KR20210106884A - Apparatus and method for emotion classification based on artificial intelligence for online data - Google Patents

Abstract

Disclosed are a device and a method for emotion classification of online data based on artificial intelligence. According to an embodiment of the present application, a method for emotion classification of online data based on artificial intelligence includes the steps of: preparing learning data including a plurality of online data to which first emotion information is given; building an artificial intelligence-based emotion classification model for recognizing emotions reflected in target online data when the target online data is received based on the learning data; and receiving the target online data and recognizing the emotion corresponding to the target online data based on the emotion classification model. Therefore, it is possible to optimize the emotion classification model based on the analysis results of factors which make it difficult to identify main emotions reflected in the online data.

Description

AI-based emotion classification device and method for online data

The present application relates to an artificial intelligence-based emotion classification apparatus and method for online data.

Emotions play an important role in conveying the speaker's intentions and expressing thoughts, and influence the formation and maintenance of interpersonal relationships and decision-making. Therefore, emotion analysis can be a means for understanding various aspects of human beings (eg, cognition, thinking, behavior, expression, etc.) present in various situations.

A representative area of emotion analysis research is the development of emotion models that perform automatic emotion classification, and research on these emotion models has received considerable attention in several academic fields, such as artificial intelligence and human computer interaction, and accordingly The model is being developed.

However, in order to build such an emotion model more reliably, various aspects should be carefully considered. First, the reliability of the emotion label should be considered as a major consideration. In relation to this, in the prior research, it was common to build an emotion model using an emotion label inferred through a heuristic method such as an emotion label or hashtag arbitrarily defined by a commenter who is a third party other than the author of the post. .

However, there is a limitation that the emotion specified by the commenter is defined by a third party and is not directly specified by the author, and heuristic labeling has a problem in that it is not possible to be sure whether the hashtag accurately represents the emotion of the author. . Therefore, in both of the above-described methods, there is a limit to the reliability of the emotion label.

Also, the second aspect is the lack of reliable data due to manual labeling, which may be related to the first aspect described above. That is, since manual labeling of emotion labels by annotators inevitably requires significant human effort such as time and cost, conventional studies adopting this manual labeling method are generally limited to 1,000 to 10,000 human-annotations. It was only at the level of constructing an emotion model using human-annotated data.

The technology that is the background of the present application is disclosed in Korean Patent Publication No. 10-1713558.

The present application is to solve the problems of the prior art described above, and is collected to include a large number of online data to which a reliable emotion label is given, and online data for constructing an emotion classification model based on learning data with a high degree of suitability for emotion analysis The purpose of the present invention is to provide an artificial intelligence-based emotion classification device and method for

However, the technical problems to be achieved by the embodiments of the present application are not limited to the technical problems as described above, and other technical problems may exist.

As a technical means for achieving the above technical task, the AI-based emotion classification method for online data according to an embodiment of the present application prepares learning data including a plurality of online data to which first emotion information is given building an artificial intelligence-based emotion classification model for recognizing emotions reflected in the target online data when the target online data is received based on the learning data; receiving the target online data; and classifying the emotions It may include recognizing an emotion corresponding to the target online data based on the model.

In addition, the preparing may include defining second emotion information corresponding to at least one emotion category based on the first emotion information and classifying the plurality of online data based on the second emotion information. may include.

In addition, the second emotion information may be defined to include a smaller number of emotion categories than the first emotion information.

Also, the second emotion information may be defined based on a Parrott emotion model.

In addition, the AI-based emotion classification method for online data according to an embodiment of the present application may include evaluating the consistency and representativeness of the plurality of online data classified in response to the second emotion information. .

In addition, the receiving may include receiving the target online data to which the first emotion information is not given.

In addition, the building step may include tokenizing each of the plurality of online data, extracting a part corresponding to a preset part-of-speech from each of the plurality of tokenized online data, and BERT based on the extracted part (Bidirectional Encoder Representations from Transformers) may include generating a based emotion classification model.

In addition, the identifying may include determining an emotion category corresponding to a main emotion reflected in the target online data from among the emotion categories included in the second emotion information based on the emotion classification model.

Also, the second emotion information may include at least one emotion category of anger, fear, sadness, surprise, joy, and love.

Also, the second emotion information may include a first emotion category classified as a positive emotion and a second emotion category classified as a negative emotion.

Also, the emotion classification model may be separately generated for the first emotion category and the second emotion category.

In addition, the constructing may include subdividing the second emotion information included in the first emotion category based on psychological definitions and emotional situations.

On the other hand, the artificial intelligence-based emotion classification apparatus for online data according to an embodiment of the present application, a collection unit for preparing learning data including a plurality of online data to which first emotion information is given, based on the learning data , When the target online data is received, a learning unit that builds an artificial intelligence-based emotion classification model for recognizing emotions reflected in the target online data and the target online data are received, and the target online data is added to the target online data based on the emotion classification model It may include an analysis unit for identifying the corresponding emotion.

In addition, the collection unit may include an emotion setting unit defining second emotion information corresponding to at least one emotion category based on the first emotion information, and data for classifying the plurality of online data based on the second emotion information It may include a classification unit.

In addition, the learning unit tokenizes each of the plurality of online data, and a data processing unit for extracting a part corresponding to a preset part-of-speech from each of the plurality of tokenized online data, and a BERT ( Bidirectional Encoder Representations from Transformers) may include a model building unit for generating the based emotion classification model.

Also, the analysis unit may determine an emotion category corresponding to the main emotion reflected in the target online data from among the emotion categories included in the second emotion information based on the emotion classification model.

The above-described problem solving means are merely exemplary, and should not be construed as limiting the present application. In addition to the exemplary embodiments described above, additional embodiments may exist in the drawings and detailed description.

According to the above-described problem solving means of the present application, artificial intelligence for online data that is collected to include a large number of online data to which a reliable emotion label is given and builds an emotion classification model based on learning data with a high degree of suitability for emotion analysis It is possible to provide an apparatus and method for classifying emotions based on the present invention.

According to the above-described problem solving means of the present application, the utility of the emotion classification model can be further improved by optimizing the emotion classification model based on the analysis result of the factors that make it difficult to grasp the main emotion reflected in the online data.

However, the effects obtainable herein are not limited to the above-described effects, and other effects may exist.

1 is a schematic configuration diagram of an online data analysis system including an artificial intelligence-based emotion classification device for online data according to an embodiment of the present application.
FIG. 2 is a diagram exemplarily illustrating a plurality of online data collected to correspond to the second emotion information and each emotion category of the second emotion information.
3 is a diagram illustrating a conventional emotion data set that is exemplarily compared in order to evaluate learning data collected by an artificial intelligence-based emotion classification apparatus for online data according to an embodiment of the present application.
4A and 4B are diagrams illustrating statistical analysis of data deviation between learning data collected by an artificial intelligence-based emotion classification apparatus for online data and a conventional emotion data set according to an embodiment of the present application.
5 is a table showing the comparison between the learning data collected by the AI-based emotion classification apparatus for online data according to an embodiment of the present application and the consistency of the conventional emotion data set.
6 is a conceptual diagram for explaining the function and structure of an artificial intelligence-based emotion classification model built by an artificial intelligence-based emotion classification apparatus for online data according to an embodiment of the present application.
7 is an experimental example linked to an artificial intelligence-based emotion classification technique for online data according to an embodiment of the present application, and is a chart illustrating performance evaluation results of an emotion classification model.
8 is an experimental example linked to an artificial intelligence-based emotion classification technique for online data according to an embodiment of the present application. It is a diagram showing the characteristics of data by way of example.
9 is a schematic configuration diagram of an artificial intelligence-based emotion classification apparatus for online data according to an embodiment of the present application.
10 is an operation flowchart of an AI-based emotion classification method for online data according to an embodiment of the present application.
11 is a detailed operation flowchart of a processing process of learning data required for construction of an emotion classification model.

Hereinafter, embodiments of the present application will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art to which the present application pertains can easily implement them. However, the present application may be embodied in several different forms and is not limited to the embodiments described herein. And in order to clearly explain the present application in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

Throughout this specification, when a part is "connected" with another part, it is not only "directly connected" but also "electrically connected" or "indirectly connected" with another element interposed therebetween. "Including cases where

Throughout this specification, when it is said that a member is positioned "on", "on", "on", "under", "under", or "under" another member, this means that a member is positioned on the other member. It includes not only the case where they are in contact, but also the case where another member exists between two members.

Throughout this specification, when a part "includes" a component, it means that other components may be further included, rather than excluding other components, unless otherwise stated.

The present application relates to an artificial intelligence-based emotion classification apparatus and method for online data.

1 is a schematic configuration diagram of an online data analysis system including an artificial intelligence-based emotion classification device for online data according to an embodiment of the present application.

Referring to FIG. 1 , the online data analysis system 10 according to an embodiment of the present application is an artificial intelligence-based emotion classification apparatus 100 for online data according to an embodiment of the present application (hereinafter, 'emotion classification apparatus') (referred to as '100)'), a storage server 200 , and a user terminal 300 .

The emotion classification apparatus 100 , the storage server 200 , and the user terminal 300 may communicate with each other through the network 20 . The network 20 refers to a connection structure in which information exchange is possible between each node, such as terminals and servers, and an example of such a network 20 includes a 3rd Generation Partnership Project (3GPP) network, a long-term LTE (LTE) network. Term Evolution) network, 5G network, WIMAX (World Interoperability for Microwave Access) network, Internet, LAN (Local Area Network), Wireless LAN (Wireless Local Area Network), WAN (Wide Area Network), PAN (Personal Area) Network), a wifi network, a Bluetooth network, a satellite broadcasting network, an analog broadcasting network, a Digital Multimedia Broadcasting (DMB) network, etc. are included, but are not limited thereto.

The user terminal 300 is, for example, a smartphone (Smartphone), a smart pad (SmartPad), a tablet PC and the like and PCS (Personal Communication System), GSM (Global System for Mobile communication), PDC (Personal Digital Cellular), PHS ( Personal Handyphone System), PDA (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)-2000, W-CDMA (W-Code Division Multiple Access), Wibro (Wireless Broadband Internet) terminals The same may be any type of wireless communication device.

In the description of the embodiment of the present application, the storage server 200 may be a server or device for storing learning data required for learning an artificial intelligence-based emotion classification model to be described in detail below.

The emotion classification apparatus 100 may prepare learning data including a plurality of online data to which the first emotion information is given. According to an embodiment of the present disclosure, the emotion classification apparatus 100 may receive learning data including a plurality of online data from the storage server 200 . As another example, the emotion classification apparatus 100 may receive, as learning data, online data created by the corresponding user terminal 300 from at least one user terminal 300 .

In this regard, the first emotion information may correspond to an emotion classification tag assigned by each creator of the online data. According to an embodiment of the present application, the online data collected as learning data for the construction of an artificial intelligence-based emotion classification model to be described later is created by the author selecting at least one of a plurality of emotion classification tags preset at the time of writing, and thus The emotion classification tag selected by the user may be data that is posted (uploaded) together with the content (content) of online data.

On the other hand, in the description of the embodiments of the present application, 'online data' refers to blogs, social network service (SNS)-based posts (data), reviews, comments, online postings, voice data, image data, etc. broadly. may include In addition, the online data herein may be data including at least a portion of data in text format, but is not limited thereto, and is text-type data by itself such as voice data, image data, etc. according to an embodiment of the present application. is not included, but may include a wide range of data and the like having a format in which the voice reflected in the data can be converted based on a predetermined conversion means (eg, STT algorithm, etc.) and processed into text data.

In this regard, in most conventional techniques for emotion analysis or emotion classification, emotion data inferred in a heuristic manner through a hashtag or the like is used as learning data by annotators independent of the creator of the online data. Therefore, not only there is a reliability problem for the emotion label given or inferred for the training data, but also a significant amount of time and effort is required for the manual labeling operation, so it was difficult to secure a large amount of data for learning.

On the other hand, the emotion classification apparatus 100 disclosed herein utilizes online data including first emotion information corresponding to the emotion classification tag directly selected (allocated) by the creator as learning data, thereby providing each emotion of the learning data. The reliability of the label is high, and since a separate manual process for allocating emotion information or giving an emotion label is not required, it is relatively easy to collect a very large amount of online data as learning data. Therefore, according to the present application, not only can a large amount of learning data be utilized, but also the accuracy (reliability) of the emotion information reflected in the online data is very high, so the accuracy of the emotion classification model built using this can be dramatically improved. there will be

As an example to help understanding, the emotion classification device 100 may collect posts created based on the online social community “TalkLife” service created for the purpose of peer support and social connection as learning data, but is not limited thereto. . In particular, the "TalkLife" service has a characteristic that requires the process of selecting the emotion of the author of the post from among dozens of predefined emotion classification tags.

Hereinafter, for convenience of explanation, based on the learning data collected based on the "TalkLife" service (in an experimental example related to an embodiment of the present application, about 900,000 multiple posts are collected as learning data). An embodiment of the present application will be described on the assumption that a process of building an emotion classification model is performed.

The emotion classification apparatus 100 may perform pre-processing of selecting online data including text corresponding to a preset language from among a plurality of collected online data. Illustratively, the emotion classification device 100 may operate to exclude online data written based on Arabic, French, Portuguese, etc. by using a Python-based isalpha function to perform preprocessing for screening online data written in English, However, the present invention is not limited thereto.

The emotion classification apparatus 100 may define second emotion information corresponding to at least one emotion category based on the first emotion information. In this regard, the second emotion information may be defined to include a smaller number of emotion categories than the first emotion information. The second emotion information may be defined from the emotion information.

With respect to the aforementioned Parrott emotion model, various psychological models are used in emotion research, and are largely classified into two perspectives, and the classification of these two perspectives is discrete emotion and dimensional emotion, respectively. First, according to discrete emotions, emotions can be classified based on physiological, behavioral, and expressive characteristics, and the Ekman model is a representative example. The Ekman model defined six emotions (specifically, happiness, sadness, surprise, fear, disgust, and anger) related to facial expressions. In addition, the Plutchik model, which is another emotion model according to discrete emotions, defined 8 emotions (specifically, joy and sadness, trust and disgust, fear and anger, surprise and expectation) as 4 opposite pairs, but 8 emotions It may not be sufficient to categorize a person's various emotions by itself. On the other hand, the Parrott model used to define the second emotion information herein has a classification system of three layers (levels) to cover a wider range of emotions, and the first layer divides the main emotions into anger, fear, and sadness. , surprise, joy, and love were classified into 6 categories, and 25 emotion categories were considered in the second layer, and the emotion categories were expanded to 115 more subdivided categories in the third layer.

In this regard, the emotion classification device 100 receives the second emotion information from the first emotion information to include at least one emotion information of anger, fear, sadness, surprise, joy, and love, which is considered in the first layer of the Parrot model. can be defined In addition, according to an embodiment of the present application, the emotion classification apparatus 100 determines each emotion of the first emotion information to define the first emotion information including a plurality of emotion categories compared to the second emotion information as the second emotion information. First emotional information to respond to any one of emotional information of anger, fear, sadness, surprise, joy, and love based on the similarity of the category with the subdivided emotional category considered in the second or third layer of the Parrot model It is possible to determine the second emotion information for each.

FIG. 2 is a diagram exemplarily illustrating a plurality of online data collected to correspond to the second emotion information and each emotion category of the second emotion information.

Referring to FIG. 2 , according to an embodiment of the present application, the emotion classification apparatus 100 receives first emotion information including a plurality of emotion classification tags provided by the author based on a Parrott emotion model including a hierarchical emotion category. It is possible to assign (group) the second emotional information corresponding to the emotional categories of anger, fear, sadness, surprise, joy, and love, respectively. More specifically, if the emotion classification apparatus 100 includes 59 predefined emotion classification tags in the learning data collected based on the "TalkLife" service, "Meh", "Numb", "Nothing", The first emotion information including 55 emotion classification tags excluding the four emotion classification tags that have little relevance to emotional expression such as "Inadequate" as the second emotion information, each of anger, fear, sadness, surprise, joy and love Pre-processing of assigning (grouping) 7, 6, 14, 3, 21, and 4 emotion classification tags to emotion categories may be performed.

Hereinafter, the evaluation result of comparing the learning data collected by the emotion classification apparatus 100 disclosed herein with a conventional emotion data set will be described with reference to FIGS. 3 to 5 .

3 shows the suitability of the learning data collected by the artificial intelligence-based emotion classification apparatus for online data according to an embodiment of the present application (more specifically, the suitability as learning data for building an artificial intelligence-based emotion classification model) It is a diagram showing a conventional emotional data set that is exemplarily compared in order to evaluate the reflecting consistency and representativeness).

Referring to FIG. 3 , as a conventional emotion data set, Crowdflower data, Emotion-Stimulus data, and ISEAR data may be considered. In addition, referring to FIG. 3 , the Crowdflower data may include 14 emotion classifications, and the Emotion-Stimulus data and ISEAR data may include 7 emotion classifications, respectively. For reference, Emotion-Stimulus data and ISEAR data do not include data corresponding to the emotion category of 'love', so for Emotion-Stimulus data and ISEAR data, the emotion categories of anger, sadness, fear, and joy are four emotions except love. was compared with the training data in the present application.

In the experiment to evaluate the suitability of the learning data of the present application, text preprocessing is applied to each emotion data, a model for each emotion data is built through the Word2Vec technique, which is an embedding method, and a predetermined analysis means (for example, Word2Vec's most_similar function, etc.), a plurality of (eg, 20) words judged to be the most semantically similar for each emotion category were extracted for each emotion data set. Afterwards, statistical tests were performed based on the questionnaire results on the consistency and representativeness of emotional data based on the words extracted from multiple respondents, followed by normality and equality of variance tests, followed by ANOVA analysis and Tukey's post-test. did.

In this regard, FIGS. 4A and 4B are tables showing statistical analysis of data deviations between the learning data collected by the AI-based emotion classification apparatus for online data and the conventional emotion data set according to an embodiment of the present application. am.

4A and 4B, for all emotion categories (anger, sadness, fear, joy, love), the word set derived from the learning data collected by the emotion classification apparatus 100 disclosed herein and three conventional It can be confirmed that there is a significant group difference between the word sets derived from the emotion data set.

In addition, FIG. 5 is a chart showing the comparison between the learning data collected by the AI-based emotion classification apparatus for online data according to an embodiment of the present application and the consistency of the conventional emotion data set.

Referring to FIG. 5 , as a result of Tukey's post-mortem analysis, a questionnaire about whether words included in a word set (keyword set) consistently describe the corresponding emotion category (more specifically, a score of 1 to 5 for each emotion category) A survey is performed to assign, but a value closer to 1 corresponds to a case that is judged to be inconsistent, and a value closer to 5 corresponds to a case judged to be consistent) As a result, learning in the present application collected based on the "TalkLife" service It can be confirmed that the consistency evaluation result value for the data is evaluated as a value larger (in other words, a value close to 5) than the evaluation result value for each of the remaining conventional emotional data sets, so that the learning data in this application is evaluated as the best in terms of consistency. can

In addition, in the above-described experimental example in connection with the emotion classification apparatus 100 according to an embodiment of the present application, as a result of a questionnaire request to select a data set that best expresses and represents emotion, the "TalkLife" service collected based on The adoption rate for the learning data in the present application was the highest at 95.5%, confirming that the learning data in the present application was evaluated as the best in terms of representativeness.

Therefore, according to the above-described experimental results, the emotional expression (in other words, classification of the emotion category reflected in the second emotion information) considered by the emotion classification apparatus 100 according to an embodiment of the present application is used for constructing an emotion classification model. It can be confirmed that the suitability is high in terms of consistency and representativeness compared to the existing emotion data set used in previous studies, and accordingly, more objective evaluation and improvement through emotion modeling performed based on the learning data collected through the above process It can be reasonably expected to be able to present a specific point of view on

Also, according to an embodiment of the present application, the emotion classification apparatus 100 may evaluate the consistency and representativeness of a plurality of online data classified in response to the second emotion information.

For example, the emotion classification apparatus 100 receives questionnaire data for each emotion category of the second emotion information of online data evaluated by a plurality of respondents, and quantifies the consistency and representativeness of the learning data based on the received questionnaire data. (arithmetic) can be done.

In addition, the emotion classification apparatus 100 may determine whether the learning data collection (preparation) process is completed based on the evaluation result of at least one of consistency and representativeness of each of the quantified (calculated) emotion categories. For example, the emotion classification apparatus 100 removes at least a portion of the online data collected in advance for an emotion category in which the evaluation result of the evaluated consistency or representativeness is less than or equal to a preset threshold level, or removes at least a portion of the online data corresponding to the emotion category. It may act to further collect data.

6 is a conceptual diagram for explaining the function and structure of an artificial intelligence-based emotion classification model built by an artificial intelligence-based emotion classification apparatus for online data according to an embodiment of the present application.

Referring to FIG. 6 , the emotion classification apparatus 100 may build an artificial intelligence-based emotion classification model for recognizing emotions reflected in the received target online data when the target online data is received, based on the collected learning data. .

Also, referring to FIG. 6 , the emotion classification apparatus 100 may build an emotion classification model based on Bidirectional Encoder Representations from Transformers (BERT).

More specifically, according to an embodiment of the present application, the emotion classification apparatus 100 may tokenize each of a plurality of online data included in the training data in order to build a BERT-based emotion classification model. In this regard, according to an embodiment of the present application, the emotion classification apparatus 100 may tokenize each of the online data using a predetermined natural language processing means (eg, Natural Language Toolkit, NLKT, etc.). have.

Also, the emotion classification apparatus 100 may extract a portion corresponding to a preset part-of-speech from each of the plurality of tokenized online data. For example, the preset part-of-speech may be set to include nouns, adjectives, and adverbs that are highly likely to include content related to emotional expression.

Also, the emotion classification apparatus 100 may generate a BERT-based emotion classification model based on the extracted part. BERT (Bidirectional Encoder Representations from Transformers) is an artificial intelligence model based on a transformer architecture of encoder-decoder structure. It builds a plurality of transformer layers for in-depth representation of input, and a masking language model that is a token sequence. It is characterized by applying a masking process to (Masking Language Model).

Unlike the existing transformer procedure that masks words sequentially by replacing the word to be predicted with a mask token, BERT includes a procedure to determine the bidirectional characteristics of the model by randomly masking the percentage of words. Therefore, BERT has the advantage of being able to obtain interactive information of the masked word through the corresponding procedure, and taking a human-like approach when selecting the missing word within the context.

In this regard, herein, a BERT model including a plurality of layers (eg, 12) of transformer blocks may be constructed. More specifically, each block on the encoder side of the BERT model herein may be designed with a structure including a 12-head self-attention layer and a 768-dimensional hidden layer, and thus the total It can operate to yield 1.1 million parameters. On the other hand, with respect to the fine-tuning of the BERT-based model of the present application, most hyperparameters except for the batch size (eg, 64) may be maintained the same as the pre-trained model.

As described above, after learning (building) of the AI-based emotion classification model is completed, the emotion classification apparatus 100 may receive target online data. Here, the target online data may include online data to which the first emotion information is not provided, but is not limited thereto. As another example, the emotion classification apparatus 100 receives, as target online data, the online data to which the first emotion information is given, including the emotion classification tag assigned (granted) by the author, but to the online data to which the first emotion information is given In response, it may operate to evaluate the level of matching or similarity between the emotion identified through the AI-based emotion classification model from the content (content) of the target online data and the first emotion information allocated to the online data.

According to an embodiment of the present application, the emotion classification apparatus 100 may receive online data written by a user of the user terminal 300 from the user terminal 300 .

Also, the emotion classification apparatus 100 may identify an emotion corresponding to the target online data based on the constructed (generated) emotion classification model. Specifically, the emotion classification apparatus 100 may determine an emotion category corresponding to the main emotion reflected in the target online data from among the emotion categories included in the second emotion information based on the emotion classification model.

More specifically, the emotion classification apparatus 100 divides text data included in the received target online data into predetermined units (eg, sentence units, paragraph units, etc.), and divides each divided unit text data into artificial intelligence. Based on the emotion classification model, the emotion category mainly reflected in the target online data is identified based on statistical/probabilistic processing for the emotion category analyzed for each unit text data, and this is at least one of the second emotion information can be matched to the types of emotions of

In other words, referring to FIG. 6 , the emotion classification apparatus 100 according to an embodiment of the present application provides an emotion classification apparatus 100 based on a BERT-based emotion classification model constructed based on the training data (Labeled Text Input, 1). ) may operate to output at least one emotion category (Emotion Label, 2) of the second emotion information reflected in the applied target online data.

Hereinafter, a technique for optimizing the emotion classification model based on the evaluation result of the emotion classification performance by the emotion classification apparatus 100 disclosed herein and the emotion classification performance will be described with reference to FIGS. 7 and 8 .

7 is an experimental example linked to an artificial intelligence-based emotion classification technique for online data according to an embodiment of the present application, and is a chart illustrating performance evaluation results of an emotion classification model.

Referring to FIG. 7 , according to an experimental example performed to evaluate the performance of the BERT-based emotion classification model of the emotion classification apparatus 100, the collected online data is divided into 80% training data and 20% verification data. As a result of performing 5-fold cross validation, it can be confirmed that the BERT-based emotion classification model disclosed herein exhibits an accuracy of 81% for the entire emotion category, which is a conventional emotion classification model. It can be seen that the accuracy increases significantly compared to the 70% performance level derived from the classification study.

8 is an experimental example linked to an artificial intelligence-based emotion classification technique for online data according to an embodiment of the present application. It is a diagram showing the characteristics of data by way of example.

Referring to FIG. 8 , emotion information (second emotion information) derived by the emotion classification model of the emotion classification apparatus 100 for predetermined target online data and a predetermined surveyor directly check and evaluate the target online data The target online data determined to be relatively inconsistent with emotion information (the survey result of FIG. 8) includes a plurality of sentences, but the emotions reflected in each of the plurality of sentences have different characteristics, or the boundaries of the plurality of emotion categories themselves are vague, or It was analyzed that the intensity of a specific emotion had ambiguous characteristics.

More specifically, according to an embodiment of the present application, the second emotion information may be defined to include a first emotion category classified as a positive emotion and a second emotion category classified as a negative emotion. Specifically, the first emotion category may include joy, love, and the like. In addition, the second emotion category may include anger, fear, and the like.

Also, in relation to this, the emotion classification apparatus 100 may operate to separately generate an emotion classification model for the first emotion category and the second emotion category. In other words, since positive emotions often have more blurred boundaries between different emotional categories (e.g., joy and love) compared to negative emotions, they are separate from the emotion classification model for identifying the emotional categories classified as negative emotions. It can be built to have detailed settings (eg, weights, parameters, etc. of the BERT model).

More specifically, according to an embodiment of the present application, the emotion classification apparatus 100 generates (constructs) an emotion classification model corresponding to a first emotion category corresponding to a positive emotion, and the second emotion category included in the first emotion category is generated (constructed). It is possible to optimize (update) the emotional category by subdividing the emotional information based on the psychological definition and emotional situation of each emotion. For example, the emotion classification apparatus 100 provides psychological definitions and emotions such as 'joy' and 'interest' instead of second emotion information that is a positive emotion such as 'love' but is difficult to match with a single emotional situation. The second emotion information for learning the emotion classification model corresponding to the first emotion category may be defined to include the emotion category that can be more clearly distinguished in light of the situation.

9 is a schematic configuration diagram of an artificial intelligence-based emotion classification apparatus for online data according to an embodiment of the present application.

Referring to FIG. 9 , the emotion classification apparatus 100 may include a collection unit 110 , a learning unit 120 , and an analysis unit 130 . Also, referring to FIG. 9 , the collection unit 110 includes an emotion setting unit 111 and a data classification unit 112 , and the learning unit 120 includes a data processing unit 121 and a model building unit 122 . may include.

The collection unit 110 may prepare learning data including a plurality of online data to which the first emotion information is given.

Specifically, the emotion setting unit 111 may define second emotion information corresponding to at least one emotion category based on the first emotion information given to a plurality of online data included in the learning data. Also, the data classification unit 112 may classify the plurality of collected online data based on the second emotion information defined by the emotion setting unit 111 .

The learning unit 120 may build an artificial intelligence-based emotion classification model for recognizing emotions reflected in the target online data when the target online data is received, based on the prepared training data.

Specifically, the data processing unit 121 may tokenize each of the plurality of online data, and extract a portion corresponding to a preset part-of-speech from each of the plurality of tokenized online data.

Also, the model building unit 122 may generate an emotion classification model based on Bidirectional Encoder Representations from Transformers (BERT) based on the extracted part of speech.

The analysis unit 130 may receive target online data. In addition, the analysis unit 130 may identify an emotion corresponding to the received target online data based on the learned emotion classification model. Specifically, the analysis unit 130 may determine an emotion category corresponding to the main emotion reflected in the target online data from among the emotion categories included in the second emotion information based on the generated emotion classification model.

Hereinafter, an operation flow of the present application will be briefly reviewed based on the details described above.

10 is an operation flowchart of an AI-based emotion classification method for online data according to an embodiment of the present application.

The AI-based emotion classification method for online data shown in FIG. 10 may be performed by the emotion classification apparatus 100 described above. Therefore, even if omitted below, the description of the emotion classification apparatus 100 may be equally applied to the description of the artificial intelligence-based emotion classification method for online data.

Referring to FIG. 10 , in step S11 , the collection unit 110 may prepare learning data including a plurality of online data to which first emotion information is given.

Next, in step S12 , the learning unit 120 may build an artificial intelligence-based emotion classification model for recognizing emotions reflected in the target online data when the target online data is received based on the prepared training data.

Specifically, in step S12 , the data processing unit 121 may tokenize each of the plurality of online data. Also, in step S12 , the data processing unit 121 may extract a portion corresponding to a preset part-of-speech from each of the plurality of tokenized online data.

Also, in step S12 , the model building unit 122 may generate an emotion classification model based on Bidirectional Encoder Representations from Transformers (BERT) based on the extracted part-of-speech part.

Meanwhile, the second emotion information may include a first emotion category classified as a positive emotion and a second emotion category classified as a negative emotion. And the second emotion information included in the first emotion category may be subdivided based on psychological definitions and emotional situations so as to individually generate an emotion classification model for the second emotion category.

Next, in step S13 , the analysis unit 130 may receive the target online data. According to an embodiment of the present application, in step S13, the analysis unit 130 may receive the target online data to which the first emotion information is not assigned.

Next, in step S14 , the analysis unit 130 may identify an emotion corresponding to the received target online data based on the learned emotion classification model.

Specifically, in step S14, the analysis unit 130 may determine an emotion category corresponding to the main emotion reflected in the target online data from among the emotion categories included in the second emotion information based on the generated emotion classification model.

In the above description, steps S11 to S14 may be further divided into additional steps or combined into fewer steps, according to an embodiment of the present application. In addition, some steps may be omitted as necessary, and the order between steps may be changed.

11 is a detailed operation flowchart of a processing process of learning data required for construction of an emotion classification model.

The AI-based emotion classification method for online data shown in FIG. 11 may be performed by the emotion classification apparatus 100 described above. Accordingly, even if omitted below, the description of the emotion classification apparatus 100 may be equally applied to the description of FIG. 11 .

Referring to FIG. 11 , in step S121 , the emotion setting unit 111 defines second emotion information corresponding to at least one emotion category based on the first emotion information given to a plurality of online data included in the learning data. can

Next, in step S122 , the data classification unit 112 may classify the plurality of collected online data based on the second emotion information defined in step S1212 .

In the above description, steps S121 and S122 may be further divided into additional steps or combined into fewer steps, according to an embodiment of the present application. In addition, some steps may be omitted as necessary, and the order between steps may be changed.

The AI-based emotion classification method for online data according to an embodiment of the present application may be implemented in the form of a program command that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the art of computer software. Examples of the computer readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

In addition, the AI-based emotion classification method for online data described above may be implemented in the form of a computer program or application executed by a computer stored in a recording medium.

The foregoing description of the present application is for illustration, and those of ordinary skill in the art to which the present application pertains will understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present application. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. For example, each component described as a single type may be implemented in a dispersed form, and likewise components described as distributed may be implemented in a combined form.

The scope of the present application is indicated by the following claims rather than the above detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalents should be construed as being included in the scope of the present application.

10: Online data analysis system
100: Artificial intelligence-based emotion classification device for online data
110: collection unit
111: emotion setting unit
112: data classification unit
120: study unit
121: data processing unit
122: model building unit
130: analysis unit
20: network
200: storage server
300: user terminal

Claims (15)

In the AI-based emotion classification method for online data,
Preparing learning data including a plurality of online data to which the first emotion information is given;
building an artificial intelligence-based emotion classification model for recognizing emotions reflected in the target online data when the target online data is received based on the learning data;
receiving the target online data; and
Recognizing an emotion corresponding to the target online data based on the emotion classification model;
Including, emotion classification method. According to claim 1,
The preparing step is
defining second emotion information corresponding to at least one emotion category based on the first emotion information; and
classifying the plurality of online data based on the second emotion information;
A method for classifying emotions that includes. 3. The method of claim 2,
Wherein the second emotion information is defined to include a smaller number of emotion categories than the first emotion information. 4. The method of claim 3,
The second emotional information,
A method for classifying emotions, which is defined based on a Parrott emotion model. 3. The method of claim 2,
evaluating the consistency and representativeness of the plurality of online data classified in response to the second emotion information;
Which will further include, emotion classification method. According to claim 1,
The receiving step is
The emotion classification method of receiving the target online data to which the first emotion information is not granted. 3. The method of claim 2,
The building step is
tokenizing each of the plurality of online data;
extracting a portion corresponding to a preset part-of-speech from each of the plurality of tokenized online data; and
Generating a BERT (Bidirectional Encoder Representations from Transformers)-based emotion classification model based on the extracted part;
A method for classifying emotions that includes. 5. The method of claim 4,
The step of figuring out is
The emotion classification method of determining an emotion category corresponding to the main emotion reflected in the target online data from among the emotion categories included in the second emotion information based on the emotion classification model. 9. The method of claim 8,
The second emotional information,
and comprising at least one emotion category of anger, fear, sadness, surprise, joy and love. 10. The method of claim 9,
The second emotional information,
a first emotion category classified as positive emotion and a second emotion category classified as negative emotion;
The emotion classification model is characterized in that it is separately generated for the first emotion category and the second emotion category, emotion classification method. 11. The method of claim 10,
The building step is
subdividing the second emotion information included in the first emotion category based on psychological definitions and emotional situations;
A method for classifying emotions that includes. In an artificial intelligence-based emotion classification device for online data,
a collecting unit for preparing learning data including a plurality of online data to which first emotion information is given;
a learning unit for constructing an artificial intelligence-based emotion classification model for recognizing emotions reflected in the target online data when the target online data is received, based on the learning data; and
an analysis unit for receiving the target online data and identifying emotions corresponding to the target online data based on the emotion classification model;
Including, emotion classification device. 13. The method of claim 12,
The collection unit,
an emotion setting unit defining second emotion information corresponding to at least one emotion category based on the first emotion information; and
a data classification unit for classifying the plurality of online data based on the second emotion information;
That comprising a, emotion classification device. 14. The method of claim 13,
The learning unit,
a data processing unit for tokenizing each of the plurality of online data and extracting a portion corresponding to a preset part-of-speech from each of the plurality of tokenized online data; and
A model building unit for generating an emotion classification model based on BERT (Bidirectional Encoder Representations from Transformers) based on the extracted part;
That comprising a, emotion classification device. 14. The method of claim 13,
The second emotion information is defined based on a Parrott emotion model,
The analysis unit,
The emotion classification apparatus that determines an emotion category corresponding to the main emotion reflected in the target online data from among the emotion categories included in the second emotion information based on the emotion classification model.

Images