KR102337290B1 - Method And Apparatus for Generating Context Category Dataset - Google Patents

Abstract

The present disclosure provides an apparatus and method for generating a context category dataset.
According to an aspect of the present disclosure, a context category dataset for predicting a context category to which a hashtag belongs by a user, and generating and updating a context category dataset by receiving an input from a user of a context category to which a hashtag belongs A generating apparatus and method are provided.

Description

Apparatus and method for generating context category dataset {Method And Apparatus for Generating Context Category Dataset}

The present invention relates to an apparatus and method for generating a context category dataset.

The content described in this section merely provides background information for the present embodiment and does not constitute the prior art.

Techniques related to natural language processing and application, such as text generation, natural language generation, and intelligent agent, are being advanced. However, it is difficult to provide a dataset that can design and train a model used for such natural language processing and application.

One of the ways to create a dataset that can design or train a model is to use a crowd-sourcing method. Crowdsourcing is a compound word of Crowd and Outsourcing, and it means to engage the public in various production activities, including data collection. Crowdsourcing can shorten the time required for the production process by involving a large number of people in production activities.

However, as the crowdsourcing method takes cost and time in proportion to the size of the data to be generated, recently, the method of producing datasets using machine prediction of artificial intelligence (AI) models has been adopted. cases are increasing. However, datasets generated by machine prediction have a disadvantage in that accuracy is lower than datasets generated by direct classification or labeling by humans.

Accordingly, recently, classification or labeling of data using a human-machine collaboration method has been introduced. This human-machine collaboration is in the form of a machine preemptively presenting classification or labeling prediction results, and humans reviewing the prediction results. This collaborative method is expected to be helpful in the fields of technology related to natural language processing and application described above, especially in that it uses feedback from humans.

Therefore, it is necessary to devise a data set creation method for classifying natural language using a human-machine collaboration method.

US Patent Publication No. 2014-0337257 (2014.11.13)

According to one aspect of the present disclosure, a context category dataset for predicting a context category to which a hashtag entered by a user will belong, and generating and updating a context category dataset by receiving a context category to which a hashtag belongs from the user The main object is to provide an apparatus and method for generating it.

According to an aspect of the present disclosure, in an apparatus for generating a context category dataset using a user interface (UI), a hashtag list for each context category is provided. List providing unit to provide; and a category prediction unit for predicting one or more context categories of hashtag information input from the user interface using word embedding vectors generated for each context category based on the hashtag list. comprising, wherein the user interface provides a predicted context category to a user, receives context category information from the user, and provides it to the list providing unit. to provide.

According to another aspect of the present disclosure, a process of generating a word embedding vector for each context category based on a hashtag list for each context category; a process of receiving hashtag information from a user interface (UI); predicting one or more context categories of the hashtag information by using the word embedding vector; providing a predicted context category to a user through the user interface; receiving context category information from the user; and generating or updating the hashtag list based on the context category information.

According to one aspect of the present disclosure, there is provided an apparatus and method for predicting a context category to which a hashtag entered by a user will belong, and for generating and updating a context category dataset by receiving an input from a user of a context category to which a hashtag belongs. By providing it, there is an effect that it is possible to preemptively provide a natural language dataset classified in consideration of the human situation and context.

The natural language classification result in consideration of the human situation and context is text generation that generates text describing a specific situation in consideration of the context of natural language, and automatically recognizes the user situation and context to generate appropriate words and sentences It can be used to design and train a machine learning model or artificial intelligence model for the purpose of intelligent agent or natural language generation.

1 is a conceptual diagram for explaining a process of generating a context category dataset of the present disclosure.
2 is a block diagram illustrating an apparatus for generating a context category dataset according to an embodiment of the present disclosure.
3A and 3B are exemplary diagrams of a word embedding vector and an embedding vector of hash tag information according to an embodiment of the present disclosure.
4 is an exemplary diagram of a user interface according to an embodiment of the present disclosure.
5 is a flowchart illustrating a method for generating a context category dataset according to an embodiment of the present disclosure.

Hereinafter, some embodiments of the present disclosure will be described in detail with reference to exemplary drawings. It should be noted that, in adding the reference numerals to the components of each drawing, the same components are to have the same reference numerals as much as possible even if they are displayed on different drawings. In addition, in describing the present disclosure, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present disclosure, the detailed description thereof will be omitted.

In addition, in describing the components of the present disclosure, terms such as second and first may be used. These terms are only for distinguishing the elements from other elements, and the essence, order, or order of the elements are not limited by the terms. Throughout the specification, when a part 'includes' or 'includes' a certain element, this means that other elements may be further included, rather than excluding other elements, unless otherwise stated. . In addition, the '... Terms such as 'unit' and 'module' mean a unit that processes at least one function or operation, which may be implemented as hardware or software or a combination of hardware and software.

DETAILED DESCRIPTION The detailed description set forth below in conjunction with the appended drawings is intended to describe exemplary embodiments of the present disclosure and is not intended to represent the only embodiments in which the present disclosure may be practiced.

In the present disclosure, a context category dataset refers to a dataset that provides a hashtag or words of a hashtag having a specific category and contextual connectivity. This context category dataset is created by presenting a machine prediction result for a context category to which a hashtag will belong to a user, and improving the accuracy of context category prediction by receiving a review result from the user. That is, the context category dataset in the present disclosure is generated and collected by a human-machine crowd sourcing method.

The context category dataset of the present disclosure will be described on the premise that it is configured as a hashtag list according to the context category. However, the data structure of the context category dataset is not necessarily limited to the list structure, and hashtags belonging to context categories such as tree and hash table are stored and managed (e.g. : If it is a data structure that can be created, deleted, searched, traversed, referenced, etc.), it may be a data structure of the context category dataset in the present disclosure.

It is assumed that each of a context category, an element of a category list, a hashtag included in the hashtag information, and a context category included in the context category information in the present disclosure may be one or more.

1 is a conceptual diagram for explaining a process of generating a context category dataset of the present disclosure.

When receiving hashtag information composed of one or more hashtags from the user, the apparatus for generating a context category dataset presents a context category to which each hashtag of the hashtag information is expected to belong to the user. Prediction of context categories is performed using a word embedding vector generated based on the hashtag list of each context category. Here, the word embedding vector is a location of a context category in an embedding vector space, and is a vector calculated based on an embedding vector obtained by word embedding of a category list element of the context category. Such a word embedding vector may be obtained, for example, by calculating a centroid from a cluster consisting of each embedding vector of a hashtag list element corresponding to each context category.

The user selects a context category for each hashtag based on the context category presented from the context category dataset generating device. That is, by making the same selection as the presented context category or by modifying all or part of the presented context category, the context category for each hashtag as context category information is provided to the context category dataset generating device. The context category dataset generating apparatus updates the context category dataset based on the provided context category information and previous hashtag information in order to improve prediction accuracy of the context category. Such updating may be accomplished by, for example, adding a hashtag or each word of a hashtag as a new element to the hashtag list of each context category according to context category information, or replacing an existing element.

2 is a block diagram illustrating an apparatus for generating a context category dataset according to an embodiment of the present disclosure.

The context category dataset generating apparatus 200 according to an embodiment of the present disclosure includes all or all of a list provider 210 , a category prediction unit 220 , and a user interface 230 . includes some The context category dataset generating apparatus 200 shown in FIG. 2 is according to an embodiment of the present disclosure, and not all components shown in FIG. 2 are essential components, and some components are added, changed, or may be deleted. For example, in another embodiment, the context category dataset generating apparatus further includes a performance estimate unit (not shown) that performs performance evaluation of context category prediction used for the regeneration of word embedding vectors of each context category. may include

2 shows the context category dataset generating apparatus 200 as a device for convenience of explanation, this is for convenience of description, and in another embodiment, the context category dataset generating device includes the components 210 to 230 It may be implemented as a software module or a processor that performs a function.

The list providing unit 210 generates and manages a hash tag list for each context category as a context category dataset, and provides it to the category prediction unit 220 . The list providing unit 210 may generate a hashtag list of a new context category based on the hashtag information and context category information input from the user interface 230 or update the hashtag list of an existing context category. have.

The list providing unit 210 may generate and update the hash tag list after preprocessing the hash tag information. Such pre-processing, for example, collectively transforms each hashtag included in the hashtag information into uppercase or lowercase letters, removes spaces or special characters included in each hashtag, or a plurality of words and a plurality of characters constituting each hashtag , may be a task of stochastically classifying a combination of words and numbers, and the like, but is not limited thereto.

The category prediction unit 220 uses the word embedding vector generated for each context category based on the hashtag list provided by the list providing unit 210 , the context category of the hashtag information input from the user interface 230 . predict more than one Here, predicting the context category of the hashtag information means predicting one or more context categories to which each hashtag included in the hashtag information belongs. Specifically, the category prediction unit 220 calculates the probability that the vector provider 222 generates or updates all or part of the word embedding vector of each context category and the hashtag information belongs to a specific context category for each context category. to include all or part of a prediction unit 224 for predicting one or more contextual categories of hashtag information. However, not all components shown in FIG. 2 are essential components, and some components may be added, changed, or deleted.

When the vector providing unit 222 generates the first hashtag list, a cluster composed of embedding vectors of the basic tags generated by performing word embedding on one or more basic tags preset in each hashtag list. is created on a predefined embedding vector space. The dimension of the embedding vector space may be redefined by a parameter preset in the vector providing unit 222 , a minimum dimension required for word embedding, or the prediction performance of the prediction unit 224 . The vector providing unit 222 sets the center of each cluster as a word embedding vector of a corresponding context category. Each set word embedding vector may be updated by reflecting the embedding vector of an element added to the corresponding hashtag list. For example, the vector providing unit 222 reconfigures the cluster whenever the number of elements newly added to the hashtag list of a specific context category becomes a preset number, finds a new center, and resets it to a word embedding vector of the corresponding context category. have.

The prediction unit 224 generates an embedding vector of each hashtag in the hashtag information and calculates the distance between the embedding vector of the hashtag and the embedding vector of the word in each context category for each hashtag, respectively, so that the corresponding hashtag is Calculate the probability of belonging to each context category, respectively. Specifically, the prediction unit 224 calculates the distance between the extracted embedding vector and each word embedding vector when the embedding vector of the hashtag can be extracted to generate the embedding vector of each hashtag, and cannot be extracted. In this case, the distance is calculated by classifying the hashtags by word and calculating the sum of the embedding vectors of each word as an embedding vector of the hashtag. The prediction unit 224 may provide the top N context categories with the shortest distance (N is a natural number equal to or greater than 1) based on the calculated distance to the user interface 230 as the predicted context category of the corresponding hashtag.

As another example, the prediction unit 224 normalizes the calculated distance for each context category, calculates the probability that the hashtag information belongs to each context category, and selects all or part of the context category in which the probability value is equal to or greater than a preset threshold. It may be provided to the user interface 230 as a predicted context category of the corresponding hashtag.

The user interface 230 receives the hashtag information from the user and provides it to the list providing unit 210 and/or the category prediction unit 220, and then provides the predicted context category provided from the category prediction unit 220 to the user. After providing, context category information is received from the user and provided to the list providing unit 210 . A specific example of the user interface 230 will be described later with reference to FIG. 4 .

3A and 3B are exemplary diagrams of a word embedding vector and an embedding vector of hash tag information according to an embodiment of the present disclosure.

A, B, and C of FIGS. 3A and 3B are word embedding vectors of context categories, respectively, showing the center of a cluster consisting of embedding vectors of corresponding hashtag list elements on a three-dimensional embedding vector space. Although the embedding vector space is shown as a three-dimensional space in FIGS. 3A and 3B, this is exemplary and it is obvious that the dimension of the embedding vector space may be three-dimensional or more or less.

In the hashtag information input by the user through the user interface (the hashtag in FIGS. 3A and 3B), FIG. 3A shows a case in which an embedding vector of hash tag information can be extracted in an embedding vector space, and FIG. 3B shows a case where it is not extractable. indicates Here, extracting the embedding vector of the hashtag information means extracting the embedding vector for each hashtag included in the hashtag information, and whether it is possible to extract even between the hashtags included in the same hashtag information may vary. have.

When the embedding vector can be extracted, the context category dataset generating apparatus extracts the embedding vector of the hashtag information. However, when the embedding vector of the hashtag information cannot be extracted, the context category dataset generating apparatus assumes that the hashtag information is composed of a plurality of words and classifies the words included in the hashtag information by using an algorithm or library ( library) and the like to extract embedding vectors (hatag#1 and hashtag#2 in FIG. 3b) of the classified words, respectively. Thereafter, the context category dataset generating apparatus obtains an embedding vector of hash tag information by summing the embedding vectors of each word.

The context category dataset generating device calculates the distance between the extracted or acquired hashtag information embedding vector and the word embedding vector of each context category, and it is predicted that the closer the distance, the higher the probability that the hashtag information belongs to the context category. do.

4 is an exemplary diagram of a user interface according to an embodiment of the present disclosure.

In the embodiment of FIG. 4 , connectivity between a hashtag and a context category is provided using an image. That is, the user inputs image and hashtag information into the user interface, and when the context category predicted by the context category dataset generating device is provided, the user selects or modifies the context category information into the user interface. The context category dataset generating apparatus increases the accuracy of prediction by regenerating or updating the word embedding vector of the hashtag list and/or context category based on the received context category information and hashtag information, and the context category Data sets can be accumulated.

Referring to FIG. 4A , the context category dataset generating apparatus receives an image and one or more hashtags related to the image from the user as hashtag information from the user interface. By receiving the image and the hashtag together, it can be expected that there is a context related to the image input together between the inputted hashtags. In order to facilitate the creation of the context category dataset, the user interface preferably requests to input a hashtag of two or more characters or requests to input a certain number of such hashtags (eg, five or more). In addition, when the user interface creates overlapping hashtag information in order to collect hashtags valid for generating the context category dataset, it is preferable to invalidate it and request a re-entry from the user.

In the example of FIG. 4 ( a ), images of trees and rivers are input to the user interface, and hashtags of #strasboug, #petitefrance, #sundaymorning, #morningwalk, and #christmasvacation are inputted in relation to each other.

Accordingly, the context category dataset generating apparatus extracts or obtains an embedding vector of each hashtag included in the hashtag information to predict the context category, and provides the predicted context category to the user through the user interface. In order to easily obtain an embedding vector of each hashtag, the context category dataset generating apparatus may pre-process each hashtag (eg, classification by word, case conversion, space or symbol removal, etc.). Referring to (b) and (c) of Figure 4, the context category includes emotion, mood, location, time, object, activity, and event. ), it can be seen that other (Other) is included, but it is self-evident that this is not limited thereto.

Referring to (b) of FIG. 4 , the context category dataset generating device is a context category for each hashtag, with a location category in #strasboug, a location category in #petitefrance, a time category in #sundaymorning, and a time category and activity in #morningwalk. The category, #christmasvacation, provides the time category and event category as predicted context categories.

4C shows a result of the user selecting a context category through the user interface. The user excluded the event category from the context category of the hashtag #christmasvacation among the provided context categories. The context category dataset generating apparatus updates the hashtag list for each context category based on context category information that is a user's selection result. For example, you can add strasboug and petitefrance to the hashtag list of the location category, sundaymorning, morningwalk, and christmasvacation to the hashtag list of the time category, and morningwalk to the hashtag list of the activity category.

As another example, the apparatus for generating a context category dataset may add a word of a hashtag to each hashtag list by performing pre-processing of classifying each hashtag by word. For example, to add sunday, morning, morningwalk, christmas, vacation instead of sundaymorning, morningwalk, and christmasvacation to the hashtag list of the time category.

5 is a flowchart illustrating a method for generating a context category dataset according to an embodiment of the present disclosure.

The context category dataset generating apparatus generates a word embedding vector for each context category by using the hashtag list for each context category (S500).

The context category dataset generating apparatus receives hashtag information using a user interface (S510). The hashtag information may be one or more hashtags related to materials (eg, images, videos, documents, etc.) uploaded through the user interface.

As the location of the embedding vector in space, it is determined whether it is possible to extract the embedding vector of the received hashtag information (S520). If it is determined that the embedding vector can be extracted, the distance between the embedding vector of the hashtag information and each word embedding vector is calculated, and the context corresponding to the top N word embedding vectors in the order of the shortest distance (N is a natural number greater than or equal to 1) The categories are calculated as predicted context categories (S530). Alternatively, context categories corresponding to embedding vectors having a distance equal to or greater than a preset threshold are calculated as predicted context categories.

If it is determined that the embedding vector cannot be extracted, the apparatus for generating the context category dataset obtains an embedding vector of the hashtag information from the embedding vector of each word in which the hashtag information is classified (S522). Thereafter, the context category is predicted by calculating the distance between the embedding vector of the hashtag information and each word embedding vector ( S530 ).

The context category dataset generating apparatus provides the predicted context category to the user through the user interface, and receives context category information from the user ( S540 ). Accordingly, the user inputs context category information into the user interface in consideration of the predicted context category.

Based on the hashtag information and the context category information, a new element is added to the hashtag list for each context category, or an element is added to the new context category and the hashtag list (S550).

Although it is described that each process is sequentially executed in FIG. 5 , this is merely illustrative of the technical idea of an embodiment of the present disclosure. In other words, those of ordinary skill in the art to which an embodiment of the present disclosure pertain may change the order described in FIG. Since it is possible to apply various modifications and variations by executing in parallel, it is not limited to the time-series order of FIG. 5 .

Various implementations of the devices, units, processes, steps, etc., described herein may include digital electronic circuits, integrated circuits, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), computer hardware, firmware, software, and/or a combination thereof. These various implementations may include being implemented in one or more computer programs executable on a programmable system. The programmable system includes at least one programmable processor (which may be a special purpose processor) coupled to receive data and instructions from, and transmit data and instructions to, a storage system, at least one input device, and at least one output device. or may be a general-purpose processor). Computer programs (also known as programs, software, software applications or code) contain instructions for a programmable processor and are stored on a "computer-readable recording medium".

The computer-readable recording medium includes all types of recording devices in which data readable by a computer system is stored. These computer-readable recording media are non-volatile or non-transitory, such as ROM, CD-ROM, magnetic tape, floppy disk, memory card, hard disk, magneto-optical disk, and storage device. It may further include a medium or a transitory medium such as a data transmission medium. In addition, the computer-readable recording medium may be distributed in network-connected computer systems, and computer-readable codes may be stored and executed in a distributed manner.

Various implementations of the systems and techniques described herein may be implemented by a programmable computer. Here, the computer includes a programmable processor, a data storage system (including volatile memory, non-volatile memory, or other types of storage systems or combinations thereof), and at least one communication interface. For example, the programmable computer may be one of a server, a network appliance, a set-top box, an embedded device, a computer expansion module, a personal computer, a laptop, a Personal Data Assistant (PDA), a cloud computing system, or a mobile device.

The above description is merely illustrative of the technical idea of this embodiment, and a person skilled in the art to which this embodiment belongs may make various modifications and variations without departing from the essential characteristics of the present embodiment. Accordingly, the present embodiments are intended to explain rather than limit the technical spirit of the present embodiment, and the scope of the technical spirit of the present embodiment is not limited by these embodiments. The protection scope of this embodiment should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be interpreted as being included in the scope of the present embodiment.

200: Context category dataset generator
210: list providing unit
220: category prediction unit
222: vector provider
224: prediction unit
230: user interface

Claims (16)

In an apparatus for generating a context category dataset using a user interface (UI),
a list providing unit that provides a hashtag list for each context category; and
A category prediction unit for predicting one or more context categories of the hashtag information input from the user interface by using the word embedding vector generated for each context category based on the hashtag list but,
The user interface is
Providing the predicted context category to the user, receiving context category information from the user and providing it to the list providing unit
Context category dataset generating device, characterized in that. According to claim 1,
The word embedding vector is
A context category dataset generating apparatus, characterized in that it is a vector indicating the location of a context category corresponding to itself on a predefined embedding vector space. According to claim 1,
The list providing unit,
Creating or updating the hashtag list by creating a hashtag list of a new context category or updating the hashtag list of an existing context category based on the hashtag information and the context category information
Context category dataset generating device, characterized in that. 4. The method of claim 3,
The list providing unit,
Creating or updating the hash tag list after performing preprocessing of classifying the hash tag information by word
Context category dataset generating device, characterized in that. According to claim 1,
The category prediction unit,
a vector providing unit for generating or updating the word embedding vector; and
A prediction unit predicting one or more context categories to which the hashtag information belongs by calculating a probability that the hashtag information belongs to the context category by using the word embedding vector
Context category dataset generating device, characterized in that it comprises a. 6. The method of claim 5,
The word embedding vector generation by the vector providing unit,
Based on one or more basic tags preset in the hashtag list, the centroid of a cluster consisting of the embedding vectors of the basic tags is assigned to the word embedding vector of the corresponding context category. A device for generating a contextual category dataset characterized by it. 6. The method of claim 5,
The word embedding vector update by the vector providing unit,
When there is an addition of an element to the hashtag list, a word embedding vector of a context category corresponding to the center of a cluster consisting of an embedding vector of an element included in the hashtag list by further considering the embedding vector of the added element Context category dataset generating apparatus, characterized in that performed by resetting (reassign) to. 8. The method of claim 7,
The reset is
The apparatus for generating a context category dataset, characterized in that it is performed whenever the number of hashtags added to the hashtag list becomes a preset number. 6. The method of claim 5,
The prediction unit,
Context category dataset creation, characterized in that by generating each embedding vector of the hashtag information and calculating a distance from each of the word embedding vectors, the probability that the hashtag information belongs to each of the context categories Device. 10. The method of claim 9,
The embedding vector of the hashtag information is,
It is generated by extracting an embedding vector from the hashtag information, and when the embedding vector cannot be extracted from the hashtag information, the hash tag information is classified by word and the embedding vector of each word is summed. Context category dataset generator. 10. The method of claim 9,
The prediction unit,
The apparatus for generating a context category dataset, characterized in that all or part of a context category in which the distance between the embedding vector of the hashtag information and each of the word embedding vectors is equal to or greater than a preset threshold is predicted as the context category of the hashtag. generating a word embedding vector for each context category based on a hashtag list for each context category;
a process of receiving hashtag information from a user interface (UI);
predicting one or more context categories of the hashtag information by using the word embedding vector;
providing a predicted context category to a user through the user interface;
receiving context category information from the user; and
A process of newly creating or updating the hashtag list based on the context category information
Context category dataset creation method comprising a. 13. The method of claim 12,
The process of generating the word embedding vector is,
A method for generating a context category dataset, characterized in that assigning a centroid of a cluster composed of embedding vectors of elements included in the hashtag list to a word embedding vector of a corresponding context category. 13. The method of claim 12,
The process of predicting one or more context categories,
A method for generating a context category dataset, comprising generating an embedding vector of the hashtag information and predicting it based on calculating a distance from each of the word embedding vectors. 15. The method of claim 14,
The prediction is
The distance from each of the word embedding vectors is normalized for each corresponding context category to calculate the probability that the hashtag information belongs to each context category, and the top N items with a high calculated probability (N is a natural number greater than or equal to 1) A method for creating a context category dataset, characterized in that it is performed by selecting a context category of A computer program stored in a computer-readable recording medium to execute each process included in the method for generating a context category dataset according to any one of claims 12 to 15.

Images