phrases as specific semantic types, then extracting the morphemic relationship between objects with labeled semantic types, and bundling embedding vector values of two extracted different objects and relationship information between the two objects into one to calculate bundling object information; a search database (DB) storing vector information of bundled objects made by the natural language processing device; and a clinical test database (DB) built based on clinical test data with bundled objects, and storing clinical test information. Accordingly, the present invention can resolve the structural difference in a sentence such as the order of various object bundles in the sentence by using only the relationship information between objects in natural language information to utilize the relationship information in a search and resolve vocabulary diversity by using a word embedding model to calculate vector values of bundled objects.

Description

Clinical information search system and method using structure information of natural language

The present invention relates to a clinical information retrieval system and method using natural language structure information, and more particularly, only natural language level entities using natural language processing and relationship information between entities without the help of databases such as ontology having semantic information. It relates to a clinical information retrieval system and method using natural language structure information that enables automatic classification of clinical trials according to patients by building a search database using .

Although there are many types of data produced in relation to the medical field, most of them are produced as unstructured text-oriented data.

Various combinations of clinical results are created for each event, such as a patient or experiment, and when they are input in a standardized form, situations that go against the input requirements of the form occur frequently, so it is recommended to enter them in the form of an atypical sentence. in many cases

A user (e.g., a doctor) wants to obtain information about a treatment or treatment according to the results of a clinical trial under similar conditions according to the patient's environment or condition. A search system is supported. In order to construct such a search database, the unstructured text data of clinical trials or literature is formalized and then constructed using a database expressing structural information.

For standardization of unstructured data, natural language processing technology is used to extract natural language level entities and relationships between entities for primary structuring, and secondary structuring is achieved using semantic information such as various ontology defining entity relationships related to medicine. However, this involves a lot of cost to build, and the scope or scale of semantic information is limited.

In addition, at the time of search, a search expression is generated through a standardized and structured input form and a search is requested to the database. Accordingly, the number of items for search increases, the degree of freedom of input is limited, and the complexity of the input form increases. will do

On the other hand, Korean Patent Application Laid-Open No. 10-2016-0030809 (Patent Document 1) discloses "a substitution-based pattern search apparatus and search method for unstructured clinical documents", and thus substitution-based patterns for unstructured clinical documents The search apparatus includes: a natural language processing unit configured to generate normalized text by normalizing an unstructured medical document received through a natural language processing (NLP) process; an entity name recognition unit for identifying Named Entity Recognition information by matching the normalized text and the received domain model; a pattern recognition unit generating structured information from the entity name recognition information using a permutation-based pattern discovery approach; and a template creation unit for creating a prescribed template based on the structured information.

In the case of Patent Document 1 as described above, by extracting structured information from unstructured medical documents, there is an advantage that clinicians or researchers can search and classify information more quickly, but By building a database that expresses relational information by using semantic information such as various ontology, and using complex search rules such as ontology search, access of general users is restricted, and considerable time is spent on learning to use the system. There are downsides to needing it.

Korean Patent Publication No. 10-2016-0030809 (2016.03.21.)

The present invention was created in consideration of the above situation comprehensively, and without the aid of a database such as an ontology having semantic information, a search database is constructed using only natural language level entities using natural language processing and relationship information between entities. Structural differences in sentences, such as the order of various entity bundles in a sentence, can be resolved by using only the relationship information between entities in natural language information to search. An object of the present invention is to provide a clinical information retrieval system and method using natural language structure information that can solve the diversity of vocabulary by calculation.

In order to achieve the above object, the clinical information retrieval system using natural language structure information according to the present invention,

After pre-processing the input natural language data according to the rule pattern, the natural language data is separated into the smallest unit of word, morphological parts-of-speech is tagged, the word or phrase is labeled with a specific semantic type, and the semantic type is labeled object. a natural language processing device for extracting a morphological relationship between the two entities, and for calculating bundled entity information by tying each of the extracted embedding vector values of two different entities and relation information between the two entities into one;

a search database (DB) connected to the natural language processing apparatus through the Internet or a local network and storing vector information of bundled objects created by the natural language processing apparatus; and

It is characterized in that it is connected to the natural language processing device by the Internet or a local network, is built based on bundled individualized clinical trial data, and includes a clinical trial database (DB) that stores clinical trial information.

Here, an interface device capable of receiving unstructured text as an input for search may be further included.

Here, also the natural language processing device,

a natural language preprocessing module for preprocessing natural language data in which letters, numbers, and special symbols are mixed and in which email and web URL rule patterns exist according to the rule patterns, and then creates and refines arbitrary rule patterns of letters, numbers, and special symbols;

a part of speech (POS) analysis module that divides the natural language data into words of a minimum unit of sentence components, and tags parts of speech such as nouns, verbs, and adjectives, which are morphological features, for the corresponding word;

an entity name extraction module for labeling a word or phrase composed of a single word or a plurality of words with a specific semantic type through named entity recognition;

an entity relationship extraction module for extracting morphological relationships between entities labeled with each semantic type with respect to the word or phrase labeled by the entity name extraction module; and

In the entities extracted by the entity relationship extraction module, the embedding vector value of entity A, the embedding vector value of entity B, and the one-hot vector, which is information on the relation between entity A and entity B, are bundled into one. It can be configured by including a bundle individuation module that is expressed as a vector value of .

In this case, the entity name extraction module extracts an entity highly related to a medical term using an entity name recognizer that extracts semantic information about an entity using a semantic type defined in UMLS (Unified Medical Language System). can

In addition, the entity relationship extraction module can extract entity relationships, which are morphological relationships between entities, using an entity relation recognizer that extracts structural relation information defined in the CONLL-U format used in natural language processing.

In addition, the natural language processing apparatus calculates a vector value through a word embedding model using the word of the object for the bundled object, and connects the structural relationship information between the objects together as a connection vector of the bundled object. The vector value can be calculated by extracting bundled objects for each document from the entire clinical/literature unstructured data, and each bundled object can be configured as a matrix for one search.

In addition, the natural language processing apparatus extracts the bundled object of the search sentence and calculates a vector value, calculates the similarity between the bundled objects through the bundled object search matrix and matrix multiplication, and then determines the bundled object with high relevance and By searching the clinical trial database to which the bundled entity belongs, the clinical trial information can be sorted using the similarity and subject frequency to the retrieved clinical trial information.

In addition, in order to achieve the above object, the clinical information search method using natural language structure information according to the present invention,

a) receiving and processing atypical clinical trial data by a natural language processing device to extract bundled individualized clinical trial data;

b) calculating a vector value using each individual word through a word embedding model for the extracted bundle individualized clinical trial data;

c) constructing a search database (DB) having vector information of a bundled entity by the vector value calculation and a clinical trial database (DB) based on the bundled individualized clinical trial data;

d) extracting the bundled individualized patient information data by receiving and processing the unstructured patient information data by the natural language processing device;

e) calculating a vector value using each individual word through a word embedding model for the extracted bundle individualized patient information data;

f) searching the search database (DB) for a bundled entity having bundled entity vector information similar to the calculated bundled entity vector value of the patient information; and

g) It is characterized in that it includes the step of retrieving clinical trial information including a bundle object having a high degree of similarity to the searched bundle object from the clinical trial database (DB).

Here, preferably, after step g), the method may further include h) arranging the retrieved clinical trial information according to a similarity score, and visualizing the search result together with bundled entity information determined to have high similarity.

In addition, in step a), the atypical clinical trial data may include clinical trial-related contents and information published in diseases or treatments, clinical findings, papers, medical journal reports, and the like.

In addition, in receiving and processing the clinical trial atypical data in step a), the natural language data in which letters, numbers, and special symbols are mixed and email and web URL rule patterns exist by the natural language preprocessing module of the natural language processing device After preprocessing by pattern, it can be refined by creating arbitrary regular patterns of letters, numbers and special symbols.

In addition, in receiving and processing the clinical trial unstructured data in step a), natural language data (clinical trial unstructured data) is separated into words of the minimum unit of sentence components by a POS (part of speech) analysis module, and the corresponding word You can tag parts-of-speech such as nouns, verbs, and adjectives, which are morphological features.

In addition, in receiving and processing the clinical trial atypical data in step a), a specific semantic type for a word or phrase consisting of a single word or a plurality of words by named entity recognition by the entity name extraction module can be labeled.

In addition, in receiving and processing the clinical trial atypical data in step a), each semantic type for the word or phrase labeled by the entity name extraction module by the entity relationship extraction module is a morpheme between the labeled entities. adversarial relationships can be extracted.

In addition, in calculating the vector value using the word of each entity through the word embedding model in step b), the embedding vector value of entity A and the embedding vector value of entity B and entity A and entity B by the bundle individualization module One-hot vectors, which are relational information of , can be bundled together and expressed as a single vector value.

According to the present invention, a search database is constructed using only natural language level objects and relationship information between objects using natural language processing without the aid of databases such as ontology having semantic information, and between objects in natural language information. By using only relational information for search, structural differences in sentences such as the order of various entity bundles can be resolved in a sentence, and vocabulary diversity can be solved by calculating vector values for bundle entities using the word embedding model. there are advantages to

1 is a diagram schematically showing the configuration of a clinical information retrieval system using natural language structure information according to the present invention.
2 is a flowchart illustrating an execution process of a clinical information retrieval method using natural language structure information according to the present invention.
3 is a diagram illustrating an overview of analyzing an entity (morpheme) and extracting relationship information, and processing only morphemes related to a medical domain as one token.
4 is a view showing the token1 vector, token2 vector, and relation information one-hot vector in a table for bundled entities.

The terms or words used in the present specification and claims should not be construed as being limited to their ordinary or dictionary meanings, and the inventor may properly define the concept of the term in order to best describe his invention. Based on the principle, it should be interpreted as meaning and concept consistent with the technical idea of the present invention.

Throughout the specification, when a part "includes" a certain element, it means that other elements may be further included, rather than excluding other elements, unless otherwise stated. In addition, terms such as “…unit”, “…group”, “module”, and “device” described in the specification mean a unit that processes at least one function or operation, which is hardware or software or a combination of hardware and software. can be implemented as

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a diagram schematically showing the configuration of a clinical information retrieval system using natural language structure information according to an embodiment of the present invention.

Referring to FIG. 1 , the clinical information search system 100 using natural language structure information according to the present invention includes a natural language processing device 110 , a search database (DB) 120 , and a clinical trial database (DB) 130 . is composed by

The natural language processing device 110 pre-processes the received natural language data (eg, clinical trial atypical data) according to a rule pattern, then divides the natural language data into words of a minimum unit and tags the morphological parts of speech, words or After labeling a phrase with a specific semantic type, the morphological relationship between the entities labeled with the semantic type is extracted, and each embedding vector value of two different extracted entities and the relation information between the two entities are bundled together. yield information.

The search database (DB) 120 is connected to the natural language processing unit 110 through the Internet or a local network, and stores vector information of bundled objects created by the natural language processing unit 110 .

The clinical trial database (DB) 130 is connected to the natural language processing device 110 through the Internet or a local network, is built based on bundled individualized clinical trial data, and stores clinical trial information.

As described above, the clinical information search system 100 using natural language structure information according to the present invention may further include an interface device (not shown) capable of receiving an unstructured text as an input for a search.

Here, the natural language processing device 110 pre-processes natural language data in which letters, numbers, and special symbols are mixed, and in which email and web URL rule patterns exist according to the rule patterns, and then randomly selects letters, numbers, and special symbols. a natural language preprocessing module 111 for creating and refining rule patterns; a POS (part of speech) analysis module 112 that divides the natural language data into words of a minimum unit of sentence components, and tags parts of speech such as nouns, verbs, and adjectives, which are morphological features, for the corresponding word; The entity name extraction module 113 for labeling a word or phrase consisting of a single word or a plurality of words with a specific semantic type (eg, institution, name, place, time, school, etc.) with named entity recognition )class; an entity relationship extraction module 114 for extracting morphological relationships between entities labeled with each semantic type for the word or phrase labeled by the entity name extraction module 113; And in the objects extracted by the entity relationship extraction module 114, the embedding vector value of the object A, the embedding vector value of the object B, and the relationship information between the object A and the B object is a one-hot vector. It may be configured to include a bundle individualization module 115 that binds to one and expresses it as one vector value.

The natural language processing apparatus 110 as described above may be configured as a computer system. In addition, the natural language preprocessing module 111, the POS analysis module 112, the entity name extraction module 113, the entity relationship extraction module 114, and the bundle individualization module 115 each perform a given specific function. Each can be configured as a program.

In this case, the entity name extraction module 113 uses an entity name recognizer that extracts semantic information about an entity using a semantic type defined in UMLS (Unified Medical Language System) to an entity highly related to a medical term. can be extracted.

In addition, the entity relationship extraction module 114 may extract entity relations, which are morphological relations between entities, using entity relation recognizers that extract structural relation information defined in the CONLL-U format used in natural language processing. .

In addition, the natural language processing device 110 calculates a vector value through a word embedding model using the word of the object for the bundled entity, and connects the structural relationship information between the entities together to form the bundled entity. It consists of a connection vector, extracts a bundled entity for each document from the entire clinical/literature unstructured data, calculates a vector value, and configures each bundled entity into a single search matrix.

In addition, the natural language processing device 110 extracts the bundled object of the search sentence, calculates a vector value, calculates the similarity between the bundled objects through the bundled object search matrix and matrix multiplication, and then selects the bundled object with high relevance By determining and searching the clinical trial database 130 to which this bundled entity belongs, the clinical trial information can be sorted using the similarity and entity frequency to the retrieved clinical trial information.

Next, a description will be given of a clinical information retrieval method based on a clinical information retrieval system using natural language structure information according to the present invention having the above configuration.

2 is a flowchart illustrating an execution process of a clinical information retrieval method using natural language structure information according to an embodiment of the present invention.

Referring to FIG. 2 , according to the clinical information search method using natural language structure information according to the present invention, the clinical trial atypical data is first received and processed by the natural language processing device 110 to extract bundled individualized clinical trial data ( step S201).

Then, a vector value is calculated using the word of each individual through the word embedding model for the extracted bundle individualized clinical trial data (step S202).

Then, a search database (DB) 120 having vector information of a bundled entity and a clinical trial database (DB) 130 based on the bundled individualized clinical trial data are constructed by the vector value calculation (step S203).

After the database construction for the natural language data input as described above is completed, the patient information atypical data is received and processed by the natural language processing device 110 to extract the bundled individualized patient information data (step S204).

Then, a vector value is calculated using each individual word through a word embedding model for the extracted bundled individualized patient information data (step S205).

Then, a bundled entity having bundled entity vector information similar to the calculated bundled entity vector value of the patient information is retrieved from the search database (DB) 120 (step S206).

Thereafter, clinical trial information including the searched bundle object and the bundle object having a high degree of similarity is retrieved from the clinical trial database (DB) 130 (step S207).

Here, preferably, after step S207, the method may further include aligning the retrieved clinical trial information according to a similarity score, and visualizing the search result together with bundled entity information determined to have high similarity.

In addition, the clinical trial atypical data in step S201 may include clinical trial-related contents and information published in diseases or treatments, clinical findings, papers, medical journal reports, and the like.

In addition, in receiving and processing the clinical trial atypical data in step S201, letters, numbers, and special symbols are mixed by the natural language preprocessing module 111 of the natural language processing device 110, and there are email and web URL rule patterns. After natural language data is pre-processed by rule patterns, arbitrary rule patterns of letters, numbers, and special symbols can be created and refined.

In addition, in receiving and processing the clinical trial unstructured data in step S201, the natural language data (clinical trial unstructured data) is separated by the POS (part of speech) analysis module 112 into words of the smallest unit of sentence components, Part-of-speech such as nouns, verbs, and adjectives, which are morphological features, can be tagged for the corresponding word.

In addition, in receiving and processing the clinical trial atypical data in step S201, a specific meaning for a word or phrase composed of a single word or a plurality of words through named entity recognition by the entity name extraction module 113 You can label them by type (eg, institution, name, place, time, school, etc.).

In addition, in receiving and processing the clinical trial atypical data in step S201, each semantic type is labeled with respect to the word or phrase labeled by the entity name extraction module 113 by the entity relationship extraction module 114 Morphological relationships between entities can be extracted.

In addition, in calculating the vector value using the word of each entity through the word embedding model in step S202, the embedding vector value of entity A, the embedding vector value of entity B, and entity A by the bundle individualization module 115 One-hot vectors, which are relational information of object B, can be bundled together and expressed as a single vector value.

On the other hand, FIG. 3 is a diagram showing an outline of analyzing an entity (morpheme) and extracting relational information and processing only medical domain related morphemes into one token according to the clinical information retrieval method using natural language structure information according to the present invention, FIG. 4 is a diagram showing the token1 vector, token2 vector, and relation information one-hot vector for bundled entities arranged in a table.

First, as shown in FIG. 3A , morpheme (individual) analysis and relationship information between the individual and the individual are extracted for the input natural language data (eg, clinical trial atypical data). Then, as in (B), only morphemes related to the medical domain are bundled and processed as one token.

In this way, the result of processing tokens for each bundle entity may be organized into a table of bundle entities, token vectors, and relation information one-hot vectors, as shown in FIG. 4 .

The token1 vector arranged in the table can be expressed as {0.23, -0.12, ..., 0.01}, and the token2 vector can be expressed as {-0.13, -0.02, ..., 0.14}, A relation one-hot vector may be expressed as {0, 0, 0, ..., 1}. Based on this, the bundled entity vector can be expressed as "token1 vector + token2 vector + relation one-hot vector".

As described above, the clinical information retrieval system and method using natural language structure information according to the present invention uses only natural language level entities and relationship information between entities using natural language processing without the aid of databases such as ontology having semantic information. By constructing a search database and using only the relationship information between objects in natural language information for search, structural differences in sentences such as the order of various object bundles in a sentence can be resolved, and bundled objects using the word embedding model There is an advantage in that the diversity of vocabulary can be resolved by calculating the vector value for .

In addition, by supporting a natural language-based search, there is an advantage that a user can use the search system without a separate learning process for using the system.

In addition, by using atypical text data such as clinical findings among patient EMR (Endoscopic Mucosal Resection) data in hospitals for search terms with only natural language processing level processing, it has the advantage of automatically performing classification of clinical trials according to patients. .

As described above, the present invention has been described in detail through preferred embodiments, but the present invention is not limited thereto, and it is common in the art that various changes and applications can be made without departing from the technical spirit of the present invention. self-explanatory to the technician. Therefore, the true 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.

100: (the present invention) clinical information retrieval system using natural language structure information
110: natural language processing unit 111: natural language preprocessing module
112: POS analysis module 113: entity name extraction module
114: entity relationship extraction module 115: bundle entityization module

Claims (15)

After pre-processing the input natural language data according to the rule pattern, the natural language data is separated into the smallest unit of word, morphological parts-of-speech is tagged, the word or phrase is labeled with a specific semantic type, and the semantic type is labeled object. a natural language processing device for extracting a morphological relationship between the two entities, and for calculating bundled entity information by tying each of the extracted embedding vector values of two different entities and relation information between the two entities into one;
a search database (DB) connected to the natural language processing apparatus through the Internet or a local network and storing vector information of bundled objects created by the natural language processing apparatus; and
A clinical information retrieval system using natural language structure information that is connected to the natural language processing device and the Internet or a local network, is built based on bundled individualized clinical trial data, and includes a clinical trial database (DB) that stores clinical trial information.
According to claim 1,
Clinical information retrieval system using natural language structure information, further comprising an interface device capable of receiving unstructured text as an input for retrieval.
According to claim 1,
The natural language processing device,
a natural language preprocessing module for preprocessing natural language data in which letters, numbers, and special symbols are mixed and in which email and web URL rule patterns exist according to the rule patterns, and then creates and refines arbitrary rule patterns of letters, numbers, and special symbols;
a part of speech (POS) analysis module that divides the natural language data into words of a minimum unit of sentence components, and tags parts of speech such as nouns, verbs, and adjectives, which are morphological features, for the corresponding word;
an entity name extraction module for labeling a word or phrase composed of a single word or a plurality of words with a specific semantic type through named entity recognition;
an entity relationship extraction module for extracting morphological relationships between entities labeled with each semantic type with respect to the word or phrase labeled by the entity name extraction module; and
In the entities extracted by the entity relationship extraction module, the embedding vector value of entity A, the embedding vector value of entity B, and the one-hot vector, which is information on the relation between entity A and entity B, are bundled into one. A clinical information retrieval system using natural language structure information, including a bundle individualization module that is expressed as a vector value of
4. The method of claim 3,
The entity name extraction module uses a semantic type defined in UMLS (Unified Medical Language System) to extract an entity highly related to a medical term using an entity name recognizer that extracts semantic information about the entity. Clinical information retrieval system using information.
4. The method of claim 3,
The entity relationship extraction module uses natural language structure information to extract entity relationships that are morphological relationships between entities using an entity relation recognizer that extracts structural relation information defined in the CONLL-U format used in natural language processing. search system.
4. The method of claim 3,
The natural language processing device calculates a vector value through a word embedding model using the words of the object for the bundled entity, and connects the structural relationship information between the entities together to form a connection vector of the bundled entity, , a clinical information retrieval system using natural language structure information that extracts bundled objects for each document from all clinical/literature atypical data, calculates vector values, and configures each bundled object into a single search matrix.
4. The method of claim 3,
The natural language processing device extracts the bundled entity of the search sentence and calculates a vector value, calculates the similarity between the bundled entities through the bundled entity search matrix and matrix multiplication, then determines the bundled entity with high relevance and determines the bundled entity A clinical information retrieval system using natural language structure information that searches the clinical trial database to which the group belongs, and aligns the clinical trial information using the similarity and individual frequency to the retrieved clinical trial information.
a) receiving and processing atypical clinical trial data by a natural language processing device to extract bundled individualized clinical trial data;
b) calculating a vector value using each individual word through a word embedding model for the extracted bundle individualized clinical trial data;
c) constructing a search database (DB) having vector information of a bundled entity by the vector value calculation and a clinical trial database (DB) based on the bundled individualized clinical trial data;
d) extracting the bundled individualized patient information data by receiving and processing the unstructured patient information data by the natural language processing device;
e) calculating a vector value using each individual word through a word embedding model for the extracted bundle individualized patient information data;
f) searching the search database (DB) for a bundled entity having bundled entity vector information similar to the calculated bundled entity vector value of the patient information; and
g) a clinical information retrieval method using natural language structure information, comprising the step of retrieving clinical trial information including a bundle object having a high degree of similarity to the searched bundle object from the clinical trial database (DB).
9. The method of claim 8,
After step g), h) arranging the retrieved clinical trial information according to the similarity score, and visualizing the search result together with the bundled entity information determined to have high similarity. Clinical information search using natural language structure information Way.
9. The method of claim 8,
In step a), the atypical data of the clinical trial is a clinical information search method using natural language structure information including the clinical trial-related content and information published in a disease or treatment, clinical findings, a thesis, or a medical journal report. 9. The method of claim 8,
In step a), in receiving and processing atypical clinical trial data, letters, numbers, and special symbols are mixed by the natural language preprocessing module of the natural language processing device, and natural language data in which email and web URL rule patterns exist is added to the rule pattern. Clinical information retrieval method using natural language structure information that creates and refines arbitrary rule patterns of letters, numbers, and special symbols after pre-processing by
9. The method of claim 8,
In receiving and processing the clinical trial atypical data in step a), the natural language data (clinical trial atypical data) is separated into words of the minimum unit of sentence components by the POS (part of speech) analysis module, and A clinical information retrieval method using natural language structure information that tags parts of speech such as nouns, verbs, and adjectives, which are morphological features.
9. The method of claim 8,
In step a), in receiving and processing the clinical trial atypical data, a word or phrase consisting of a single word or a plurality of words is labeled with a specific semantic type by named entity recognition by the entity name extraction module. Clinical information retrieval method using natural language structure information.
14. The method of claim 13,
In receiving and processing the clinical trial atypical data in step a), the morphological relationship between the entities labeled by each semantic type with respect to the word or phrase labeled by the entity name extraction module by the entity relationship extraction module Clinical information retrieval method using natural language structure information to extract
9. The method of claim 8,
In calculating the vector value using the word of each entity through the word embedding model in step b), the embedding vector value of entity A and the embedding vector value of entity B and the relationship between entity A and entity B by the bundle individualization module A clinical information retrieval method using natural language structure information that binds one-hot vectors, which are information, and expresses them as one vector value.

Images